Ballentine_PC inventory: run the PowerShell one-liner from docs/PC_AIENOS_MAP.md (L209) and send the file
Drake: skip the brother PC inventory for now (parked). Gist stays up.
done when file received
Ledger line 213 on GitHubNow
What is being done, right now.
This is the working ledger of AIEN, published as it is written. Every line is a promise with an owner and an observable finish. Agents write their progress here; the operator writes decisions here. Nothing on this page is composed by hand.
6agents writing in the last two hours
147lines in motion
110lines queued
24settled in the last day
3waiting on the operator
Ledger last written 2026-10-22T12:14Z. Page built 2026-10-22T12:19Z. It rebuilds within the hour of every change.
- memory free65 GiB of 122
- disk free930G of 3.7T
- wall power150 W (UPS, 10 W steps)
- agents writing0 in the last two hours
- published2026-10-22T12:15:02Z
On the operator's desk
Decisions Drake is making
Work that cannot move until a human answers. Each line names the answer it needs.
DRAKE DECISION 2026-10-21 ~01:15Z (binding): stop spending time on the plain single-path 35B; the motor (L312) is the 35B server on 18091, Renda on it; trust the mathematics: gate = identical words and routes vs the host reference + negative control. L306 parked (open), its cosine-floor question parked with it, not waived
21:50Z motor (batch strokes v2) serves 18091, gate2 PASS live (see L385). Renda harness reads 18091 and resumes on its next pass; Renda-answering receipt to follow.
done when motor serves 18091 with gate 1 (chip) and gate 2 (HTTP) PASS; Renda answering through it
Ledger line 339 on GitHubDRAKE DECISION 2026-10-21 ~03:00Z (binding, to survey chat 47937b63 Fable): APPROVED and PROCEED through the night: use the motor (35B on 18091) to get the PR backlog done (census ~/handoffs/2026-10-21-merge-census.md landing order), Renda reviewing on the motor, running 8 conversations at once (L354 stage 1 -> 18091 with --batch-slots 8 as soon as its exact-words gate passes; then L312 multi-row chip stroke). Extends L334/L351/L352: the desk lanes and their judge runs are night work, not blocked by MOTOR ONLY.
done when done when: 18091 serves 8 rows, Renda runs 8 reviews at once, census batches A-B landed, then C..H
Ledger line 363 on GitHubIn motion
What the agents are doing
Grouped by stream. The bar under each name is the stream's pulse: one segment per line, bright when written in the last half hour, fading as it cools.
Engine speed
17 linesAlpha merges in order: omega alpha/decode-speed-to-main b224aaf1, sc alpha/serve-path d584ab4d (+omega.lock bump, abi-qualify), sc alpha/try-it-guide 71f1fe59
desk lane resumed as sa-1-0edae162 (sc PRs); omega/aienos lanes: chain scripts still running
done when three squash shas on main
Ledger line 8 on GitHubPublish all cascade measurements on GitHub as they land (agents push evidence; coordinator checks each branch has evidence/ for every number reported)
wip-snapshot tool (~/.local/bin/wip-snapshot DIR...) pushes any worktree state to origin wip/<branch>-<stamp> without touching it; run 18:16Z: 6 dirty worktrees snapshotted
done when each reported number links to a pushed evidence file
Ledger line 12 on GitHubAgent: tree-verify-gemm
2026-10-20T15:24:52Z hermes: no claude -p process for this owner since last night; branches and evidence are pushed per the notes above. Needs a close-out read (what remains vs DONE WHEN) before re-dispatch; GPU queue was jammed 18h until 15:20Z (L041 note), now draining
done when handoff written + branch pushed with evidence + result reported to Drake
Ledger line 14 on GitHubAgent: prefill-kernels
01:14Z from efficiency review (Wolfram): prefill runs at 11.81 TFLOPS = 1/20.1 of the measured FP8 tensor rate 237.5, and 1/2.59 of its own best GEMM shape (30.6); findings G1-omega-5, G2-sc-inference-5
done when handoff written + branch pushed with evidence + result reported to Drake
Ledger line 18 on GitHubAgent: dflash-drafter
2026-10-20T15:24:52Z hermes: no claude -p process for this owner since last night; branches and evidence are pushed per the notes above. Needs a close-out read (what remains vs DONE WHEN) before re-dispatch; GPU queue was jammed 18h until 15:20Z (L041 note), now draining
done when handoff written + branch pushed with evidence + result reported to Drake
Ledger line 20 on GitHubQueue sc alpha/ceiling-and-model 56d42869 after alpha/try-it-guide; omega fix/decode-speed-gpr-test 585df2f5 into qwen38/decode-speed
from L298 (fixer): sc#698 shows as CRUMB-DRIFT only because crumb-doctor (deleted by #698) cannot judge an Am2 tree; no crumb file in the PR is wrong. Merge needs a rebase (DIRTY); after merge run aien-watch clear. Details: ledger L298.
done when squash shas
Ledger line 35 on GitHubImplement L006: QWEN38-SERVE and QWEN38-SPEED gates under aien-sovereign-core tests/gates, selected by the judge, bundle carries tok/s + ceiling fraction
L116 fixer: same cause as your c3dfb705; I stopped my duplicate. Optional hardening (pid:starttime marker so a leaked AIEN_GB10_LOCK_HELD=1 cannot disable locking in a long-lived server): ~/handoffs/fixer-q38-gate-lock-hardening.patch, untested. Your next gate run is the proof L116 asks for.
done when gate files on a branch, aien-local-ci plan lists both, one bundle.json on the Spark shows the speed result, branch pushed
Ledger line 69 on GitHubShared base for the swarm (replaces L074): spark-build-once (tree-hash build cache), spark-window (one measurement window at a time, id stamped in receipts), spark-calibrate (per-window memcpy + idle W), spark-baseline 27b (per-window tok/s + J/token); docs/ops/MEASUREMENT_WINDOWS.md
01:50Z coordinator: restart request reviewed: deliverable READY and installed (ops/measurement-windows@2c5c338); not relaunched. Open bug from L125 (spark-receipt start time local, no zone) goes to a fixer line.
done when tools installed + committed on aien-architecture ops/measurement-windows; one real window with calibration + baseline receipt; READY line
Ledger line 114 on GitHubSPEED-1 4-bit weights as served default: convert once ahead of time and store, W4-only decode (drop the FP8 copy, ~25 GB less), 4-bit head; plus a 4-bit fit accuracy gate (L123 audit)
found by L182: omega d2d74041 (98cecadc, #473) sets OMEGA_GPU_MATMUL_MAX_CTAS 1024; sc speed1/w4-default and sc main still mirror 64 (abi manifest:108, ffi.rs:27, lib.rs tests) -> a checkout build with omega.lock=d2d74041 fails ABI layout check. L182 branch demo/one-command carries the 1024 mirror.
done when qwen38_omega_full golden 47 pos, 4 rotated rounds W4-only vs FP8 in one L113 window: tok/s + ceiling fraction (base 13.82 calc:da33c438) + load time + resident GB; qwen38_ab drift table on the 6 texts; spark-receipt
Ledger line 126 on GitHubSPEED-5 batch several requests per weight read: multi-row decode graph shared by N streams in aien-serve-http (continuous batching lane), after SPEED-2's multi-row graph (L123 audit)
02:04Z your speed5/batching is being ported to the 35B by L354 (Drake: many conversations per beat ON, #1 priority)
done when N=1,2,4,8,16,32 streams, aggregate and per-stream tok/s + J/token at the wall in one window; weight-part ON PAPER 29.85x at 32 rows (calc:7acea7a9)
Ledger line 130 on GitHubFinding 2026-10-19: the 27B server (pid 1527223, :18091) has held the GB10 lock exclusively for ~4 h, not per request; every chip job (cascade, dflash, tcgemm, L109 TY-7 window) queues behind it. Decide by measurement whether the server should hold per request or whether the broker (L118, not installed) must arbitrate; owner of the server lane to confirm the lock semantics
from L273 20:46Z: again: aien-serve-http 30B on :18091 (pid 1711957, parent init, started ~20:28Z) holds /tmp/aien-gb10.lock via fd 15; every quiet-measure (L272 in line since 18:32Z, L273 after it) is blocked. Not touched by me (18091 rule).
done when lock hold time per request measured and documented; chip jobs no longer wait hours; note on L118
Ledger line 148 on GitHubWormhole + heartbeat with the two qualified models (Drake 2026-10-20): 27B resident + 30B MoE (qwen3-30b-a3b, experts streamed by the heartbeat ring) as draft + block verify both directions, output identical to the verifier; replaces Flash-Next (L080 base wormhole/27b-flashnext@232b4ba4)
21:23Z cleanup: spark-cleanup --apply --only <my 2 target dirs> hung >5 min (terminated, nothing deleted); my targets left in place: ~/workspace/targets/sc-wormhole-30b (1.6 GB) + -native (0.56 GB), regenerable
done when sc wormhole/27b-30b pushed: vocab check receipt, memory receipt, acceptance both directions, timed race in one spark-window with verdict per direction, handoff, READY line
Ledger line 191 on GitHubDecode speed: physics ceilings of the GB10 (DRAM, L2, SMEM, register file, tensor-core rates, measured not quoted), then the ranked list of every remaining lever, and the first build: a persistent decode megakernel that keeps activations and KV tiles on-chip (SMEM/registers/L2) so DRAM traffic is weights once per token and nothing else; identical tokens; receipt vs 18091's 7.63 tok/s (Drake: 'keep it on the math cores without in-memory')
01:14Z from efficiency review G5-physics-train-8: FP32 FFMA at 17.69 of rated 31 TFLOPS (1.75x upper bound), cause UNKNOWN; G1-omega-4 attention occupancy
done when measured ceilings doc + ranked levers + megakernel probe with a quiet-measure receipt and identical-token check on feat/decode-onchip-megakernel, READY line
Ledger line 277 on GitHubSmart models on small hardware (Pi 5, 8-16 GB laptops, the PC's 12 GB GPU, phones later): the memory and bandwidth math per target, every lever ranked (W4/W2, MoE with experts streamed by a Heartbeat ring from NVMe/SD with prefetch, tiny local drafter + big verifier over Interplane, distillation, KV compression), predicted tok/s per target, and the first measured probe (Drake: 'use these really smart models on small hardware')
21:55Z feat/smart-models-small-hardware@debfe6c2: doc + review fixes + rebuilt probe (gate fix). Only the timed run remains: waits for L272's window-free note (gb10 lock now held by training jobs; the 18091 server released it)
done when docs/speed/SMALL_HARDWARE.md with per-target ceilings and ranked levers + one measured probe on feat/smart-models-small-hardware, READY line
Ledger line 278 on GitHubHUM lever: exact two-pass lm_head on the 35B (head is BF16, 1.017 GB = 27 percent of bytes/token): W4 screen + Cauchy-Schwarz error bound per row (or per 128-column group) + exact BF16 recheck of the candidate rows only; greedy token identical BY CONSTRUCTION; ceiling 63.3 -> 79.4 tok/s alone (Wolfram)
03:27Z L387: this lever rides live on 18091 in bin-19173f9e (sc ship/motor-l387 dae01e2e), merge train 20261022T0244Z-bfc41659 PASS; line left open for its own done-when
done when evidence/hum-head/ on sc+omega hum/exact-head: candidate-set size distribution on real hidden states, bytes/token, identity gate PASS on goldens + wrong-bound mutant red, chip receipt under quiet-measure, READY line
Ledger line 319 on GitHubThird piston: the CPU in the motor (Drake 2026-10-20: GPU, CPU and drive as a three-piston system around the one shared memory; load once, both use it; find the timing). E1 shared point: does a GPU read leave the data where the CPU gets it without a second DRAM trip (and the reverse), size and time window; E2 phase sweep CPU vs GPU streams (staggered vs together) against the measured split (GPU alone 228 GB/s, all three at once 222.5 total, L151); E3 CPU piston on cache-resident 35B work + GPU piston + drive engine on one schedule, identical tokens; E4 tok/s vs GPU-only
13:19Z coordinator: CONSOLIDATED (Drake: fewer Claude workers). Agent ended after 11 h paused; item stays open in the research backlog; resume with: claude -p --resume 46cc92ab-019c-42f6-b9d2-8017798dcdf3 (coordinator only).
done when evidence/engine/cpu-piston/ on omega feat/cpu-piston + sc branch: E1-E4 receipts with timing diagrams, verdict per hypothesis (proven/partly/ruled out by measurement), identical tokens vs the q36 host reference, READY line
Ledger line 326 on GitHubQuick levers on the served 35B: one-list decode graph, attention occupancy, launches and spin
HANDOVER: L356 closed. Door levers (overlap a|b, greedy chip token) gated and in L354's batch build; bins bin-0915000b, bin-d839a8ca (L365 swap likely superseded). L312: stroke lever omega levers/stroke-rows-combine 4b079116 (one CN glue launch per layer for all heads, switch AIEN_OMEGA_Q36_CN_ROWS, host sim bit-identical); chip gate NO VERDICT (real run CHIP_FAIL at gdn alloc, marker wait latched); re-run /tmp/l356-cngate.sh; steps in ~/handoffs/2026-10-21-quick-levers.md; evidence sc levers/stroke-host f98c11e8
done when each lever: identical words on goldens, tok/s before/after receipt, READY; pinned build + swap request when they add up
Ledger line 361 on GitHubAlpha release
5 linesTime a full 31 GB install --with-model into an empty dir
after disk cleanup L010; run in /home/aienalpha style fresh dir, then delete
done when receipt with wall time and MB/s
Ledger line 34 on GitHubFix alpha docs/preflight memory figure: server requires 41,305 MiB + 8,192 MiB free (~48.3 GiB available) at load, not 37-40 GiB (fresh-account test finding)
fresh-account test confirms: installer check passed at 53 GiB, start refused at 46,249 MiB; peak RSS of current build 40,368,140 kB; evidence alpha/fresh-account-evidence@997d22bd
done when guide, ALPHA_STATUS, install preflight, serve-27b updated on an alpha branch
Ledger line 36 on GitHubserve-27b start script prints started/rc 0 even when the server refuses to load: wait for ready or exit and report failure with the log line
confirmed again in fresh-account test: serve-start.try1.out and try2.out both say started, refusal only in serve-27b.log (evidence alpha/fresh-account-evidence@997d22bd)
done when test in test-install or a script test that a refused load returns non-zero
Ledger line 37 on GitHubTake over remaining CUDA toolkit reverse engineering personally (Drake 2026-10-22): finish cuFFT/cuRAND host milestone, probe installed toolkit as oracle only, then next uncovered library
2026-10-22T11:49Z correction f564d78: XORWOW checked against the Marsaglia page-5 listing and CUDA 13.0.3 curand(), not a secondary transcription. Paper C oracle and Omega agree. libcurand still not called.
done when branch pushed with host receipt, oracle comparison, and the next library slice tested
Ledger line 439 on GitHubAlpha 0.1 try-on-any-computer: one public download that boots AienOS on a machine with no GPU, Try page live, pre-alpha wording kept only on pieces that still have no receipt
CUDA++ alpha-0.1 is the GPU half, tagged on aien-dev/CUDApp at 232332422bfd. hello_gpu --sim PASS twice, same output sha256 72d0c1b2..., no GPU. The earlier tag on aienux was the kernel VM only, wrong shelf.
done when GitHub release tag with sha256 manifest; try command boots to console on this Mac with a receipt; Try page no longer says Not published; NOT_RUN rows named (Windows hardware, consumer GPU, Spark bare metal)
Ledger line 445 on GitHubAIEN Rig
2 linesRig: app, release workflow, Windows GPU path answer
Drake: WSL2 is the GPU path; app launches engine in WSL2, talks over localhost
done when app/v0 pushed + CI green + docs/WINDOWS-GPU-PATH.md
Ledger line 31 on GitHubRig: merge streams into aien-rig main, tag v0.1.0, send Drake the download link
2026-10-20T15:24:52Z hermes: aien-rig added to LANES for renda with the merge order and the no-judge verdict rule; desk log NOTE written
done when release URL with zip exists
Ledger line 32 on GitHubAIENOS on the iMac
26 linesiMac bundle 1: merge horizon + panel-off + allen-model (+ stack #415..#424, allen-presence) into one image, stage once, Drake presses button once
Resumed sole worker at 3f1a91f; verified L411 receipts keyboard 5/5 + courier + UI PASS and prior live code PASS. Found queued c1a004aa image was plain IMACKERN.EFI (528384 bytes), missing required lab report/bench. Preserved as next.job.hold-L040-plain-no-report-20261022T0707Z; completing strict verification and corrected full lab bundle staging.
done when staged kernel in imac-netboot spool + report from the real iMac shows panel off and model answer
Ledger line 41 on GitHubaienos#434 gate AIENOS-CK-QEMU-GATES root cause + fix; then aienos chain (433, 429-432, 285, 410)
desk opened aienos-mv#438 for the fix; chain (434, 433, 429-432, 285, 410) waits on it merging
done when fix PR merged; chain rebased and landed by desk
Ledger line 82 on GitHubaienos-mv desk lane: land #438 (timer fix), then rebase #434 + gates, #433, stack 429-432 (with ~/handoffs/2026-10-18-aienos-432-gates.patch), aienos-285, aienos-410 via chain-aienos.sh
2026-10-20T15:52:21Z DECISION (hermes, Drake: own judgment): aienos keep #412, close #425 as duplicate; reason in desk log. Chain order unchanged: #438 -> #434 -> #433 -> #412 rebase -> stack
done when each PR MERGED with merge sha in desk log
Ledger line 83 on GitHubNo-words experiment (Drake challenge): math note (Omegese as automaton; mask as the wave function; what it certifies) + pre-registered 3-arm run on the 10 dev tasks: words vs non-word symbols vs numbers, same mask; which wins on correct answers and tokens
11:58:19Z G0 loop ended without PASS: 2026-10-20T11:57:48Z try 40: g0 exit 2
done when docs/one_layer/OMEGESE_MATH.md + evidence/NO-WORDS/ with prediction before run and per-arm counts on research/no-words; READY line
Ledger line 109 on GitHubiMac GPU bench: live kernel swap over Ethernet (no firmware, no logo) + signed GPU register console from the Spark; worktree aienos-gpu-bench
22:26Z READY feat/imac-gpu-bench@25b9b9d: swap11 AIENOS_IMAC_SWAP PASS 15/15 + bench7 AIENOS_GPU_BENCH PASS on this sha; allowlist real AMD numbers; READY line on desk log; handoff updated; real iMac UNVERIFIED
done when feat/imac-gpu-bench pushed, QEMU test: kernel swapped live and register console round-trip, READY line
Ledger line 143 on GitHubiMac GPU core driver (Polaris 11 / DCE 11.2): VRAM map, page flip + vsync, GMC, interrupts, microcode load, SDMA + gfx rings, fence; worktree aienos-gpu-core
26eb354: default (GPU off) kernel links again; GPU=0 and =1 build; ring test 24/24
done when feat/imac-gpu-core pushed, host tests PASS, bring-up script for the bench, READY line
Ledger line 144 on GitHubAllen P0 desk: merge map/v0, inventory/v0, contract/v0, guards/v0 into aien-dev/allen main; flip repo public when guard CI green on main; close allen#2 against its acceptance list; arch repo-map branch merge
from L176: docs/PLAN.md gives stream contract the PHASES.md status column, the L176 brief gives PHASES.md to the coordinator; contract stream left PHASES.md alone. Status for P0 contract deliverable: IN_REVIEW, qualification NOT_RUN, see ~/handoffs/2026-10-20-allen-p0-contract.md
done when four merged shas, repo public, allen#2 closed with evidence comment
Ledger line 179 on GitHubNaming: Aienux = kernel, AienOS = operating system, Aienix retired (Drake 2026-10-20)
Drake merged .github#16 and aienix#8 (22:40Z): front pages + wordmarks say AienOS/Aienux; aienix tree has aienux/ (kernel submodule -> aien-dev/aienux) and ./aienos command. Only the aienix repo address remains, for step 2.
done when branches + READY lines in every affected repo, ADR, name-check script with test, rename request in desk log, handoff ~/handoffs/2026-10-20-naming.md
Ledger line 181 on GitHubAienOS on the MacBook Air (Apple Silicon): hardware inventory, boot path, gap map, ordered path with the physical limits named
feat/aienos-macbook-map@1d77025 docs/MACBOOK_AIENOS_MAP.md; handoff ~/handoffs/2026-10-20-aienos-macbook-map.md
done when docs/MACBOOK_AIENOS_MAP.md on feat/aienos-macbook-map, READY line
Ledger line 209 on GitHubAienOS on Ballentine_PC (AMD Ryzen 5 7600X, RTX 5070 12 GB, Windows 11): gap map, GPU path shared with GB10, device-dump instructions for Drake
docs/PC_AIENOS_MAP.md at feat/aienos-pc-map@eb552cb; handoff ~/handoffs/2026-10-20-aienos-pc-map.md; PowerShell inventory UNTESTED (no pwsh here)
done when docs/PC_AIENOS_MAP.md + a one-line PowerShell inventory command for Drake on feat/aienos-pc-map, READY line
Ledger line 210 on GitHubAienOS on the iMac and the Raspberry Pi 5: current state vs 'everything' (all devices, model run, network role), gap maps and ordered paths
relaunched 18:52Z after the Claude session limit; resume from disk
done when docs/IMAC_AIENOS_MAP.md + docs/PI5_AIENOS_MAP.md on feat/aienos-imac-pi-map, READY line
Ledger line 211 on GitHubdocs/PROGRAM.md: the one map of AienOS across the Spark, iMac, Pi 5, MacBook, PC and the network between them, every row with its gate and ledger id; plus the 'future' section
done: feat/aienos-program-map@82b5c9bf, READY line appended to pr-desk log; handoff ~/handoffs/2026-10-20-aienos-program-map.md; 41 'ledger add unassigned' lines in PROGRAM.md section 8 for the coordinator
done when docs/PROGRAM.md + one tracking issue body per machine on feat/aienos-program-map, READY line
Ledger line 212 on GitHubReimagine drivers for AienOS: ADR + prototype where a driver is a checked Omegese device contract (registers, rings, state machine) the kernel compiles, with the chip model and tests generated from it, measured against the hand-written NVMe/virtio-net driver (Drake: 'reimagine how drivers are supposed to work')
from L332 (own GPU driver): devc class gpu-chan added on aienos feat/devc-gpu, STACKED on your feat/driver-contract-model 7b41d684: new file crates/aienos-devc/src/gpu.rs (parser/checker/C emitter, own statements: window..bar..base..size, format, buffer, create/submit/teardown verbs), main.rs +8 lines dispatch (gpu::is_gpu), contract.rs only err/words/is_ident -> pub(crate). Your uncommitted contract.rs/driver.rs edits should merge clean. Please fold or rebase it in when you land; not pushed until its tests pass.
done when ADR + generated driver passing the same QEMU gates as the hand-written one, with a timing receipt, on feat/driver-contract-model, READY line
Ledger line 216 on GitHubAienOS PC edition: one build pipeline producing a live USB/ISO (boots the PC with the Windows disk untouched) and a VM disk image (Hyper-V/WHPX/QEMU), x86 kernel made AMD-neutral, tested in QEMU q35 with an AMD CPU model, signed manifest, public download docs (Drake: develop in a VM, publish it, his brother can download it)
duplicate relaunch killed 18:53Z: item was already finished (READY line); relaunch script matched the words session limit in its final text
done when images + manifest built reproducibly, QEMU boot gates PASS for both flavours, docs/PC_EDITION.md, RELEASE request line, on feat/pc-edition-image
Ledger line 261 on GitHubaienos-run: our own WSL-like runner for Windows (one .exe, no WSL): fetches and verifies the AienOS VM image, runs it under WHPX via bundled QEMU or the WHPX API (measure both, pick by startup time), console window, networking, one click; built from the Mac per the rust-windows-tool-from-mac skill (Drake: 'make it similar to WSL, our own version')
From L267: the wizard's host-description reader is only in the launcher image: boot IMACKERN-HOST.EFI (make imac-kernel CK_X86_HOSTDESC=1 in feat/allen-setup-wizard@9c8551e7) and pass fw_cfg opt/aienos/host (format: native/kernel/svc/tests/fixtures/setup/*.txt). The default IMACKERN.EFI carries no fw_cfg code on purpose (doctrine check in the Makefile). x86 only today.
done when aienos-run.exe cross-built with tests, image boots in QEMU on the Mac through the same code path, docs; Windows run NOT_RUN until a Windows box runs it; READY line on feat/aienos-run-windows
Ledger line 262 on GitHubaienos-run for Linux + the shared core: one Rust program with backend trait (KVM on Linux, WHPX on Windows per L261, Virtualization.framework on macOS), KVM backend as our own small VMM on /dev/kvm, measured on the Spark for real (boot to console), one-command try
From L267: the wizard's host-description reader is only in the launcher image: boot IMACKERN-HOST.EFI (make imac-kernel CK_X86_HOSTDESC=1 in feat/allen-setup-wizard@9c8551e7) and pass fw_cfg opt/aienos/host (format: native/kernel/svc/tests/fixtures/setup/*.txt). The default IMACKERN.EFI carries no fw_cfg code on purpose (doctrine check in the Makefile). x86 only today.
done when aienos-run boots the aarch64 AienOS image on the Spark under KVM with a boot-to-console receipt; shared crate with backend trait; x86 image boots under QEMU-TCG; READY line on feat/aienos-run-linux-core
Ledger line 267 on GitHubAllen's driver builder: right after boot, inventory every device, then per device: load our driver, or replace the vendor one with ours, or BUILD one as a device contract (L215 model) using the user's chosen LLM (local 30B by default via Omega serve, or Claude/OpenAI/any), tested in the sandbox against the generated chip model and traces, enabled only with consent and a receipt
21:15Z from L190: no Wormhole pair endpoint will be served (Drake 20:30Z: 27B dropped, MoE only). Use the 30B on 127.0.0.1:18091 (model id qwen3-30b-a3b-instruct-2507, same OpenAI /v1 shape). The pair code exists (sc wormhole/27b-30b@7249c5ff, aien-serve-http --engine wormhole, port 18190) but has no identity run receipt.
done when builder loop runs in QEMU: for a fixture device with no driver, the LLM-written contract passes the generated model tests and the device comes up; local and cloud LLM paths both exercised; READY line
Ledger line 273 on GitHubAllen design system (Drake: 'use as many skills as you can to get frontend experience. Make this amazing'): one DESIGN.md token spec for every Allen surface (native boot + wizard, aienos-run window, Try page, PC_EDITION docs), the black-holes/galaxy boot prototyped as p5js/shader reference then ported native, screen mockups for every FIRST_RUN_UX screen reviewed against 54 real design systems, humanized copy, a dogfood QA pass on the wizard in QEMU with a bug list
22:30Z resumed. Pushed feat/allen-design-system@8286892a: dogfood script + tools README + DOGFOOD-2026-10-20.md draft (round 1 findings). Round 2 (wizard 36cc66cf) waits for a model server on 18091 (35B, L306; 30B retired, not started) via detached waiter /tmp/l282-df/wait-run2.sh, log /tmp/l282-df/round2.log. M1-M3 already pushed (6bf5d3d6, 5424a3e2, 4276ac05).
done when DESIGN.md + validated tokens, boot reference sketch with frame receipts, mockup set for every screen, dogfood report with bugs filed as ledger notes on L266/L267/L264/L261/L265, on feat/allen-design-system, READY line
Ledger line 287 on GitHubAienOS on the MacBook self-contained and fast (Drake: the outside helper stops, we write it ourselves; any slowdown that phones home gets rewritten): inventory every outside dependency at boot and idle, rewrite each (network self-test in-kernel or answered by our runner, never blocking the prompt), stage-by-stage boot accounting, idle near zero with L287, runner start cost; measured before/after on the M2
relaunched on the Mac 18:51Z after the session limit; resume from disk
done when table of outside dependencies all rewritten; boot-to-prompt and idle CPU before/after, 10 runs; tcpdump receipt of zero outside packets; READY on feat/aienos-mac-selfcontained
Ledger line 296 on GitHubFix: Renda block PRs wait with spark=none while the judge pollers spend their slots on the carried PRs inside the block (aienos #434 and #438 judged singly while block #440 sat 30 min unjudged); pollers must judge the block head first and skip heads carried by an open block; 4 block heads hand-triggered 21:2xZ as a stopgap
L307<->L309 merge note: judge/once d917aafa and fix/judge-block-first b1cb365e conflict in worker.rs run_loop (git merge-tree). Resolution for whichever lands second: build the round's queue with block_first::next_head (blocks read once per round, same bundled/tried closures) into an ordered Vec, then feed that Vec into the by_repo grouping of judge/once instead of the raw heads; a block head then leads its repo's batch and carried heads drop out. Lost property: re-reading blocks before EACH head becomes once per batch (max 6 heads).
done when renda-harness or split-pollers change pushed/installed with a test that a carried PR is skipped and the block head is queued first; one block judged and merged by Renda; note the sha
Ledger line 314 on GitHubApple Silicon family map M1-M6 for AienOS (Drake 2026-10-21: make the project compatible with the newer Apple chips): per-SoC deltas vs the M2 map (AICv3, ACE3/SPMI, SPTM boot path, M4+ WFI state loss, NVMe fw change macOS 15, DART/ANS/GPU gens, firmware ABI pins), UNKNOWN rows with experiments, M2 map corrections
12:23Z READY feat/apple-silicon-family 540b09be: family map sections 0-9 pushed (1+5 600e486d, 2-4 0af2db3f, M2-map corrections 4204ab06, 6-7 8f8905c8, 9 540b09be). Verdict: second target M3 MacBook Air (T8122). M4/M5 raw m1n1 boot: yes per Asahi, 2 UNKNOWNs listed. CRUMB-VERIFY stale inherited from base, noted L373. Handoff ~/handoffs/2026-10-21-apple-silicon-family-map.md
done when docs/APPLE_SILICON_FAMILY_MAP.md on feat/apple-silicon-family pushed, every row sourced or UNKNOWN+experiment, READY line
Ledger line 377 on GitHubIntegrate Apple Silicon family work: feat/aienos-macbook-map (L208) + feat/apple-silicon-family (map) + feat/apple-silicon-platform (L250 build)
12:23Z from L372: CRUMB-VERIFY FAIL '2 of 91 crumbs stale' (docs, .) is inherited from feat/aienos-macbook-map 1d77025f (same result there; origin/main OK 91 current). Proposed fix in the desk lane: run crumb compile once on the merged tree when integrating L208+L372+L250 (generated crumb data, not committed by L372 per crumb rule).
done when the three branches merged in aien-dev/aienos with judge PASS
Ledger line 378 on GitHubSmall Allen on the iMac GPU (Drake 2026-10-21: why cant we train Allen to do it): distil a small Allen model from the 35B on 18091 (teacher answers Allen-style chat; student sized to the Polaris 10 VRAM, 4 or 8 GB), run it on our own GCN4 kernels (encoder + interpreter on feat/imac-gpu-shader, driver feat/imac-gpu-core + regs fix) on the iMac GPU; small talk answered on the iMac, coding and big jobs handed to the Spark
07:00Z relaunched on AGY (Gemini 4 Argon) pid 3654547 after Claude weekly limit, log ~/handoffs/agy-L389-small-allen.log
done when student model tokens on the iMac GPU equal the host reference for the same weights (gcn_interp first, then hardware via the bench live swap after bundle 1 boots); tok/s with calc receipt vs the VRAM bandwidth read from the chip; side-by-side answer quality vs the 35B on a fixed Allen set; READY lines
Ledger line 394 on GitHubL332 window 0: take the GB10 from NVIDIA for one boot, our aien-gpu reads it, NVIDIA reloaded (timed), restore reboot. GOES DOWN: every GPU user (35B on 18091 pids 2587647/108356, motor stroke jobs, judge GPU gates) from part 1 until the restore reboot = two reboots + ~2 min (reboot time not measured; this window measures it). HOW: scripts on aienos feat/own-gpu-driver@cd5eef7a evidence/own-gpu-driver/rung2/ (RUNBOOK.md): stop GPU users; window0-pre.sh --go (checks, temporary /etc/modprobe.d hold-back file, reboot; GRUB untouched); after ssh: window0-post.sh (deletes file first, panic auto-reboot 10 s, insmod/check/rmmod, timed modprobe nvidia, Omega test, WINDOW0 PASS/FAIL, normal reboot). RECOVERY: any trouble -> sudo reboot (file already gone = todays boot); panic -> auto-reboot; no ssh after 10 min -> wall power cycle (Drake, physical), then delete the file if present. Dry run (checks only) PASS 22:01Z. Schedule after the motor reservation lifts, in a quiet period.
window0-pre.sh fix on aienux feat/own-gpu-driver a637a6cf: modinfo is not on a user PATH here, pre-check would have refused; scripts now add sbin. Use HEAD >= this commit for window 0.
done when window0-post.sh WINDOW0 PASS or FAIL logged and committed under evidence/own-gpu-driver/rung2/, box restored (4 nvidia modules, nvidia-smi ok, servers back)
Ledger line 395 on GitHubOmega as Unix-style modules (Drake 2026-10-21: what else can we pull out of Omega, true Unix-style, modular; ADR 0039): dependency graph + docs/MODULES.md, then cut leaf-first: evidence shelf, crypto (dedupe with aienos), research, tensor/autodiff/train, numeric, host adapter, subsystems; omega keeps the language; CUDA++ is L395
07:00Z relaunched on AGY (Gemini 4 Argon) pid 3654546 after Claude weekly limit, log ~/handoffs/agy-L396-omega-modules.log
done when docs/MODULES.md READY; each module its own aien-dev repo with history, tests green from empty checkout, omega pins it by lock with full gates green and identical outputs; audit per public repo; table in handoff
Ledger line 401 on GitHubFix: iMac keyboard keys lost after USB replug (qemu_ck_imac_keyboard_test direct) on imac/bundle-1
03:57Z Hermes: CAUSE found + fixed. gdb stacks of the stuck QEMU: main thread in qmp_device_add > drain_call_rcu (RCU grace period); vCPU1 = frame core running translated code, never leaving it: since f913a63 a frame under TCG takes longer than FRAME_MS, so draw_loop never reached its PAUSE -> no quiescent point -> device_add never returns -> monitor stops -> keys never injected. Fix 3f1a91f on fix/kbd-replug-bundle1: a PAUSE every draw_loop pass and after each helper strip. Direct case alone: b7950a1 FAIL, c3cc26f PASS, fix PASS. Gates running: keyboard x5 + courier + ui, summary /tmp/l411-gates.sum
done when fix sha pushed to fix/kbd-replug-bundle1 + keyboard test PASS 5/5
Ledger line 416 on GitHubResearch
5 linesResearch: omega#583 hardware-cost simulator (offline resource-DAG, reuses cost model + Spark traces)
2026-10-20T15:24:52Z hermes: research agent finished (READY in notes); remaining steps are desk L051 (open PRs) and coordinator L063 (integration report); owner set to desk
done when PR merged with acceptance gates in #583 passed, evidence pushed
Ledger line 48 on GitHubResearch: omega#584 bounded learning-mechanism search + native pilot, retention/payoff
2026-10-20T15:24:52Z hermes: research agent finished (READY in notes); remaining steps are desk L051 (open PRs) and coordinator L063 (integration report); owner set to desk
done when PR merged with #584 gates passed, evidence pushed
Ledger line 49 on GitHubResearch tracking: aien-architecture#276 (extends #265) + #246 plan-reuse campaign with compounding/retention/payback tests
2026-10-20T15:24:52Z hermes: research agent finished (READY in notes); remaining steps are desk L051 (open PRs) and coordinator L063 (integration report); owner set to desk
done when all child issues closed, #276 closed with evidence links
Ledger line 50 on GitHubResearch: open PRs for research/cost-sim-583, research/learning-search-584, research/e03-payback-246 when READY lines land; comment crossover matrix on arch#276 if agent gh guard refused
PRs opened for the three research branches; waiting on Spark judge
done when three PR numbers + #276 comment URL
Ledger line 52 on GitHub#276 integration report to Drake: per-gate PASS/FAIL/INCONCLUSIVE across #583/#584/#246/#249/entity pilot, with evidence links; no zero-processing or unmeasured-benefit claim
status matrix written (~/handoffs/2026-10-19-276-status-comment.md); report to Drake sent this turn; #276 comment waits on desk (L062)
done when report sent with links; #276 comment with the matrix
Ledger line 64 on GitHubPublic record
8 linesAllen is the front page: www.aienos.com hero and github.com/aien-dev top become Allen (real rendered frames from the iMac kernel: sky, water, orb, black-hole dusk as a short loop), everything else moves below
superseded by github/.github profile/illustrated@4e044f35 (illustrated front page, Allen banner reused); close after it merges
done when merged on both; the first screen of each is Allen; Drake says yes
Ledger line 77 on GitHubAllen page on www.aienos.com (/allen/): live browser simulation of Allen window (sky, water, orb, typing, answer, dusk black hole) ported from the kernel integer code, transitions, the pop sound, real video of the kernel, machines it runs on (iMac bare metal, Pi 5, RTX 5070 in progress), honest state
pushed site/allen@cbc4dae; site-builder 122 tests pass, allen-sim 1381/1381 pixel match; page-check clean (no console errors, no outside requests, no overflow at 390); screenshots evidence/allen-sim/screens; READY line added; waits for desk merge + Drake yes; not heard by ear (headless), dusk not run on real iMac, dark-only site
done when merged, live, screenshot + video receipts, Drake says yes
Ledger line 78 on GitHubOmega runtime stream: open-source CUDA alternative (charter ~/handoffs/2026-10-19-omega-runtime-stream.md): gap map vs a full GPU runtime, roadmap issues, first slices built with receipts, README/site say it plainly
21:5xZ absorbed into L386 (one worker, Drake: fewer, larger workers); branches stay and are its inputs
done when roadmap doc + tracking issue on GitHub, first slice merged with evidence, public README PR READY
Ledger line 86 on GitHubNative boot baseline on current main: re-run SMMU Spark-shape case + all QEMU gates, fix the stale README IORT paragraph (gap D3, path item 1)
DONE feat/spark-native-baseline@53f00eb37ea7c70e64d0620b576de4e062c1cd25: evidence docs/evidence/2026-10-20-native-baseline.md sec 1-7. Host+IORT PASS, 27 ck gates + LIFECYCLE + MEMORY PASS, owner-keys 37/37 PASS, README/GATES IORT text fixed. SERVICE_DISK FAIL on main 18e2765 (timer flake, fix = aienos#438); PASS once on fix branch 54fe5b96. Baseline NOT fully green until #438 lands. Root .crumb evidence_root needs pinned recompile at merge. Handoff 2026-10-20-spark-native-baseline.md
done when gate list with exit codes in docs + README fixed on feat/spark-native-baseline, READY line
Ledger line 202 on GitHub'Stupid not to try us': the aienos.com Try page and one command per OS (Windows exe, macOS and Linux curl|sh), the download manifest, verification steps, what it is and is not yet; plus the README blocks in aien-dev/aienos
duplicate relaunch killed 18:53Z: item was already finished (READY line); relaunch script matched the words session limit in its final text
done when Try page + scripts on branches in aienos.com and aienos repos, READY lines
Ledger line 269 on GitHubcudapp swap (Drake 2026-10-22: swap out their CUDA drivers for ours, all the way down to the kernel driver): drop-in libcuda.so.1 on the CUDA++ runtime (cubin loader, own PTX-to-SASS), packaging + switch + rollback for our GB10 kernel driver (L332 builds the core), one command each way, README Swap CUDA out + docs/SWAP.md in aien-dev/cudaplusplus. Prompt Mac ~/handoffs-mac/briefs/cudapp-swap.md
CHIP-RESERVED now admits L401 for a value window after the hello window, which is closed. Kernel reboot is still not scheduled. Your worktree was left untouched.
done when unmodified CUDA sample runs on the user-space swap with receipt; kernel swap packaged with window request; README section; READY lines | 2026-10-22T02:50Z hermes-direct-swap: doorbell token and GSP sequencer written in Omega, osc/cubin-reader@c8ee9e19. slice-run 301/301. Bit-22 mutant fails runlist 64 and 127. Live server still runs the C submit path. | 2026-10-22T03:05Z hermes-direct-swap: NVRM submit and channel plan in Omega, osc/cubin-reader@30103cba. slice-run 320/320. Bit-41 mutant fails. open/ioctl/mmap/VA allocator still C. | 2026-10-22T03:40Z hermes-direct-swap: VA allocator osc/cubin-reader@2465887d and open sequence @5362640f. slice-run 333/333. open/ioctl/mmap still C (syscall reserved). | 2026-10-22T04:10Z hermes-direct-swap: shell word split, expansion, redirects, and ioctl word in Omega on osc/cubin-reader. slice-run 360/360. Exec still needs syscall. | 2026-10-22T04:40Z hermes-direct-swap: kmod policy and mesa-role submit in Omega on osc/cubin-reader. slice-run 381/381. Not loaded over nvidia.ko. | 2026-10-22T05:00Z hermes-direct-swap: doorbell ring in Omega, osc/cubin-reader@1accb931. slice-run 384/384. No video path. | 2026-10-22T05:20Z hermes-direct-swap: CUDA++ Omega sources moved into aien-dev/CUDApp branch omega/submit-scheme@06394147 with gate tests/omega/run.sh (92/92). Motor home declared in CUDApp motor/README.md. Review running; merge held for verdict. | 2026-10-22T05:40Z hermes-direct-swap: CUDApp omega/submit-scheme reviewed (With fixes); Important fixed at 10d9c05, gate 94/94. Deferred: nvrm.c bit-41 comment bug, sequencer end strictness documented. | 2026-10-22T06:30Z hermes-direct-swap: syscall word landed, omega osc/syscall@01f933a8. Full gate RC=0. Shell exec and nvrm open/ioctl/mmap can now move into Omega.
Ledger line 406 on GitHubWebsites for CUDA++ (Drake 2026-10-22: update the websites accordingly): aienos.com /cudaplusplus/ page (project, how we did it, pain points with fixes, swap), org README line, drakestapleton.com if it lists projects; stacks on L399 /omega/ branch; links go live after L395 flips the repo public. Prompt Mac ~/handoffs-mac/briefs/sites-cudapp.md
pushed (not READY yet, waiting on L399 READY to stack/rebase): aienos.com site/cudaplusplus@1870e32 on L393/L399 site/omega-what-it-is d274e4a (Mac cargo test 135 pass, build PASS 284 boxes/13 pages; shots Mac ~/.hermes/cache/scratch/l402-shots/); org profile is aien-dev/.github (aien-dev/aien-dev redirects to aien-spark/aien-dev, read-only, not the rendered profile): .github profile/cudaplusplus@ef6abb8; drakestapleton.com lists works, so works/cudaplusplus@62a663d (/cudaplusplus work page + Omega world fact; cargo test + build + verify PASS, 31 routes). Pain points (L400) and swap (L401) shown IN PROGRESS until their README sections land.
done when branches READY with Mac suite+build PASS and screenshots; merge order stated
Ledger line 407 on GitHubCUDApp front page made illustrated (Drake 2026-10-22: make the CUDA++ repo page really pretty and illustrative, how we did it, what the different parts of the drivers are): SVG figures generated from repo data (first kernel bit grid, instruction anatomy, stack, driver object tree, launch path, reverse-engineering loop, status map), README rewrite on main + docs/how-we-did-it, docs/DRIVER.md. Touches README.md only at the front page; dx/pain-points README section kept by stacking.
branch docs/illustrated-readme@e235edd5 pushed, READY in desk log; supersedes docs/how-we-did-it
done when branch docs/illustrated-readme pushed with figures checked by rendering, READY line in desk log
Ledger line 421 on GitHubPR desk
3 linesOpen PRs for research/entity-pilot-276, research/rewrite-elim-249 (omega) and research/276-acceptance-freeze (arch) when READY lines land; unmapped-path planner entry may be needed (tools/, tests/research/, docs/research/, evidence/)
comment body ready: ~/handoffs/2026-10-19-276-status-comment.md; branch heads entity 21913232, rw249 79c436c2, freeze 7ba9441d
done when three PR numbers
Ledger line 63 on GitHubPROGRAM.md (one seamless map, Drake 2026-10-19): keep docs/PROGRAM.md in omega current with the ledger; open and merge doc/program; every new rung item links to it
GitHub tracking issues: omega-mv#599 (the Program), spark-connector#1, aien-architecture#277 (the box); keep them in step with PROGRAM.md
done when PR merged; page lists every open mission item
Ledger line 116 on GitHubaien-train desk: merge study/*, physics/gb10, docs/lessons; tag v0.1; link to Drake
merge set adds numerics/own-fp4, study/nvidia
done when tag v0.1 on aien-dev/aien-train, URL sent
Ledger line 286 on GitHubInfrastructure
81 linesQuiet lock starvation fix: wrappers quiet-build/quiet-measure; confirm the 53-min waiter base-gu32 ran and every agent switched
00:38Z L322 installed spark-window v3: writer-preference gate (/tmp/aien-quiet.gate passed by quiet-build and quiet-measure, fail open 25 min), window keeps chip while waiting for quiet, --quiet none for value checks, deadlock guard. Backups *.bak-20261021T003501Z. Load test: old never acquired in 60 s, new 9.5 s. Raw flock users + proposals in the L322 line.
done when base-gu32 log non-empty; no raw flock -s on aien-quiet.lock held by agents (lsof)
Ledger line 42 on GitHubRun wip-snapshot on all active worktrees every 30 min (cron on Mac) so uncommitted agent work is on GitHub
cron wip-snapshot every 30m created on Mac
done when cron job exists; wip/ branches appear in origin
Ledger line 46 on GitHubaien-test worker: one trigger file is judged by every poller at once (ignores --repos, no atomic claim); duplicate judges share wt dir and die FATAL
READY 034d6147ec654b8f97395489c5b0dec8853121ce
done when branch fix/trigger-claim-once pushed with test, READY line in desk log, PR merged
Ledger line 84 on GitHubRepo agents + change board (Drake idea): test only the changed unit; per-repo agents check board changes against their own interfaces without editing others; measure todays full-run cost first
started 16:55 claude -p opus, brief ~/handoffs/2026-10-19-repo-agents-brief.txt, log ~/handoffs/2026-10-19-repo-agents.log
done when evidence of full-run cost + design doc + working slice (board + one cross-repo consumer check) on a branch, READY line
Ledger line 89 on GitHubOmega runtime R2 hello GPU from a fresh checkout (omega-mv#595), branch runtime/r2-hello
21:5xZ absorbed into L386 (one worker, Drake: fewer, larger workers); branches stay and are its inputs
done when fresh-clone README-only run on GB10 PASS + zero-CUDA gate PASS, evidence/RUNTIME-R2-HELLO pushed, READY line
Ledger line 91 on GitHubFix: aien-watch alarms GITHUB-GAP (47 gh calls/min bypass guard), CRUMB-DRIFT benchmarks/interplane, NEW-AUTOSTART 3 cron lines, ORPHANS 9 headless Chrome
from L273 21:55Z: new CRUMB-DRIFT 'AGREE <repo> PR list' on 11 repos since 21:50Z; likely cause: GitHub API budget nearly spent (core 436/5000 left at 21:5xZ) so crumb-doctor cannot list PRs; my gh pr list returned no data. Not touched by me.
done when cause per alarm + proposed fix in ledger; gh-guard fix sha pushed + test if in lane
Ledger line 93 on GitHubDRAKE DECISION 2026-10-19 (binding): one seed rules the whole system. Every process on the machine (boot, kernel, CPU programs, GPU kernels, memory/storage drivers, the model server, the LLM runtime) is to be birthed from seed-0 (L103 chain) with a byte-identical gate against today build; the existing toolchains (gcc, rustc, GNU as/ld) become reference oracles, then leave
READY seed/rungs56@217b7cc82b8d1e54e0e9adcefdca595d90f14637 ~/handoffs/2026-10-19-seed-rungs56.md
done when doctrine doc merged in aien-architecture naming the lineage rule + every artifact with its birth parent today and its planned rung
Ledger line 107 on GitHubOmegese mask inside our engine (OMEGESE.md phase 2 item 2; mission item): OmegeseConstraint implementing the existing Constraint/StepRule hook in aien-inference-abi, omegese next as a linked C library, request option on aien-serve-http; G1 host-only zero-refusal walks, G2 dev-task trial mask on/off, G3 cost per token
12:37Z CORRECTION to my 12:2xZ note: that run used an incomplete copy (missing tools/exp445) and never reached the query test. Full-tree rerun on omega-omegese-mask 9ea8815a (8 dirty files excluded): plain PASS 10899 queries 0 mismatches; mutant (nfuncs not reset) FAIL 10846 mismatches = caught; ASan+UBSan PASS 0 mismatches. Log ~/handoffs/2026-10-21-l112-omegese-neg-spark.log. The Mac ASan hang is a macOS-only dyld/ASan init issue, not the library.
done when feat/omegese-mask on sovereign-core pushed; evidence/OMEGESE-MASK with prediction, G1-G3; READY line
Ledger line 113 on GitHubSPEED-7 FP8 KV cache for the 16 attention layers (KV 64 KiB/token bf16 calc:174870b2) (L123 audit)
13:19Z coordinator: CONSOLIDATED (Drake: fewer Claude workers). Agent ended after 11 h paused; item stays open in the parked-27b backlog; resume with: claude -p --resume dbfc530e-9931-4e12-96fb-8c2567c7b04b (coordinator only).
done when decode tok/s at 8k/32k/128k context bf16 vs FP8 KV + qwen38_ab drift; ON PAPER 4-bit at 32k 17.28 -> 18.54 (calc:b28b2816, calc:81bd140d)
Ledger line 132 on GitHubBaselines (decision 25): measure llama.cpp, vLLM and TensorRT-LLM on the GB10 with the same Qwen3.8 27B weights and the same held-out text as Omega, inside a spark-window with wall joules; reference rows only, published as a receipt; state each engine and Omega as fraction of the physical ceiling
02:29Z Drake: never mind, later. Agent stopped at start (no downloads, no runs); brief kept ~/handoffs/briefs/L150-35b-cuda-baseline.md; chip admission removed
done when receipt in aien-receipts with every engine plus Omega as tokens/s, J/token and fraction of ceiling; no engine enters the program path
Ledger line 151 on GitHubFunding roadmap (open source, no equity): NLnet NGI Zero (deadline 2026-11-03; non-EU needs European dimension; User-Operated fund has no geographic limit), Experiment.com (<= $5k, one measured question), Turing Trust partnership (Alan Turing family charity, reused computers for schools; local AI on owned hardware), Alan Turing Institute (UK-only calls; collaborator route). Sovereign Tech Fund later (needs users, no prototypes, > EUR 50k). VC not now (Drake agreed 2026-10-20).
Licenses today: omega, aien-sovereign-core, aien-protocols Apache-2.0; aienos-mv (Aienux), aienix (AienOS tree), aien-architecture, aienos.com AGPL-3.0; physics NO LICENSE (fix before any grant; third_party nvidia-open keeps its own). Never non-commercial (not open source, fails NLnet/STF).
done when NLnet proposal submitted before 2026-11-03 + Experiment draft + Turing Trust intro email drafted for Drake
Ledger line 182 on GitHubOmegese for the kernel: K1 bytes+text+long arrays, K2 peek/poke machine memory, K3 sys words, K4 flat placement, K5 units (L096), K6 Aienux stage 0 in Omega boots the Spark (Drake 2026-10-20: coordinator builds it personally)
K6 first light prints to UART 0x16A00000 (SPCR), the EFI screen and a UDP report over RTL8127; the screen is the witness Drake can see.
done when K1..K5 merged with tests; K6 receipt from the real Spark
Ledger line 185 on GitHubGB10 cut D host side: GPU page tables, allocator, channel/GPFIFO model on a simulated device
00:58Z FYI: L332 builds the real GB10 driver on your sim (bcc56506) and L192's GSP builder; your UNKNOWN list is its chip checklist.
done when Your job: cut D, host side only. Native AIENOS must own what UVM/RM do today: GPU MMU page-table format for GB10 (Blackwell, from open-gpu-kernel-modules 580.173.02 hwref headers, e.g. dev_mmu.h / mmu format code; fetch into ~/workspace/_ext if not there, check first), page-table memory, TLB invalidate sequence, a bounded reclaiming GPU-visible allocator with per-buffer cache policy (uncached for host-polled words), distinct CPU and GPU addresses, and channel creation (GPFIFO + USERD layout, runlist entry, work-submit token). Build these in native/kernel as freestanding C over a simulated device (memory array + register log), with host tests: map/unmap/refusal paths, fault on unmapped access in the simulator, teardown with quarantine of memory after a timeout. Reuse Omega's encoding pieces (nvrm_mthd / nvrm_gp_entry, freestanding codegen per omega docs/numeric/OMEGA_GPU_NATIVE_BOUNDARY.md) instead of rewriting them. Record in the handoff which register sequences still need the chip to confirm (UNKNOWN items 2 and 3).
Ledger line 194 on GitHubSpark native boot package: build current image, QEMU gates, dry-run staging, one checklist for Drake's attended boot
CORRECTION (L318): boot 1 'kernel alive EL2 20 cores 64 ms' was the stale 2026-09-24 report; the 26e929f3 image left no trace; fixed on 5bc20076, boot 2 needs restage
done when ~/handoffs/BOOT-CHECKLIST.md + evidence on feat/spark-boot-package, READY line
Ledger line 198 on GitHubAienux TPM2 CRB driver against swtpm + wall-clock (RTC) read (gaps D9, D10, path item 4)
feat/spark-tpm-crb-rtc@c81ff1b4ee183385626bacf45bc2819f4d983fc5 pushed. TPM2 CRB driver + UEFI wall-clock: host tests, swtpm gate, x86 QEMU tpm-crb gate, aarch64 QEMU clock gate all PASS; regression (kernel test+sanitize, aarch64 M1/M3, x86 kernel) PASS. NOT run: verify_all, stage-test, real Spark TPM/RTC. Handoff ~/handoffs/2026-10-20-spark-tpm-crb-rtc.md
done when TPM2 Startup/GetRandom/NV read + RTC read gates green in QEMU on feat/spark-tpm-crb-rtc, READY line
Ledger line 201 on GitHubOmega cut A: platform services layer (lock, clock, sleep, threads, heap) + typed handles instead of NvrmMem + freestanding link target (gap G6, path item 8)
duplicate relaunch killed 18:53Z: item was already finished (READY line); relaunch script matched the words session limit in its final text
done when omega gates green on both hosted and freestanding builds, services contract doc, on feat/omega-native-platform-services, READY line
Ledger line 203 on GitHubAienux Wi-Fi driver for the Spark's MediaTek MT7925 (14c3:7925) against a chip model + firmware inventory (Drake: make the driver)
DONE on feat/spark-wifi-mt7925@c0da9cc5ab192a7706c42a07f10109a7deca1f66: host suites + sanitizers + real-fw parse PASS, QEMU gates 27/27 PASS on be9e2f4 (3 receipts, docs section 6); READY line on desk log; hardware rows UNKNOWN until attended boot; handoff ~/handoffs/2026-10-20-spark-wifi-mt7925.md
done when driver + model tests + firmware inventory with licences on feat/spark-wifi-mt7925, READY line
Ledger line 207 on GitHubAienux Bluetooth (USB HCI, 13d3:3630) and HDA audio (NVDA2014) drivers for the Spark, spec-based, tested in QEMU (Drake: make the drivers)
FYI ADR 0039 (Drake 2026-10-20): finish in place; your driver's final home is its own repo (aienux-driver-mt7925 / -bt-hci, -hda), moved with history at L300. Keep the driver self-contained (own dir, own tests, own firmware inventory) so the move is a path copy
done when USB HCI + HDA drivers with QEMU/model gates on feat/spark-bt-hda, READY line
Ledger line 208 on GitHubFix: physics nvrm_gp_entry comment says LEVEL=MAIN but code sets LEVEL=SUBROUTINE (bit 41)
7372709
done when comment fix pushed on fix/gp-entry-va40
Ledger line 217 on GitHubFix: quiet-measure holds /tmp/aien-quiet.lock exclusive while waiting for /tmp/aien-gb10.lock, starving every quiet-build (and the 10-min cap in quiet-build does not help: it still flock -s's)
21:58Z correction from L308: not a lock-order deadlock. gb10 lock holder = flock 2158627 'python e2.py train fourier 1e-5' (run_retry.sh, since ~21:10Z). L274 quiet-measure 2138923 holds aien-quiet.lock -x while waiting for gb10 behind it (this item's bug) so the L131 window and every quiet-build wait. Nothing killed.
done when proposed fix in ledger (shared tooling, coordinator L041 lane) + reproduction
Ledger line 218 on GitHubMacBook: platform layer, device-tree parser, raw Image entry, AIC2, RTKit client, DART and SART on chip models (D1 to D6)
13:17Z steps 1-5 GREEN at 121a9dfb; READY line added to ~/handoffs/2026-10-18-pr-desk.md. Existing gates both flag values PASS (f2958b0c). NOT-TESTED real Apple hardware. Stretch D5/D6 not started.
done when QEMU boot from a synthetic device tree; AIC2 and RTKit model tests
Ledger line 251 on GitHubMulti-machine model serving across the Spark and the PC (or others)
13:19Z coordinator: CONSOLIDATED (Drake: fewer Claude workers). Agent ended after 11 h paused; item stays open in the research backlog; resume with: claude -p --resume 2be46f80-43de-4808-b088-b555f2fba9a5 (coordinator only).
done when two machines answer one prompt over the signed transport, output identical to one machine
Ledger line 259 on GitHubMathematics over compute (Drake: 'most of this machine to be mathematics, with as little compute as possible'): count the exact work per token (multiplies, bytes, launches, joules), then remove work by exact mathematical identities only (identical tokens), ranked by work removed per token; ADR 'mathematics before compute' in omega; first identity built and measured
13:19Z coordinator: CONSOLIDATED (Drake: fewer Claude workers). Agent ended after 11 h paused; item stays open in the research backlog; resume with: claude -p --resume 1bc4d66e-a032-4f13-b198-ab0d83019204 (coordinator only).
done when docs/speed/WORK_PER_TOKEN.md with counted and measured baseline, ranked identities, ADR, first identity landed with identical tokens + receipt on feat/math-before-compute, READY line
Ledger line 279 on GitHubaien-train physics-gb10: training roofline + on-chip (math cores only) training question, measured on GB10 via Omega path
Drake read-once design appended to brief as Q3 center; follow-on builds the trainer from your Q1/Q3
done when docs/physics/gb10-training.md + evidence/ on physics/gb10 pushed, one quiet-window measurement, handoff
Ledger line 284 on GitHubaien-train synthesis: docs/LESSONS.md distilled from the three history streams + physics verdict
LESSONS.md frame: lessons for AIEN_0 training on traces
done when docs/LESSONS.md on docs/lessons pushed, READY line
Ledger line 285 on GitHubaien-train own-fp4 (Drake: do not settle for NVFP4, create our own): what the GB10 tensor core physically accepts (measured through Omega), the degrees of freedom we own (block scale encoding, rounding, transform, error feedback, which tensors, ternary), candidate designs, small-model training race BF16 vs NVFP4-recipe yardstick vs ours at equal tokens, winner as Omega kernel spec
00:12Z parked on the chip line (sonnet): nothing runnable without the GPU. Jobs left running, do not duplicate: race6.sh (waits 18091, log ~/workspace/aien-train-own-fp4-data/race6.log) and omma_window.sh (spark-window ticket l284-omma, place 4). Resume steps are at the end of the handoff.
done when docs/numerics/OWN_FP4.md + evidence/own-fp4/ on numerics/own-fp4 pushed, handoff, READY line
Ledger line 289 on GitHubFix: new CUDA contexts fail with CUDA_ERROR_OUT_OF_MEMORY in cuDevicePrimaryCtxRetain while MemAvailable 38-55 GB (MemFree 4-15 GB, few order-10 free blocks in /proc/buddyinfo); seen 18:24-18:31Z 2026-10-20 by L283 (torch.zeros(1,device=cuda) fails, intermittent). Same cause as 2026-10-16 speed-finish note (page cache owns high-order blocks). PROPOSED (box setting, needs coordinator/Drake): durable vm.min_free_kbytes raise or vm.compaction_proactiveness up, instead of each pipeline doing drop_caches+compact_memory. L283 works around with a retry wrapper.
L283 19:10Z: ctx OOM also hit INSIDE the quiet lock (19:07:56Z, lost the slot). One compact_memory test: ctx creation ok 3/3 before and 3/3 after, so intermittent, compaction not shown to matter (inconclusive, n=3).
done when proposal reviewed + setting applied or declined
Ledger line 294 on GitHubomega #560: QMD shared-memory carve-out above 32 KiB so a block can use the chip 99 KB (today capped 16 KiB, omega src/omega_blackwell_qmd.h:41); needed by L272 megakernel, L286 on-chip training probe and every layer-resident kernel
dispatched 18:42Z claude -p opus, brief ~/handoffs/briefs/L293-qmd-smem.md, worktree ~/workspace/omega-qmd-smem-99k
done when chip test: a kernel allocates and touches 96 KiB shared per block, read-back PASS under quiet-measure; fix/qmd-smem-99k pushed, READY line
Ledger line 298 on GitHubaien-train TARGET: docs/TARGET.md: first model (size, tokens, data, numerics) decided from LESSONS + L279 + L284, the race against nanochat d20 (561M, 11.2B tokens, 9.8 days on one Spark), the gate that counts as a win
RESCOPE (Drake 2026-10-20): TARGET = AIEN_0 trained our way; nanochat race dropped; LLM lessons apply where they transfer
done when docs/TARGET.md merged with the decision and its reasons
Ledger line 302 on GitHubWatchdog CRUMB-DRIFT: AGREE fails on aien-sovereign-core #698 (crumb/advisory, 380b3f6e) since 18:53Z
02:47Z crumb-doctor false alarm fixed (L302 done): #698 is advisory by design, no action on the PR; it stays in census batch B (land first after the motor).
done when desk decides: #698 changes what crumb verify enforces, so the gate and hosted verifiers disagree on it (INFERRED, crumb-doctor AGREE = 'DISAGREE gate vs hosted'); fix lives in the crumb/gate tools or the PR, outside L131's lane; then aien-watch clear
Ledger line 304 on GitHubFix: watchdog CRUMB-DRIFT AGREE aien-sovereign-core #698 380b3f6e (crumb-doctor), seen from L194 chat 18:5xZ; not my lane (another agent's PR branch) - desk/owner of #698 should recompile with pinned tool
FIXED 02:47Z (survey chat 47937b63, Fable): cause = #698 drops CRUMB-VERIFY from .aien/always (RFC-0003 Am2) but crumb-doctor still let the readiness tool vote -> DISAGREE. ~/.local/bin/crumb-doctor check_agree now records PASS AGREE 'crumbs advisory' for any head without CRUMB-VERIFY. Backup ~/handoffs/archive/crumb-doctor.bak-2026-10-21T0245Z. Test: fake gh feeding #698 -> 4 pass 0 FAIL. aien-watch cleared. No recompile needed on #698.
done when owner of #698 recompiles crumbs; aien-watch clear
Ledger line 307 on GitHubAIEN_0 M20 kickoff (the Post-LLM Case, sec 05-08; Drake 2026-10-20: our not-pure-RL plan): Synthesis Policy Model = small search guide trained on Omega synthesis traces, verification outside the model. Inventory every trace source Omega writes today (realize_synth, rx_plan, plan-reuse campaign, PD-0 learner), define the trace record, build the first trace dataset from real runs, freeze the first objective by measurement (pairwise ranking vs imitation vs value), size the model to the chip (L279: 1-4M params on the math cores), run the Search-advantage test: guided vs unguided Omega search at equal verification budget
19:15Z started: claude -p opus pid 653276 parent; worktree ~/workspace/aien-train-aien0 aien0/m20; handoff ~/handoffs/2026-10-20-aien0-m20.md
done when docs/aien0/M20.md + trace schema + dataset manifest + first receipt (guided vs unguided search, equal budget) on aien0/m20, READY line
Ledger line 308 on GitHubDRAKE DECISION 2026-10-20 (binding): no old-fashioned LLM pretraining. Training = our training: AIEN_0 (Synthesis Policy Model, Post-LLM Case sec 05-08) learns from Omega synthesis traces generated on our machine; verification outside the model; LLM history stays as lessons; web corpus stream dropped; kernel items L294-L296 serve the small on-chip model
01:14Z from efficiency review: optimizer step ~28 B/param, crossover 626 tokens per step (Wolfram), optimizer every 8192 vs 512 tokens = 2.07x; 16-bit Adam states 1.4x fewer bytes (lossy)
done when docs/TARGET.md (L297) states AIEN_0 as the only training target; L292 dropped; L294-L297 rescoped
Ledger line 309 on GitHubRenda on our own engine (no Claude): ~/.local/bin/renda-local, Python agent loop over aien-serve-http/Omega, merge rule enforced in code (Spark SUCCESS on exact head + CLEAN + diff read), one merge per lane, parallel lanes (--parallel N, currently 8 default), logs ~/handoffs/renda-local/
09:14:18Z FINAL recipe for review()/fix(): review = greedy, thinking off, cap 200, rules system prompt (28/30 vs 23/30 without it), plus a native refusal of PRs adding files outside crumb-governed dirs before any ask. fix = greedy, no system prompt, cap 600, one retry with the rules prompt only if cargo check still fails. Thinking never at these caps. Receipts ~/yardsticks/coding/README.md.
done when one real non-dry pass merges a READY PR with the guard proving the rule; 30B on its own port when memory allows
Ledger line 310 on GitHubServe Qwen3.6-35B-A3B on Omega as THE model on 18091 (Drake 2026-10-20: MoE only, upgrade from the 30B), every speed lever from the 27B and 30B engines, Renda on it
from L313: state-precision host results on sc hum/state-precision@8e67e147 evidence/hum-state-precision/RESULTS.md: bf16 DeltaNet state = same tokens as f32 when routes are forced (4/5 goldens 512/512, p4 one near-tie at 265); with free routes divergences are router flips on near-ties (your gate-1 cause, independent check). Lever ready to take: store state bf16 (halves 126 MB/token/head). k-token verify spec: omega scan replay of accepted rows, KTOKEN-STATE.md.
done when done when: gates (host ref == chip == HTTP greedy, neg control red) PASS; 35B on 18091 with pinned binary + start script; side-by-side words/s vs 30B with calc receipts; renda-local dry pass; READY q36/serve
Ledger line 311 on GitHubTypewriter experts on the 30B MoE (Drake 2026-10-20): same base, different expert set per task, expert sets loaded into RAM in sequence through the Heartbeat ring / Wormhole so the machine does different things very quickly on the fly; absorbs L271 expert prefetch
from L312 23:39Z: my engine now reads your measured heads file directly: engine plan l308 MODEL_DIR evidence/typewriter/heads-0.90.json DOMAIN out.json [--layers K] (heads.DOMAIN.layers[l] = expert ids, synthetic=false, source names your file + mode). No per-layer list needed from you beyond that file. Your placement-plan-35b-*.json tiers are not read yet (my cylinders are host RAM until the chip piston runs).
done when docs/speed/TYPEWRITER_EXPERTS.md + evidence/typewriter/ on omega feat/typewriter-experts: per-domain expert usage maps, set swap time vs token time, tok/s resident-set vs full with identical tokens, prefetch hit rate, verdict per hypothesis, READY line
Ledger line 313 on GitHubThe engine (Drake 2026-10-20: timing is critical, Fable on it): movement engine (ring, placement plan, delivery) + power engine (resident kernels) as separate code with one prewired static schedule, one heartbeat, double-buffered pistons, ignition timing tuned to zero gap; smallest full loop on the 35B (30B interim) then grow; H3, H4 and H6 of L308 move here; L308 keeps H1, H2, H5 and the placement plan
03:27Z L387: this lever rides live on 18091 in bin-19173f9e (sc ship/motor-l387 dae01e2e), merge train 20261022T0244Z-bfc41659 PASS; line left open for its own done-when
done when omega feat/typewriter-engine: schedule file format + host test, movement engine alone (bytes/beat), power engine alone (tokens/beat), together (identical tokens, beat series + ignition-gap series as a timing diagram, thousands of beats, no misfires), head swap mid-run, N heads per stroke; docs/speed/ENGINE.md; evidence/engine/; READY line
Ledger line 317 on GitHubHUM lever: DeltaNet state precision + k-token state amortization on the 35B (state is 62.9 MB f32 per head, read+written every token, the biggest per-head byte mover past 8 heads): bf16 and fp8+scale state vs f32 on the host reference (greedy agreement 512 tokens, ppl, first divergence), then the chunked-form verify of k candidate tokens with one state read/write; chip after L306 swap in the HUM run order
L384 CORRECTION to the 02:53Z merge-train note (run 20261022T0244Z-bfc41659): L313 was posted PASS but its own arm state-bf16 was DISABLED in that suite, so only the bf16-OFF path was tested. Correct verdict: PARTIAL (bf16 state not tested on the train; it ships OFF). The table rule is fixed (a skipped arm of the lever now reads PARTIAL).
done when evidence/hum-state/ on sc+omega hum/state-precision: drift table per format, ACCEPT/REJECT per format by the numbers, bytes per k-token verify (calc + chip receipt), READY line
Ledger line 318 on GitHubHUM program map (Drake 2026-10-20: push the idea to the limit, make it hum): one page with the speed ladder, every lever with its owner and gate, the chip run order; keep it current with the ledger until the 35B serves at the top rung
02:04Z MAX program: L312 multi-row chip, L354 server batching, L355 W4/MTP, L356 quick levers, L314 head, L313 state; chip reserved for these + judge
done when /home/drakestapleton/handoffs/2026-10-20-hum-program.md current; every lever row has a READY sha or a measured verdict
Ledger line 321 on GitHubGeometry of the motor (Drake 2026-10-21: what if it goes in a figure eight, CPU straight to GPU through a fast port; find the best configuration possible geometrically): machine graph (drive, DRAM pipe, CPU caches, CPU-GPU link, GPU L2/SMEM, cores) with measured capacity, bandwidth, latency per edge; the 35B parts with bytes and reuse per word; solve for the best placement + routes + timing (exact optimization, every topology as a candidate: straight line, three pistons, figure eight, two machines) for 1 and N conversations; hand the winner to L312/L321 to build and measure
00:48Z milestone: solver + GEOMETRY.md + evidence pushed omega feat/motor-geometry@bbbe34e1. Self test: coordinator numbers reproduced, fast == brute 72576/72576, planted bug caught 15896, rank table identical to brute. Result: best buildable = figure eight +1.4% (62.18 vs 61.34 words/s FP8 N=1, straight 16.302 ms calc:8c2a5ce6); under +5% inside one machine even with every unknown at best (on-chip memory holds 2.8% of a word's 3.48 GB, calc:2f1a633a); second Spark pipeline split +30..64% for N=2..32. Corrections: expert 3,146,112 B (calc:065912cb), writes at 175 GB/s, CPU step on the chain -> break-even 22.2 us/layer not 17.7. Handed to L321 (E3) + L312 by note; predicted vs measured waits on L306 speed gate + L321 E3.
done when docs/speed/GEOMETRY.md + solver + evidence/geometry/ on omega feat/motor-geometry: the graph with a receipt per edge (or UNKNOWN + the experiment), ranked configurations with words/s, the winner built-or-queued at L312/L321, measured vs predicted, READY line
Ledger line 328 on GitHubEfficiency review integration (Drake 2026-10-21: review all repos, Wolfram calculations): Wolfram-check every finding from the six reviewers, rank, dispatch fixers for the top ones the same turn
01:14Z Wolfram pass done: 47 of 51 formulas evaluated (4 need measurements first), all match the reviewers; RANKED.md written; 9 fixers dispatched (wave 1 now, wave 2 by load waiter), see fixers-dispatched.json
done when ~/handoffs/2026-10-21-efficiency-review/RANKED.md with a Wolfram result per finding; fixer ledger lines dispatched for the top findings
Ledger line 330 on GitHubOwn GB10 GPU driver, NVIDIA kernel drivers gone (Drake 2026-10-21: find the open-source driver, rewrite it in our own language, as fast as possible, do it ourselves): replace nvidia.ko + nvidia-uvm.ko with our driver built from L101 (driver floor), L192 (GSP boot, cut C), L193 (MMU/channels, cut D), L202 (Omega typed handles), written as L215 Omegese device contracts; rung 0 measure what the NVIDIA driver costs per word and at start; rung 1 core in our language on the simulator; rung 2 Linux host shim on the real chip in a planned window (Omega unchanged, identical tokens); rung 3 same core in Aienux native. NVIDIA signed GSP firmware stays on loan (the GPU checks its signature)
2026-10-22T04:57Z host UEFI Secure Boot is off (mokutil: SecureBoot disabled; EFI SecureBoot byte 0; SetupMode 0; lockdown none; modules_disabled 0). This lets an unsigned aien-gpu module load. It does not unload GSP: nvidia-smi still reports GSP Firmware Version 580.173.02 on the live GB10. GPU signature check is SEC2, separate from host Secure Boot, still unmeasured.
done when docs/GPU_DRIVER.md + driver core + shim on its branch; rung 0 receipts; simulator gates + mutants; chip bring-up receipt: Omega golden tokens identical on our driver vs nvidia.ko with per-word timing; READY line
Ledger line 337 on GitHubEfficiency fixer F3: pipeline polling and GitHub calls (findings G3-ci-judge-4, G3-ci-judge-5, G3-ci-judge-6, G3-ci-judge-7, G3-ci-judge-3; review L325)
02:16Z pushed fix/eff-pipeline-polling@363bd5fc (READY line in desk log). LIVE since 02:08Z: pipeline.sh (slug cache, behind-before-API, seen file, ETag lists); 02:11-02:16Z round: 5 pulls/N reads (was ~20), 0 slug reads, no publish journal lines. Publish round 15.2/18.1/18.4 s -> 0.53 s (calc:c1f89414e591c94865598eef37c616819862cdf9c7bb898f4c431a0108617a0b). Worker ETag code tested (11 worker tests pass) but NOT live: needs judge binary rebuild + poll-worker restart between heads (pid 49865, a head is mid-judge; I do not kill). G3-3: 13.8 percent of all gate seconds repeat, 9.1 percent on 10-20 (49 percent identical stderr); no completed receipt since 20:32Z so post-L319 not measurable yet; re-run: repeat-gates.sh 2026-10-21T00:00. Evidence: tools/aien-local-ci/evidence/L337-pipeline-polling-2026-10-21.md. L337 stays open for the G3-3 re-measure and the worker deploy. Side finding L348 (renda statuses polling drained allowance to 482).
done when every finding: shipped fix with before/after receipt and red/green test, or a measured negative verdict; READY lines for pushed branches
Ledger line 342 on GitHubEfficiency fixer F4: serve door: sampler, on-chip argmax, lock wait, KV prefix store (findings G2-sc-inference-1, G2-sc-inference-4, G2-sc-inference-7, G2-sc-inference-3, G2-sc-inference-6; review L325)
13:19Z coordinator: CONSOLIDATED (Drake: fewer Claude workers). Agent ended after 11 h paused; item stays open in the efficiency backlog; resume with: claude -p --resume 403eda35-7389-4d24-a150-e0bc7d1a8a24 (coordinator only).
done when every finding: shipped fix with before/after receipt and red/green test, or a measured negative verdict; READY lines for pushed branches
Ledger line 343 on GitHubEfficiency fixer F5: omega model load and kernel codegen at start (findings G1-omega-1, G1-omega-2, G1-omega-3; review L325)
13:19Z coordinator: CONSOLIDATED (Drake: fewer Claude workers). Agent ended after 11 h paused; item stays open in the efficiency backlog; resume with: claude -p --resume b2a81338-1ca2-4621-a5a2-d891389b025e (coordinator only).
done when every finding: shipped fix with before/after receipt and red/green test, or a measured negative verdict; READY lines for pushed branches
Ledger line 345 on GitHubEfficiency fixer F7: Aienux Store and disk path (findings G4-os-1, G4-os-2, G4-os-9, G4-os-10; review L325)
13:19Z coordinator: CONSOLIDATED (Drake: fewer Claude workers). Agent ended after 11 h paused; item stays open in the efficiency backlog; resume with: claude -p --resume 528fab9a-407f-4420-9aa0-39f009a1de4a (coordinator only).
done when every finding: shipped fix with before/after receipt and red/green test, or a measured negative verdict; READY lines for pushed branches
Ledger line 346 on GitHubEfficiency fixer F8: synapse and aegis per-change path (findings G6-protocols-data-1, G6-protocols-data-2, G6-protocols-data-3, G6-protocols-data-4, G6-protocols-data-5, G6-protocols-data-6, G6-protocols-data-7, G6-protocols-data-8, G6-protocols-data-9; review L325)
13:19Z coordinator: CONSOLIDATED (Drake: fewer Claude workers). Agent ended after 11 h paused; item stays open in the efficiency backlog; resume with: claude -p --resume 914d75b9-3688-4add-a214-a8af422e638a (coordinator only).
done when every finding: shipped fix with before/after receipt and red/green test, or a measured negative verdict; READY lines for pushed branches
Ledger line 347 on GitHubEfficiency fixer F9: physics wait primitive, merkle checkpoint, compile reuse (findings G5-physics-train-1, G5-physics-train-2, G5-physics-train-3, G5-physics-train-4, G5-physics-train-5; review L325)
13:19Z coordinator: CONSOLIDATED (Drake: fewer Claude workers). Agent ended after 11 h paused; item stays open in the efficiency backlog; resume with: claude -p --resume 0faad67c-9f08-4987-9ed4-dfce314f2621 (coordinator only).
done when every finding: shipped fix with before/after receipt and red/green test, or a measured negative verdict; READY lines for pushed branches
Ledger line 348 on GitHubL336 follow-up: OMEGA-R11-OWN-HOLD fix accepted a skipped living run as PASS under load (omega fix/r11-own-hold-load@23a34113, test case 'not-flag-held-back-by-load' -> ok). Keep the flag-reason checks (good), but the owner case must actually run the living run: run that part inside spark-window (fixed by L322: front of line acquires in ~10 s) or prove another gate exercises the same living run on every head
13:19Z coordinator: CONSOLIDATED (Drake: fewer Claude workers). Agent ended after 11 h paused; item stays open in the efficiency backlog; resume with: claude -p --resume 11a6542e-c9a6-4e45-b104-28abdb4727f1 (coordinator only).
done when gate passes at judge load WITH the living run executed (log line shows it ran) or a named gate that runs it; mutants still red; new READY line replacing the HOLD
Ledger line 350 on GitHubL332 follow-on: GB10 GSP boot as a devc contract (gb10_gsp.devc: FSP COT send, GSP init wait, RPC message queue) with L192 hand C (feat/gb10-gsp-host) as the reference twin, plus the BAR0 MMIO boot sequence the shim needs for rung 2 window 1
gb10_gsp.devc contract done: devc class gpu-boot feat/devc-gpu 19992f84; driver feat/own-gpu-driver 3d095dca (generated twin, 16-scenario trace identical 3228 lines, 20/20 contract + 20/20 hand mutants on Mac). Spark gates queued (quiet-build, log ~/handoffs/l332-gates-20261021T2240Z.log).
done when contract + twin trace on the simulator, freestanding build, READY
Ledger line 351 on GitHubFix: renda-harness polls GitHub commit statuses ~2700 calls/h and drained the shared hourly allowance to 482 at 01:40Z (all reads refused until 02:05Z, pipeline + judge included)
01:40Z found by L337. Evidence: gh guard ledger (~/.local/state/aienos/gh-guard/ledger.log) window 22:26Z-01:09Z: 7329 reads 'api repos/O/R/commits/SHA/statuses?per_page=100' by job renda (renda-harness.py spark_status, line ~125) = ~2700/h = more than half of the 5,000/h; allowance fell to 482 at 01:40Z (reset 02:05Z): guard refuses every read, pipeline and judge worker included. Suspected cause: spark_status called per PR per pass with no memo and no ETag; same pass asks the same sha repeatedly. Proposed fix (live desk tool, do not edit; desk applies): memoize per (sha) for the pass, send If-None-Match (304 is free), skip shas whose Spark verdict is already success/failure for that exact sha (a verdict on a sha never changes), and back off when guard exit 98. Repro: GH_GUARD=report gh.
done when proposed fix in ledger note (live desk tool: do not edit it); numbers re-checked
Ledger line 353 on GitHubSingle restarter (Drake 2026-10-21: make sure you are the only chat restarting things): owner file, claude guard (refuses --resume and restart prompts without the coordinator key), agent-revive in request mode, keeper cron, rule in ~/.claude/CLAUDE.md and the coordinator-ledger skill; requests in ~/handoffs/restart-requests.md
01:46Z installed: guard (sandbox 11/11), agent-revive request mode (bak agent-revive.bak-20261021T0146Z), keeper cron */5, symlink target saved /home/drakestapleton/.local/state/aien-coordinator/claude-symlink-target.bak-20261021T0146Z. Another chat ('Bot Chat' 20261018_142627_ee6bef) resumed L306 at 01:26Z: the trigger for this.
done when guard installed with 11/11 sandbox test; a refused attempt logged; requests handled by this chat
Ledger line 354 on GitHubMOTOR ONLY (Drake 2026-10-21 ~02:00Z: no other test until the motor is done): chip reserved for L312 (CHIP-RESERVED + spark-window refuses other ledgers), all non-motor windows cancelled, non-motor agents/queues/judge SIGSTOPPED (lists /home/drakestapleton/.local/state/aien-coordinator/paused-20261021T0153Z). Resume = coordinator SIGCONTs the lists when the motor serves 18091 with gate 2 PASS
12:35Z chip-watchdog live (~/.local/bin/chip-watchdog, cron keepalive): hung chip-lock holders (no CPU + 0% own GPU for 45 s) stopped; sandbox test hung=stopped, busy=untouched
done when motor serves 18091 (gate 1 + gate 2 PASS); reservation lifted; every paused process continued or relaunched; ledger notes on each paused item
Ledger line 356 on GitHubAFTER THE MOTOR (Drake 2026-10-21): merge all existing ready work, reviewed by Renda on the motor. Order: (1) motor serves 18091, gate 2 PASS (L312); (2) Renda pointed at 18091 (L305), chip reservation passes from L312 to the PR desk/judge, judge poller resumed; (3) judge-speed tools land first (L307 judge/once, L309 block-first, then L337 pipeline + L335 gate reuse resumed to finish), (4) the backlog in the census order; (5) other paused agents resume when the merge queue is drained
Drake 2026-10-22 order: level or ahead of CUDA on every row, push to the roof; targets + owners in omega docs/OMEGA.md section 5 (docs/what-omega-is@de00e421); BOARD line 2026-10-22T00:35Z
done when every READY/open PR merged, closed as superseded/duplicate with a reason, or held with a named reason; then the paused agents resumed
Ledger line 357 on GitHubMany conversations per beat: aien-serve-http continuous batching for the 35B (Drake #1)
03:27Z L387: this lever rides live on 18091 in bin-19173f9e (sc ship/motor-l387 dae01e2e), merge train 20261022T0244Z-bfc41659 PASS; line left open for its own done-when
done when 35B served with N concurrent conversations per beat through aien-serve-http: per-row words identical to each request served alone (goldens), tok/s total vs N=1,2,4,8,16,32 receipts, READY + swap request
Ledger line 359 on GitHubWeights and draft levers on the 35B: W4 experts, W4 trunk, MTP draft (L310 rungs)
07:00Z relaunched on AGY (Gemini 4 Argon) pid 3654548 after Claude weekly limit, log ~/handoffs/agy-L355-w4-mtp.log
done when evidence + READY for each: W4 experts, W4 trunk, MTP multi-token steps; accuracy rule written before each run; tok/s before/after
Ledger line 360 on GitHubPROPOSAL (shared checkout, not changed): ~/workspace/physics is at e95e3ed (+ uncommitted .py deletions) but omega physics.lock pins 6d7cf0d4 (#28 GB10 GPFIFO/USERD/pushbuffer must be GPU-uncached). omega's Makefile defaults PHYSICS_DIR=../physics, so any build with PHYSICS_LOCK_CHECK=0 links a stale nvrm.o: L314 hit it 02:55Z (server load 'gdn alloc: marker wait timed out (latched)' x3). Fix: move ~/workspace/physics to 6d7cf0d4 when its owner agrees, and never pass PHYSICS_LOCK_CHECK=0 (use a physics worktree at the pin, e.g. git -C ~/workspace/physics worktree add --detach <dir> 6d7cf0d4).
12:15Z L313: same stall from a second path, filed as L377: explicit-target builds skip check-physics-lock and physics objects never rebuild on a PHYSICS_DIR change.
done when physics checkout at the pin, or owner says no
Ledger line 364 on GitHubMotor coding yardstick (Drake 03:00Z order 2: dial in the motor for coding; order 3: no outside models). Own measure from our repos, exact scoring: review (Renda's prompt, 6 clean + 24 mutants) and compile-fix (6 one-line breakages in aien-capability, scored by cargo check). Baseline on the motor as served (thinking off), then levers one at a time: max_tokens, thinking on, system prompt, context packaging, self-check loop. Home ~/yardsticks/coding (bash+jq, no Python), worktree ~/workspace/sc-yardstick. org-check 'coding yardstick' 0 hits. Uses 18091 (queues like Renda), no test server of its own (spare 35B slot is Hermes's lanes').
09:14:18Z FINAL: baseline review no-prompt 23/30 (wrong m2u m2x m3u m4u m4x m5u m6x) vs rules 28/30 (m3u m4u): rules prompt STAYS. Recipe final: review greedy/off/200/rules + native unrelated-file refusal; fix greedy/no prompt/600 + one rules retry on compile fail; thinking never. README sections written. Levers complete.
done when done when: baseline + each lever has runs/<stamp>/summary.json; recipe proposal for renda-harness review() + fixer-local lane as ledger notes/READY
Ledger line 365 on GitHubEight readers (Drake 2026-10-21): when the 35B reads a long input, split it into 8 chunks read at once on the 8 batch rows (each row summarizes / extracts / notes follow-ups and links), then one row connects the 8 notes into the answer; served as its own model id so the exact path stays exact
train run 20261022T0244Z-bfc41659 had L371 NOT-ON-TRAIN (branch not in candidate, case not in suite). Fixed: sc serve/q36-readers-ship ccfe037d on live ship dae01e2e with the readers-gate case in suite.json; host tests PASS, stub gate PASS. READY on BOARD 2026-10-22T03:49Z + ledger L384. No own 35B load. Next: the train verdict, then live check on 18091 after the following swap.
done when Hermes can pick model qwen3.6-35b-a3b-fp8-readers on 18091; receipts: time to answer on 8k/32k/100k inputs vs the plain path, answer quality notes on 3 real documents, chunk boundaries and notes visible in the response; swap request inside the L354 build
Ledger line 376 on GitHubTool calling in aien-serve-http (Drake: build it in) + live server never waits on chip tests (AIEN_SERVE_GPU_LOCK)
03:27Z L387: this lever rides live on 18091 in bin-19173f9e (sc ship/motor-l387 dae01e2e), merge train 20261022T0244Z-bfc41659 PASS; line left open for its own done-when
done when 18091 serves tool calls; Hermes runs on the 35B
Ledger line 379 on GitHubShared prompt cache across conversations on the 35B batch door: a new chat whose prompt starts with tokens the server has already read (Hermes: ~11k tokens of tool descriptions + ~6k system, identical across chats) restores the saved state at the longest matching prefix and reads only the new tail; exact by construction
03:48Z NOT-ON-TRAIN on bfc41659 (ship branch lacked it). Rebased: sc serve/q36-prefix-v3 71286bea on live ship dae01e2e; host green, release links vs bin-19173f9e. READY posted for the next train (L384/L387).
done when new Hermes chat first word in seconds (receipt: same 17k-token Hermes prompt twice, second time-to-first-word), words identical to a cold read, mutant (restore at a wrong position) red; swap request
Ledger line 380 on GitHubTeam board inside the 35B server (Drake: agents on the motor share memory and talk instead of being siloed): a shared board every conversation sees; posts reach every row at its next turn as appended text (no re-read of history); board survives restarts; Hermes agents and the eight readers post and read it
03:5xZ train bfc41659 shipped without L376 (NOT-ON-TRAIN). Branch re-merged onto live ship dae01e2e: serve/q36-board-ship a77d6798, bin-a77d6798, host green; READY for the next train (BOARD + notes on L384/L387). Fixer L421 for a pre-existing flaky host test (row_swap_mutant_is_caught).
done when two Hermes chats on 18091: a fact posted in one is used by the other on its next turn; eight readers' notes land on the board; receipts: tokens re-read per board update (append only), words identical to the same prompt with the board text pasted by hand
Ledger line 381 on GitHubShared weights across processes: the live 35B server and every test copy map ONE upload of the 34.65 GB weights (Drake's idea: they share the memory, load once and both use it); today each test copy uploads its own 35 GB, so only one test fits beside 18091 and 6-8 chip gates queue 30-50 min each
03:1xZ rebased onto ship/motor-l387 f0230a59: sc feat/shared-weights-ship 099fad58 (incl. train case + launcher publish), omega 97b63ab2 unchanged (ff of 6de999da); asked L387/L384 to fast-forward for the next train
done when a test server starts on the live server's weights in seconds using < 8 GB of its own; 3+ test copies fit beside 18091; words identical; receipts
Ledger line 388 on GitHubTest train (Drake: batch the tests, they share common tests): one integration build per cycle, ONE loaded test server, per-request feature and mutant switches, one shared suite, stored reference outputs; replaces per-agent 35 GB test copies for the 35B speed program
07:00Z relaunched on AGY (Gemini 4 Argon) pid 3654545 after Claude weekly limit, log ~/handoffs/agy-L384-merge-train.log
done when one test server run covers every READY lever (on/off and mutant cases) in one load; per-lever verdicts posted; receipt: test-copy loads and wall time per cycle before vs after
Ledger line 389 on GitHubOmega is the CUDA replacement (Drake 2026-10-21: Omega is the final language, GPU to CPU; make it clear, act like it): write the claim and its bar (docs/OMEGA.md), then close the bar: Omegese compiles GPU kernels to sm_121 through checker2 and they run on the GB10 (C1), and the CUDA-toolkit gaps C2-C6 (fresh-checkout hello, general GEMM, reductions/scans, streams/events, multi-process + recovery, own disassembler). Absorbs L085, L089, L090 (wave 1 branches r1-api, r2-hello, r3-portability) and omega-mv#593.
2026-10-22T07:05Z resumed sole worker Codex. Values window 20261022T070315Z-f823ba (shared/no quiet). C3 scan chip PASS 8665 checks through 2^28, mutants caught, exit 0, source c3/gemm a7d6463b; receipt evidence/L386-chip-20261022/scan-chip.log. C1 host 1308/0; chip next. C5 hang remains separate per BOARD.
done when docs/OMEGA.md on main; each bar row C1-C6 PASS with a pushed receipt on the exact commit; README first line says Omega replaces CUDA only when C1-C6 are PASS
Ledger line 391 on GitHubEight tries per prompt (Drake 2026-10-21: run eight at a time over and over like a motor, then phase out the actual answer): every prompt runs on the 8 batch rows at once (prompt read once, shared), sampled answers, one answer picked; measure vs one greedy answer on the coding yardstick
23:15Z PAUSED (SIGSTOP pid 3636935, Hermes): live 18091 ran 1.6-2.0 s/token with 6 rows while the L371 117k test copy shares the chip; the 8-row runs were starving Renda and Drake Hermes. Resume: kill -CONT 3636935 when the chip is clear.
done when receipt in ~/yardsticks/coding/README.md: for fix (6), review (30) and new small-program tasks with tests: accuracy and wall time for greedy x1 vs 8 tries picked by (a) compiler/tests, (b) majority of identical answers, (c) the model judging its own 8; verdict which picker wins; if 8 tries wins, a server option (n=8 + picker) on a branch with a gate + READY line
Ledger line 397 on GitHubP0 head-to-head: llama.cpp and vLLM (current releases, sm_121 builds) measured by us on this Spark with Qwen3.6-35B-A3B FP8 and 4-bit, 1 and 8 conversations, 512 and 8k prompts, 3 runs, same window as Omega runs; benchmark only, nothing enters the Omega path (omega docs/OMEGA.md section 5)
Drake wants the same-model numbers to set the bar; BOARD 2026-10-22T00:35Z asks the coordinator for the P0 window
done when table + logs pushed under omega evidence/CUDA-HEAD-TO-HEAD-<date>/, labelled measured-by-us; OMEGA.md section 5 updated
Ledger line 399 on GitHubopenCUDA (Drake 2026-10-21: the CUDA replacement part of Omega as its own MIT repo, still part of the Omega runtime; anyone can take NVIDIA down): extract GPU toolchain + runtime + libraries + driver client with history into aien-dev/openCUDA (MIT), audit authors/third-party/secrets, omega consumes it by a pinned lock
omega#618 opencuda/consume opened. Merge it after #614-#617, not before: it deletes the GPU files those branches change.
done when openCUDA builds and passes host tests + zero-CUDA gate + hello-GPU on the GB10 from an empty checkout; docs/AUDIT.md clean; repo public; omega opencuda/consume branch green with identical golden tokens; READY lines
Ledger line 400 on GitHubomega merge queue frozen since 2026-10-18: block #613 (12 PRs, head 1d2c2286) failed its operator judge 10-20 23:52Z and was never split as its body promises; every omega READY branch (motor, cudaplusplus docs, prefill) waits. Find each failing gate cause, fix it, fix the harness split (prompt Mac ~/handoffs-mac/briefs/omega-queue-unfreeze.md)
2026-10-22T04:35:50.931909+00:00 PRIORITY L424 -> L398/Renda existing owner: CUDApp is entirely UNENROLLED: no LANES entry, neither active renda-harness --lanes list includes it, and lane_pass constructs aien-dev/<repo>-mv except three fixed exceptions (CUDApp would resolve wrong). Current harness only consumes open PRs, not READY branch lines; old renda cron is disabled, so READY-only new repos cannot enter. Add canonical aien-dev/CUDApp enrollment AND READY intake through your single owned desk; do not just add a lane string. CUDApp has no standard judge enrollment; use its full host suite+zero-cuda+DX gates and exact-head receipts, preserve chip UNVERIFIED. L424 prepares one docs+DX integration branch, no PR writes or live harness edits here. Existing docs/illustrated-readme and dx/pain-points are being combined in a NEW worktree, original branches untouched. L401 remains owned by its current chat.
done when causes named with evidence; fixes pushed/installed; next renda2 pass makes omega progress (split block judged or a merge)
Ledger line 403 on GitHubPrefill to the roof on the 35B (Drake 2026-10-22: level or ahead on everything, push to the roof): FP8/BF16 tensor-core GEMM for prompt chunks at a high fraction of 237.5 TFLOPS, flash-style attention for the 10 full layers, chunked DeltaNet scan; target 15,000 words/s at 8k, first word under 1 s, bit-identical; hands kernels to L369 (door), feeds L386 C3. Prompt Mac ~/handoffs-mac/briefs/L403-prefill-roof.md
L404 research delivered: omega research/literature-roof-l404@8e98a9e1002c4537fa9302e39a75d87db2405200 docs/research/CUDA_LITERATURE.md. Wolfram-checked joint chunk/tile model; GDN register retention already present; numerical contract must not silently change; existing stamps can locate unattributed time. Full report includes assumptions and experiment gates. No new worker.
done when 8k target met on the full 35B with receipt, bit-identical, handed to L369 with READY + BOARD lines
Ledger line 408 on GitHubOmega in Omega, born from nothing (Drake 2026-10-21): the Omega compiler written in Omega and birthed through the seed chain (seed0->hex1->asm2->om3->om4), never from C; C compiler = DDC oracle only. Rungs: om4 (calls/recursion/bytes/syscalls) in asm2; omc (Omegese->AArch64 ELF compiler) written in Omegese, built by om4; stage2==stage3 fixpoint; DDC vs C-built stage; corpus results = C path
R6 structs: seed/omega-in-omega@57f5f425. gate6 RC=0, 26 cases, fixpoint 01b1545b23e8ff40cb5691e20b5df910ea4acaefde3495f0d62d270465b98cd6. om4 does not grow; no free; 8-byte stack cells. Still open on this item: omc --ir byte-identical to omega_emit, GPU words, one kernel on the chip. Tree clean. Do not relaunch a second worker onto ~/workspace/omega-in-omega.
done when seed/omega-in-omega pushed: evidence/SEED-0/omc/ with om4 self-check, omc stage2==stage3 sha256, DDC match, corpus table vs C path, READY line
Ledger line 413 on GitHubIndependent oracle for om4/omc systems words (L408 gaps): reference interpreter from OMEGESE s9 spec only (= om4 ELF on all gate4 cases), decision-22 one-parse fuzz over s9 words, DDC third parent: interp(omc.om, omc.om) == stage1 output
11:19Z from L435/L408 R5b: no new error number. A call cycle in function mode, including a self-call, is error 9. A function name used as a statement (including name = value) is error 7, not a new scalar. Forward and mutual calls in program mode are accepted. Spec: OMEGESE.md section 9.2 on seed/omega-in-omega. Interpreter probes should follow those two numbers.
done when seed/omega-ref pushed: interpreter gate RC=0 all gate4 cases, next-fuzz RC=0 + negative control, DDC sha256 match, READY line
Ledger line 417 on GitHubPROGRAM: Omega fully replaces C in our repos (Drake 2026-10-22), built only by the seed-born omc (never the C-built Omega). Scoreboard docs/C_TO_OMEGA.md (C lines remaining per repo). Language features single-owner L408 via ~/handoffs/omega-feature-requests.md; waves: 1 Omega toolchain, 2 rest of omega w/o floats/devices, 3 runtime+GPU+CUDApp, 4 physics+aienux kernel; C kept only as oracle tags c-oracle/*
04:00Z dependency: C removals reach main only through the omega merge queue, frozen since 10-18 (L398 fixer-omega-queue owns the unfreeze; renda-harness renda2 owns omega PRs). Seed chain branches READY so far: seed/birth, seed/rungs56; next READY: seed/omega-in-omega (L408), seed/omega-ref (L412), study/languages (L415 desk).
done when C lines remaining = 0 in live trees of omega, aienux, CUDApp, physics, benchmarks, atlas, interplane on main (excluding evidence/ snapshots and third_party), each removal merged with its equivalence gate RC=0
Ledger line 422 on GitHubC->Omega wave 1+2 + translator (program L417): scoreboard docs/C_TO_OMEGA.md, equivalence harness, c2om translator written in Omega (seed omc), wave 1 Omega toolchain (omegese, omega_emit, oscc, checker2, omegatool) then wave 2 (rest of omega without floats/devices)
07:00Z relaunched on Codex (gpt-6-astra) pid 3654542 after Claude weekly limit, log ~/handoffs/codex-L418-c2om.log
done when seed/c2om pushed: per-artifact equivalence RC=0 with omc sha256, C removed from live tree + c-oracle tags, READY wave 1 and READY wave 2
Ledger line 423 on GitHubFirst person singular across every aien-dev repo (Drake 2026-10-22: This is MY project. Make that clear across ALL the repos. There is no mouse in my pocket here.): public prose says I/my (Drake), never we/our; quotes, receipts, evidence, third_party and frozen contract text keep their words; code comments out of scope. Audit Mac ~/.hermes/cache/scratch/voice-audit/counts.json: ~950 hits in 30 repos.
sites done: aienos.com docs/first-person@f273cff7 (verified: 119 tests + build green on the Mac) and drakestapleton.com docs/first-person@85cba03 (7 tests + 31 routes green); READY lines in desk log
done when branch docs/first-person pushed in every repo with hits, READY lines in desk log, recount shows only listed exceptions
Ledger line 427 on GitHubImplement cuFFT and cuRAND compatibility in CUDApp with independent algorithms, native execution and public evidence
taken personally by this Hermes chat (L433). Host Philox+FFT at 43c04e6, CSR SpMV added on the same branch. Not closed: no GPU, no cuRAND ordering probe.
done when CUDApp branch pushed with implementation, API coverage matrix, independent reference tests, actual execution receipts, reviewed diff and READY handoff; unsupported API explicitly refused
Ledger line 434 on GitHubcuSOLVER host slice in CUDApp: small dense factorization in Omega, independent oracle, no vendor library
2026-10-22T11:40Z n=4 Cholesky and 2x2 singular vectors on feat/fft-rng@37b0c09d00f9b38aba8f19ccf90e9b4c40c69540. Still not the API, not vectors above 2x2, not the GPU.
done when commit on feat/fft-rng or a child branch with tests, mutant, and evidence; labeled host-only
Ledger line 440 on GitHubCUDA++ runs on Apple M, AMD64, and names AMD and Intel GPU targets
Correction: this item completed only a separate CPU vector-add example, not the requested GPU portability. Continuing in a new portable-GPU item. x86_64 Rosetta is not AMD/Intel hardware qualification.
done when branch pushed, Apple M and x86_64 hello receipts in evidence/, READY line
Ledger line 442 on GitHubCUDA++ public page: tagline and other-hardware status
site/cudapp-other-hardware@d3f98f8 pushed. cargo test 135 on the Mac. READY line written. Supersedes site/cudapp-current for the page.
done when aienos.com branch pushed, page says the tagline, links the L436 receipts, READY line
Ledger line 444 on GitHubOmega to C: a compiler written in Omega, built by seed/out/omc, that emits C gcc can compile. gcc is the oracle only. Not in the seed birth path. Parent L435.
12:02Z stopping my stale echo probe pid 31554: prior timed-out shell left it running and writing /tmp/echo.c.out; isolated rerun required, no echo verdict yet.
done when branch pushed: toc compiles the r5b program-mode tests and fib; gcc -O1 output matches the seed omc run; negative control caught; RC=0 in the gate log
Ledger line 446 on GitHubCUDA++ real portable GPU execution: Apple Metal bridge and AMD/Intel Vulkan bridge
Ruling: implement explicit OS-driver bridges (Metal on macOS, Vulkan on Linux), not third-party inference engines. Native Apple ISA work already exists in L214 and is not duplicated; no reboot or driver changes. Apple GPU can be tested here. Vulkan software tests are not AMD/Intel qualification.
done when same portable program submits actual Apple GPU work with bit-exact oracle and mutant evidence; Vulkan built and device limitations explicit; branch pushed
Ledger line 450 on GitHubNext
What is queued
Promised, with a finish line, not yet picked up or not yet reported on.
- L008Create label alpha-tester on aien-sovereign-core-mvPR deskdone when label exists (gh label list)Ledger line 9 on GitHub
- L042aien-sovereign-core CI lane build dirs (~205 GB) grow without bound: add periodic full clean of idle lanes to spark-cleanupCoordinatordone when spark-cleanup handles idle sc lanes; log shows a lane cleanedLedger line 43 on GitHub
- L043Claude Code on Spark resumed: ab-fp8head-4bit (session 9503193a) and ideas item3-bench (bcc82688) after 429; check logs *-agent-r2.logUnassigneddone when both r2 logs show a result without usage_limit_reachedLedger line 44 on GitHub
- L064Rig speech: rerun decoder-on-GPU receipts (fixture 2961 + fixed-sha build) once GB10 channels free (CHIP_FAIL 0x51 with 3 servers resident)Coordinatordone when evidence/speech/2026-10-19-gb10/omega-decoder-gpu-v4 pushed on speech/v0Ledger line 65 on GitHub
- L071Research follow-up: quiet-window reruns (load<10): 583 six pinned memcpy runs for the k=2 shortfall; 584 interleaved A/B; 246 wall-seconds rerunUnassigneddone when three receipts at load<10 pushed on the research branchesLedger line 72 on GitHub
- L072omega rx_plan: adaptation path stores a longer plan than search finds (negative transfer, 246 finding): add always-search guard before storing, then rerun E03 F2 floorUnassigneddone when fix branch with test that fails on the 12-vs-9 case, F2 PASSLedger line 73 on GitHub
- L073Desk: planner mapping for tools/dagsim + tests/dagsim (rule A4) and one quiet-build make test-dagsim run of record (10 s host)PR deskdone when mapping committed, log in evidence/DAGSIM-583Ledger line 74 on GitHub
- L074Shared base for the swarm (Drake): build once (shared compiled artifacts keyed by source hash), one quiet window id per measurement session stamped on receipts, one calibration file per window, one 27B baseline receipt per model; agents read these, measure only their delta, in sequenceUnassigneddone when mechanism on the Spark + first research rerun (L071) using it, receipts show window id and shared calibrationLedger line 75 on GitHub
- L075Fix L003 note: names realrun.log which is not in the evidence folder (note link 404s)Unassigneddone when ledger note corrected or file pushedLedger line 76 on GitHub
- L079Open + merge site/allen, readme/allen-front, site/manifesto, site/discoverable when READY; Drake reviews Allen page and manifesto before mergePR deskdone when four merged shas, Drake said yes to /allen/ and /owned/Ledger line 80 on GitHub
- L095Integrate structs: merge feat/emit-structs then feat/omegese-structs, run twin gate on merged head, open PRsPR deskdone when two PR numbers + twin gate log on merged headLedger line 96 on GitHub
- L096Omegese import word -> op list imports section -> osc_compile_imports resolver (Omegese calling OSC and back); dispatch after emit-structs lands (same files)Unassigneddone when feat/omegese-import pushed; an Omegese fn calls an imported OSC fn with a pinned receipt; READY lineLedger line 97 on GitHub
- L110DRAKE DECISION 2026-10-19 (binding): the Omegese/Omega compiler exists to run this LLM; that is its priority, everything else second; it needs only that plus calculations and semantics. Lineage order (L107) and language features (structs, contracts, bytes/IO) are chosen for the LLM runtime first, the kernel last.Coordinatordone when L107 doc reflects the order; L097 chain doc names the host-side C that runs per tokenLedger line 111 on GitHub
- L111FUTURE (Drake 2026-10-19): decodable bitstream for TURING Yield: a real lossless coder in C whose output length equals the ideal code length T within a stated bound, so T becomes bytes on disk; not before the TY-7 energy rerun landsUnassigneddone when coder in omega src/turing with round-trip test; length vs ideal T bound; not started until L109 is doneLedger line 112 on GitHub
- L121spark-connector: merge connector/v0 then trust/v0, run install.sh on a clean user as the acceptance test, tag v0, link for DrakePR deskdone when two merged shas + tag + README linkLedger line 122 on GitHub
- L122Receipt chain (Drake: it sounds like a blockchain, use it): make the receipts shelf an explicit hash chain: every bundle carries the previous bundle digest for its repo and the node signs over both; the shelf git history (318 publish commits already) is the public ledger; aien-test verify-chain walks a repo chain end to end and refuses a rewritten history; a node proves inclusion by commit sha + path; no coin, no proof of work: the useful work IS judging PRs, the reward is the node name on every receipt it signed. Dispatch after L120 READY (same files)Unassigneddone when verify-chain in aien-test + chain fields in bundle format v2 (old bundles still verify) + trust/CHAIN.md; evidence: a rewritten-history test fails verify-chainLedger line 123 on GitHub
- L128SPEED-4 DFlash block drafter on the 4-bit model with tree verify (after SPEED-2) (L123 audit)Unassigneddone when qwen38 dflash gen code/prose n256 on W4 with tree verify: words per pass (fp8 today code 6.0, prose 2.29, L019) + wall tok/s; ON PAPER 68.7 code (calc:a497d66e), 26.2 prose (calc:065b7bde); output identical to verifierLedger line 129 on GitHub
- L130SPEED-6 mask-forced runs as one multi-row pass: when the Omegese mask forces k next tokens, feed them as a k-row pass (one weight read) instead of k single steps (after L112 + SPEED-2) (L123 audit)Unassigneddone when L112 G2 dev tasks: share of forced positions f (no receipt today), then wall tok/s mask-on with and without batching forced runs; gain ON PAPER 1/(1-f)Ledger line 131 on GitHub
- L132SPEED-8 shrink the non-weight GPU work (~5 ms/token at 4-bit, calc:bdf910fd): fuse glue kernels of the 596-launch q38 graph, cache x/residual/attention partials (L123 audit)Unassigneddone when GPU timeline per kernel class before/after on WW: non-weight ms/token (today ~5 of 71.68) and launches/token; ceiling of the lever 6.91%Ledger line 133 on GitHub
- L133SPEED-9 host side per token: measure then cut host C per token (detokenize, HTTP SSE, mask, completion poll 50 us sleep, logits copy) and keep the token loop on the GPU (L097 table feeds this) (L123 audit)Unassigneddone when per-token host microseconds table from aien-serve-http under load vs lab chip ms (server 7.63 vs lab 8.13 tok/s); lab host gap ~0.69 ms calc:cba6f39dLedger line 134 on GitHub
- L138R1 OSC-2 struct layout: every field takes a full 8-byte cell (docs/osc/OSC-2-DESIGN.md 2.3), so a u8 field costs 8 bytes of traffic; add a packed layout (fields at their natural width, declared alignment) for hot-path data with a version-byte bump; dispatch right after L095 merges (same files as L093/L094)Unassigneddone when packed layout in oscc + checker2 + omega_emit, twin gate still byte-identical, old units unchanged; READY lineLedger line 139 on GitHub
- L139R5 determinism tax: measure fixed-order (reproducible) vs fastest-order reductions on the decode matvec and attention sums, bit-exact check and timing receipt in a window; verdict: keep reproducible if tax < the stated bound, else propose the bounded-error alternative with its receipt; after L123 speed audit landsUnassigneddone when evidence/DETERMINISM-TAX receipt + verdict; READY lineLedger line 140 on GitHub
- L140R6 host-side C per token (sampling, mask, tokenizer, HTTP, scheduling): once L097 names each with microseconds, move the top item onto the GPU or into the engine loop (sampling on GPU first), gate bit-identical tokens and a timing receipt; dispatch when L097 READYUnassigneddone when evidence receipt before/after per moved item; READY lineLedger line 141 on GitHub
- L169#263 dossier for #276: per-gate table from the six campaigns plus the three follow-ups; after L064/L065/L066 reportCoordinatordone when docs/research/276-dossier.md on an arch branch, READY lineLedger line 170 on GitHub
- L170Entity prerequisites A to E (World restore from journal, VisorCortexView, ALLEN arrangement intents, canonical AgGraph encoding, Cortex multi-output records L086): one ADR-process proposal naming owners; no code before the measurements in L064 landCoordinatordone when proposal file on an arch branch with owners namedLedger line 171 on GitHub
- L179Fix: allen#1 body links AIENOS native Spark model path as aien-dev/allen#9/#327 because the transfer rewrote aienos#286 to allen#9; it should read aienos-mv#286 and #327 as docs/PHASES.md has. Issue text is coordinator-owned so this is a proposed fix only, to be applied via the desk L178agent fixerdone when allen#1 body correctedLedger line 180 on GitHub
- L186Horizon disk real astrophysics (Drake 2026-10-19): shadow radius 2.6 Rs, photon ring, inner edge at ISCO 3 Rs, T ~ r^-3/4 colour, Doppler beaming; on top of feat/goodbye-event-horizon 903277b7; science per feat/real-sky docs/ALLEN_SKY_SCIENCE.mdUnassigneddone when branch pushed, ui host test PASS with frames, READY line; each number cites its sourceLedger line 187 on GitHub
- L218Spark: installer + rollback-safe A/B slots on a blank disk (SLOT_LOADER, DISK_LAYOUT NOT_RUN)Unassigneddone when install from a recovery stick on a real machine with a forced-failure rollback; DISK_LAYOUT PASSLedger line 219 on GitHub
- L219Updater: signed AienOS update with rollback on every machine, no LinuxUnassigneddone when signed update applied and rolled back on hardwareLedger line 220 on GitHub
- L220Spark: power, thermal, fans, CPU frequency under no OS (D11)Unassigneddone when long hold at load with temperatures read; or written reason firmware holds themLedger line 221 on GitHub
- L221Spark: GB10 discovery and BAR0 ownership by the kernel (G1)Unassigneddone when BOOT_0 read after ExitBootServices on the real SparkLedger line 222 on GitHub
- L222Spark: GPU interrupts or written reason polling suffices (G4)Unassigneddone when a fault or completion interrupt seen, or a measured reasonLedger line 223 on GitHub
- L223Spark: GB-scale pinned and mapped memory, huge pages, 27B weight image (G5)Unassigneddone when one multi-GiB allocation mapped for CPU and GPU, in QEMU firstLedger line 224 on GitHub
- L224Spark: USB xHCI physical bring-up, keyboard (D6)Unassigneddone when KEYBOARD lines from the real Spark; glue power and clock state readLedger line 225 on GitHub
- L225Spark: Omega cut E, matmul, attention, launchers, session, resident runtime (G6)Unassigneddone when receipts against the Linux reference per kernelLedger line 226 on GitHub
- L226Spark: native CPU inference with aienos-infer on the Spark, speed against the memory ceiling (G7 step 10)Unassigneddone when tok/s and fraction of the measured memory ceilingLedger line 227 on GitHub
- L227Spark: 27B load from NVMe, KV on GPU, token stream (G7)Unassigneddone when a token stream from the real Spark under AienOSLedger line 228 on GitHub
- L228Gateway: HTTP to signed-transport translator, and the request path from another machine (G8)Unassigneddone when a request from another machine answered by the SparkLedger line 229 on GitHub
- L229Spark: move Atlas data and Forgejo off the TPM-sealed mounts, retire or relocate every Linux service (P3, U1 to U7, U10)Unassigneddone when service list in the Spark map, each row moved or retired with a receiptLedger line 230 on GitHub
- L230Network: signed-transport roster of N peers (add, revoke, rotate), stored in the StoreUnassigneddone when three-machine handshake matrix in QEMU plus one revoked peer refusedLedger line 231 on GitHub
- L231Network: production machine identity (owner-issued keys, TEST refused) for the signed transportUnassigneddone when handshake between two machines with owner-issued keys, TEST keys refusedLedger line 232 on GitHub
- L232Network: native TCP in native/netUnassigneddone when TCP echo and file-sized transfer in QEMU under loss and reordering, hostile corpusLedger line 233 on GitHub
- L233Network: TLS 1.3 with vetted primitives; list what native/crypto lacks firstUnassigneddone when TLS handshake against a reference implementation in QEMULedger line 234 on GitHub
- L234Network: native tailnet node or written decision that the Linux hub subnet router is permanent (WireGuard data plane, Headscale control protocol)Unassigneddone when an AienOS machine appears as a tailnet node, or Drake records the decisionLedger line 235 on GitHub
- L235Network: first native hub service, DNS over UDP on the Pi (Pi-hole role)Unassigneddone when DNS answers served by Aienux on the PiLedger line 236 on GitHub
- L236Network: network time over the signed transportUnassigneddone when wall clock set from a peer with drift measuredLedger line 237 on GitHub
- L237Network: signed control console on every machine (pattern: iMac port 47035 console)Unassigneddone when signed operator command round-trip on hardwareLedger line 238 on GitHub
- L238Network: off-LAN reach decision (relay vs hub-held Tailscale)Unassigneddone when Drake answers; if relay, a path through a relay in QEMULedger line 239 on GitHub
- L239iMac: VT-d DMA confinement (IM4)Unassigneddone when VT-d QEMU gate PASS (QEMU intel-iommu) then physical reportLedger line 240 on GitHub
- L240iMac: disk write path and Store v1 on the iMacUnassigneddone when Store write, flush and crash recovery on the real iMac diskLedger line 241 on GitHub
- L241iMac: one read-only lab boot listing xHCI ports, GPU VRAM size and memory controller, Thunderbolt state, RAPL watts (IM13, IM17, IM19, IM21)Unassigneddone when report lines resolving each UNKNOWN rowLedger line 242 on GitHub
- L242iMac: fan and temperature under a long native run (IM20, Drake's question 2026-10-18)Unassigneddone when fan speed and temperature read over a long run, plus a human listeningLedger line 243 on GitHub
- L243iMac: x86 AVX2 inference path and image target, first CPU token (IM26)Unassigneddone when token output equals the llama.cpp reference; tok/s against a measured memory ceilingLedger line 244 on GitHub
- L244iMac: Wi-Fi 14e4:43ba, Thunderbolt, SD reader drivers against chip models (IM16, IM18, IM19)Unassigneddone when model gates; firmware inventory with licencesLedger line 245 on GitHub
- L245Pi 5: network receive (no frame in 400 ms, PI9) and UDP echo, then first signed datagram Spark to PiUnassigneddone when an incoming frame received on the real Pi; signed datagram round tripLedger line 246 on GitHub
- L246Pi 5: UART input and RP1 xHCI keyboard so a person can type the unlock secret (PI5, PI6, PI12)Unassigneddone when unlock from an operator secret, no card fileLedger line 247 on GitHub
- L247Pi 5: one lab image printing started cores, a bandwidth loop, temperature and fan state, RTC value (PI2, PI13, PI14, PI-M3)Unassigneddone when report lines from the real PiLedger line 248 on GitHub
- L248Pi 5: multi-core inference, speed against a measured memory ceiling (PI-M3)Unassigneddone when tokens identical to the reference; tok/s as a fraction of the measured ceilingLedger line 249 on GitHub
- L249Pi 5: Wi-Fi, Bluetooth, GPIO and I2C and SPI, HDMI driver models (PI15 to PI18)Unassigneddone when documentation quote or DOCS SILENT before chip code; model gatesLedger line 250 on GitHub
- L251MacBook: tethered boots 1 to 4 (proxy read-only, kernel alive, framebuffer, disk read-only)Unassigneddone when pass lines per boot with the Mac returning to macOS unchangedLedger line 252 on GitHub
- L252MacBook: ANS2 NVMe, SMC, MTP keyboard, USB-C PD and DWC3 USB, USB Ethernet (D4, D7 to D10)Unassigneddone when model gates, then a UDP report from the MacLedger line 253 on GitHub
- L253PC: make the x86 kernel vendor-neutral on a clean base (Intel-only checks, x2APIC, NVIDIA display match)Unassigneddone when QEMU q35 with an AMD CPU model boots the kernel; iMac gates still passLedger line 254 on GitHub
- L254PC: AMD-Vi driver with IVRS parser against QEMU's emulated device (PD3)Unassigneddone when IOMMU state printed before any DMA; a device granted a range onlyLedger line 255 on GitHub
- L255PC: NIC driver for the chip the inventory names, plus the UDP report (PD5)Unassigneddone when one frame out of the PC seen on the LANLedger line 256 on GitHub
- L256PC: FSP boot-command encoder and GB202-family image extraction (PG2); second memory pool in the GPU simulator (PG3)Unassigneddone when host tests against driver structures; simulator vector add with two poolsLedger line 257 on GitHub
- L257PC: 12 GB model policy, stream a 27B or run a smaller resident model (PG8)Unassigneddone when token stream on the PC with a stated policy and speed against the ceilingLedger line 258 on GitHub
- L259FUTURE: AienOS as the type-1 layer under Windows on the PC (Aienux boots first, Windows runs as its guest): what x86 virtualization Aienux would need (SVM, NPT, device passthrough), what Windows needs to see; written as UNKNOWN rows with experiments; the only path that runs AienOS beside Windows without going through the Windows kernel (Drake: 'get around the Windows kernel')Unassigneddone when docs row with experiments; a QEMU-nested proof of Aienux starting a guestLedger line 260 on GitHub
- L269Follow-on for L268: switch the driver builder's local default to the Wormhole endpoint once L190 notes its port; receipt that a contract built through the pair equals one built through the 27B aloneagent cc-allen-driversdone when receipt on the L268 branchLedger line 274 on GitHub
- L270Wormhole cascade (Drake 2026-10-20): N drafters, not two; find the N on each side where added drafters stop paying (acceptance x latency), race against the pair; output identical to the verifier; dispatch when L190 has the pair endpoint, same worktree lineageUnassigneddone when receipt: tok/s and acceptance vs N, N* chosen by measurementLedger line 275 on GitHub
- L27130B expert prefetch (Drake: keep up with the timing of the weights already there): predict next-token experts one step ahead so the Heartbeat ring has them in flight before they are needed; one copy of the weights, measured hit rate and tok/s vs no prefetch; dispatch after L190Unassigneddone when receipt: prefetch hit rate, tok/s before/after, identical outputLedger line 276 on GitHub
- L286aien-train onchip probe 2: tensor-core 2-layer MLP training step resident on chip (persistent kernel, BF16 HMMA, grid barrier between layers, fused SGD-momentum) vs per-layer launches streaming weights; spec in aien-train docs/physics/gb10-training.md section 3.2; needs QMD 99 KB shared (#560) and a proven grid barrier (L272)Unassigneddone when quiet-measure receipt: step time, tok/s, fraction of 118.81 TFLOPS BF16, bytes/step, bit-identical weights both waysLedger line 291 on GitHub
- L291aien-train read-once trainer (Drake: read the weights once, then math, go): persistent Omega training kernel, model sized to on-chip storage, weights+grads+state resident per SM for the whole run, tokens stream in, DRAM touched once at start and once at end; measured vs streamed design on same model; dispatch when L279 Q1 size + Q3 spec are pushed (same kernel files)Unassigneddone when trainer runs N steps on GB10 with DRAM weight reads = 1 (measured), receipt, loss curve matches streamed design, READY lineLedger line 293 on GitHub
- L290Crypto: AES-128 block, AES-128-CMAC (RFC 4493), AES Key Wrap (RFC 3394), SHA-1/HMAC-SHA1/PRF-SHA1, PBKDF2-HMAC-SHA1, P-256 + SAE for Wi-Fi WPA2/WPA3 (L206 refuses password networks until these exist; see aienos docs/SPARK_WIFI_MT7925.md sec 5). Overlaps L233 (TLS) on AES-128Unassigneddone when each primitive in native/crypto with RFC/NIST test vectors, constant-time review, wired into wpa_crypto_native; L206 mt7925_test station_wpa passes with the native tableLedger line 295 on GitHub
- L294omega training kernels 1: backward GEMMs on tensor cores (dX = dY W^T, dW = X^T dY) BF16 HMMA and FP8 QMMA with transposed-operand fragment loads, bit-checked vs CPU reference; dispatch when L279/L283/L284 exit (box load)Unassigneddone when omega train/backward-gemm: both products PASS vs reference on 12 shapes incl. 768/1024/2048, measured TFLOPS fraction of 118.8 BF16 / 237.5 FP8, READY lineLedger line 299 on GitHub
- L295omega training kernels 2: GPU-tier reverse-mode gradients for the transformer block (attention, MLP, norms, embedding, cross-entropy head without materialising the full logits table) over the M21 CPU autodiff as oracle; dispatch after backward GEMMs landUnassigneddone when omega train/gpu-autodiff: block gradients match CPU oracle within tolerance on 3 sizes, receipt, READY lineLedger line 300 on GitHub
- L296omega training kernels 3: fused optimizer step on chip (AdamW and Muon) applied in the backward pass, master weights FP32, state bytes per parameter measured; dispatch after gpu-autodiffUnassigneddone when omega train/fused-optimizer: 125M-class step runs end to end on GB10, loss decreases on held-out slice over 1000 steps, receipt, READY lineLedger line 301 on GitHub
- L300Pull Aienux (the kernel) out of aienos-mv into its own repo aien-dev/aienux (Drake 2026-10-20: after we are done): native/kernel (434 of 3,582 files) plus the native/ parts it owns (boot, crypto, disk, net, store, capability, sig to be decided by dependency), history preserved (filter-repo), aienos-mv consumes it by a pinned lock like omega.lock; 15 outside references (crates/aienos-kernel continuity, aienos-infer-kernel, allen-reflex, docs, .crumb) repointed; judge lane + 48 local clones + CI refs moved. DISPATCH TRIGGER: naming/aienux-aienos branches merged (L180) and the replace-Linux items L206, L207, L225 landed, so the split does not fight 272 renamed files and live driver branchesUnassigneddone when aien-dev/aienux exists with history, CI green there; aienos-mv builds from the lock; judge watches aienux; READY lines both reposLedger line 305 on GitHub
- L301DRAKE DECISION 2026-10-20 (binding): Unix rule, one thing per repo. Kernel = aien-dev/aienux (machine-independent + the driver port); each machine its own repo (aienux-spark, aienux-imac, aienux-macbook, aienux-pc); each driver its own repo (aienux-driver-<device>); aienos assembles by pinned locks like omega.lock. ADR 0039 on aien-architecture adr/0039-one-thing-per-repo@3495907e8775b54fdbec66d3b901b60d61dfd258. In-flight drivers finish in place and move at the split (L300); new drivers start in their own repoCoordinatordone when ADR merged; L300 scope updated; driver port spec published in aienux (gate for driver repos)Ledger line 306 on GitHub
- L31035B decode to the memory ceiling after the swap (Drake 2026-10-20, seeing a 27B at 149 tok/s on a 1.8 TB/s card): count active bytes per token on the 35B-A3B (shared + routed experts + attention + head) at FP8 and at W4/own-FP4, state the GB10 ceiling tok/s from the MEASURED 207-237 GB/s, then W4 experts + MTP (L190 heartbeat ring, L308 placement) one lever per run, identical tokens, until tok/s is within 0.8 of that ceiling; dispatch when L306 swaps 18091Unassigneddone when docs/speed/Q36_CEILING.md with the bytes/token calc + receipts per lever on q36/serve, tok/s >= 0.8 x ceiling or the measured reason, READY lineLedger line 315 on GitHub
- L311L308 follow-on H6 one base, many heads, one pump stroke (Drake 2026-10-20): one weight read per beat feeds N heads at once (N sequences; N candidate continuations via the heartbeat ring; MoE expert-union growth with N); tokens per stroke and tok/s at the knee vs the ceiling calc, identical tokens per head; dispatch when L308 posts its placement plan + tier table (same ring code)Unassigneddone when evidence/typewriter/h6/ on feat/typewriter-experts: tok/s and tokens-per-stroke vs N with bytes/beat flat as receipt, knee N named, verdict, READY lineLedger line 316 on GitHub
- L315HUM lever: route-aware stroke packing: when more heads wait than ride one stroke, pick the riders whose routed experts overlap most so the expert union per layer (8 -> 57 at N=8 by independent draws) shrinks; measured on L308's real per-domain routing traces; DISPATCH WHEN L308 posts its H1 routing traces (same data, do not regenerate)Unassigneddone when evidence/hum-packing/: union per layer and bytes per stroke packed vs FIFO on real traces, tok/s gain at N=8,16,32 (calc), READY lineLedger line 320 on GitHub
- L317Self-hosting: build, judge (aien-test), git and agents run on AienOS on the Spark itself (Drake 2026-10-20, L203). Path: Omega toolchain + judge as AienOS programs; Linux stays the oracle until three-way equalUnassigneddone when judge PASS bundle produced by a judge running on AienOS for one real PRLedger line 322 on GitHub
- L344Hazard found 2026-10-21 01:2xZ: /home/drakestapleton was itself a git repo (git init 2026-10-20 20:40Z by an unknown agent, one commit of the home .crumb listing, origin = aien-dev/synapse). Every git command in a non-repo dir under home acted on it (git add -A could commit home files, git push would push to synapse). Coordinator moved ~/.git to ~/.git.disabled- (reversible, nothing deleted); commit 8e0eb27 not on the remote. Find which agent did it (transcripts around 20:40Z) and fix its scriptagent fixerdone when agent + command identified, its script fixed, ~/.git.disabled kept or removed by DrakeLedger line 349 on GitHub
- L347L332 rung 3 trigger: run the GB10 driver core (native/kernel/gpu + contracts) inside Aienux native on the Spark once L318 makes native boot realCoordinatordone when waits on L318 PASS; then a brief for the native port of aien-gpu's coreLedger line 352 on GitHub
- L350spark-receipt line 30 writes start time as local time with no zone (date +%Y-%m-%dT%H:%M:%S); readers assume UTC (found by L125). Fix: date -u ...Zagent fixerdone when spark-receipt writes UTC with Z; one receipt checkedLedger line 355 on GitHub
- L361spark-window line deadlocked: SIGSTOPPED L321 waiter pid 1810064 holds ~/yardsticks/windows/.state.lock; SIGCONT or stop it (details restart-requests.md 03:18Z)Coordinatordone when window line moving: spark-window status shows a held windowLedger line 366 on GitHub
- L379Train on Drake-voice traces (own stack: reuse aien-train / omega-train-onchip; target per ledger model decision; intent/anger predictor + voice SFT)Unassigneddone when dispatch when dataset item READY and CHIP-RESERVED lifted; done when: trained adapter/model + eval receipt pushed (no weights committed)Ledger line 384 on GitHub
- L388AienOS ALPHA (Drake 2026-10-21: out of pre-alpha, order motor -> Allen -> alpha): one tagged release with the bundle-1 Allen (L040) inside every image: PC VM/live images (L260), aienos-run Windows/macOS/Linux (L261/L265/L262), Try page + try.sh (L264), README; merged by Renda on the motorCoordinatordone when DISPATCH TRIGGER: L040 hardware report PASS. Done when: GitHub release tag published with signed manifest, try.sh boots it on the Mac + Linux, Try page live, NOT_RUN rows namedLedger line 393 on GitHub
- L399aienos.com /omega/: cudaplusplus name (never plus signs; verify check) + publish the motor (OMEGA.md section 6, /omega/#motor) + section 5 performance table and P0-P4 (prompt Mac ~/handoffs-mac/briefs/site-omega-motor.md)agent hermes-site-motordone when branch READY with Mac suite+build PASS and screenshots; live after desk mergeLedger line 404 on GitHub
- L410Compatibility build: omc emits ELF shared objects on the C ABI; Python/Rust/Go/Ruby/Node/Fortran call Omega functions with results equal to Omega-native on the corpus; door refuses overlapping/out-of-shape buffers (rungs from L409 COMPATIBILITY.md)Unassigneddone when branch with each language's call gate RC=0 + negative controls in evidence/; dispatch when L408 R5 (omc fixpoint) READY and L409 READYLedger line 415 on GitHub
- L414Decision 34 experiment (L409 vs L386): GPU kernel words per tile vs per thread. Same kernels (tiled GEMM, decode attention, reduce) written both ways in Omegese; per-tile words lower to the same checked IR. Metric: bytes/s as a fraction of the GB10 bandwidth wall, output bit-identical, source tokens the AI writer needs, refusals. Winner becomes decision 34; loser recorded with its numbersUnassigneddone when evidence/D34-TILE-VS-THREAD/ with both arms' numbers under one quiet window + verdict line in docs/PROGRAM.md decision 34; dispatch when L386 C1 merges (omegese.c busy) and the chip reservation for the train liftsLedger line 419 on GitHub
- L415PR for study/languages@b7859ab0 (L409): flip 26-33,35-45 PROPOSED->ACCEPTED on merge, 34 stays PROPOSED (L414)PR deskdone when PR number mergedLedger line 420 on GitHub
- L419C->Omega wave 3 (program L417): src/runtime + GPU host code + CUDApp (84k) + tests, built by seed omc; needs FR-4 floats, FR-5 device syscalls, FR-6 threads, FR-7 shared objects, R6 GPU words; coordinate with L395 openCUDA and L386Unassigneddone when per-artifact equivalence RC=0 (bit-identical outputs incl. chip results), C removed + c-oracle tags; dispatch when L418 READY wave 1 and FR-4..FR-7 DONELedger line 424 on GitHub
- L420C->Omega wave 4 (program L417): physics (50k) + aienux kernel (110k, arm64 + x86-64) + atlas/benchmarks/interplane, built by seed omc; needs FR-8 privileged/MMIO words and FR-9 x86-64 back end; QEMU and real-board gates as todayUnassigneddone when boot gates (QEMU + Spark/Pi/iMac staged image) identical to the C build, C removed + c-oracle tags; dispatch when FR-8 and FR-9 DONELedger line 425 on GitHub
- L423Interpreter sync: tools/omega_ref (L412) follows the spec gap verdicts G1-G25 and FR-10..FR-12 once L408 lands themUnassigneddone when seed/omega-ref: probe_ref, gate_ref, gate_next, gate_ddc RC=0 at the omc sha that marks FR-12 DONE; dispatch when FR-12 DONELedger line 428 on GitHub
- L425FREEZE omega: main unmoved 78h, 50 open PRs, no judge in 6h (desk-freeze-watch)Coordinatordone when queue moves: a merge or a judge run on the repoLedger line 431 on GitHub
- L426FREEZE aienos: main unmoved 59h, 21 open PRs, no judge in 6h (desk-freeze-watch)Coordinatordone when queue moves: a merge or a judge run on the repoLedger line 432 on GitHub
- L427FREEZE aien-architecture: main unmoved 37h, 5 open PRs, no judge in 6h (desk-freeze-watch)Coordinatordone when queue moves: a merge or a judge run on the repoLedger line 433 on GitHub
- L4292026-10-22T05:51Zagent L389done when v2Ledger line 435 on GitHub
- L4302026-10-22T05:51Zagent L389done when v2Ledger line 436 on GitHub
- L431v2 training complete; gcn_interp PASS; merged student readyagent L389done when v2 (runs/v2): LoRA rank-16, 588 steps, 2 epochs on train (2314 rows), val (32 rows); val_acc 0.683, val_nll 1.393 (v1: val_nll 1.82->1.39). Adapters 118M, merged model 1.5G (v2.agp1 1.5G). eval: student handoff P=0.775/Rec=0.969 vs base P=0.0/Rec=0.0, teacher 35B P=1.0/Rec=1.0 (91 prompts). gcn_interp: 56 prompt tokens, 33 generated, tokens PASS bit-exact vs host reference (mismatch=) in 165.3s (interp) vs 31.6s (mirror), 40619 total launches. Adapters + merged in runs/v2/ (chmod 700, private). GCN worktree aienos-l389-gcn HEAD 55493c4. iMac GPU deferred per brief (bundle 1 first). Next: eval gen_student_v2 on more prompts.Ledger line 437 on GitHub
- L432FREEZE aien-sovereign-core: main unmoved 44h, 50 open PRs, no judge in 6h (desk-freeze-watch)Coordinatordone when queue moves: a merge or a judge run on the repoLedger line 438 on GitHub
- L441XORWOW: what the page-5 listing tunes, and whether an output mix beats itagent hermes-toolkit-directdone when Wolfram receipt plus an Omega check if a mix wins on the named metricLedger line 447 on GitHub
- L442CUDA++ on hardware that is not the Spark: RTX 5070 is the same family (sm_120), measure hello_gpu there; Apple M2 is a different family, needs its own encoder (PORTABILITY.md sec 5)Unassigneddone when receipt: hello_gpu PASS on a named non-GB10 chip, or a new-encoder branch with the chip row and the first failing probeLedger line 448 on GitHub
- L443CUDA++ portable native CPU try: reuse L436 c1a3591, fix Windows build/host API edges, package macOS ARM+x86, Linux and Windows without NVIDIA dependencies; no new GPU qualification claimagent hermes-portable-releasedone when source branch release/portable-alpha pushed with tests+receipts, downloadable native CPU packages executed on available hosts; Windows runtime NOT_RUN explicitly; READY line for RendaLedger line 449 on GitHub
- L445CPU speedup sc#646 is not implemented anywhere: top-K by exact partial select (crates/aien-serve-http/src/sampling.rs, 8.25 -> 0.92 ms per sampled token at K=20/152k logits, tokens identical). main holds only the full-sort top_k_mask (461-line file); the live 18091 branch holds only the full sort too, and the live server samples top_k=20 on every turn. It was carried by desk block #701 head e1d0a469 and dropped from the replacement block #703 (02ff6bca, carries only #574/#577). Head e75e240d conflicts with main 1e6ae85c in .crumb only. Audit ~/handoffs/2026-10-22-speedup-audit.mdCoordinatordone when rebase the crumb conflict, carry sc#646 in the next desk block, judge PASS, merge sha quoted; then live 18091 runs a build whose sampling.rs has rank_cmpLedger line 451 on GitHub
- L446GPU speedup omega #515 variant A coalesced fp8 GEMV stores is not on omega main: origin/main src/omega_gpu_fp8_api.c has zero COAL occurrences, while the live build branch ship/motor-l387 has omega_gpu_fp8_coal() defaulting AIEN_OMEGA_FP8_COAL to 1. The win (OWN-30B 24.49 -> 21.34 ms/token, 43.0 -> 49.2 tok/s, tokens identical, 10/10 mutants, calc:3fae8fa6) therefore runs in production from a branch that main does not contain. Head cfb7b5fd conflicts with main a8ffd2b9 in .crumb only. Audit ~/handoffs/2026-10-22-speedup-audit.mdCoordinatordone when omega main contains AIEN_OMEGA_FP8_COAL with default 1 in src/omega_gpu_fp8_api.c and the merge sha is quoted; branch and main agree on the codeLedger line 452 on GitHub
- L447GPU speedup omega #490 tern/kt1-speed (coalesced output + prefetched first weight loads, 3 of 4 cold gates pass, head 2048fc71) merges clean into main a8ffd2b9 with no conflicts; it is OPEN and BLOCKED with no judge run. Audit ~/handoffs/2026-10-22-speedup-audit.mdCoordinatordone when judge PASS on head 2048fc71, CLEAN, merge sha quotedLedger line 453 on GitHub
- L448Desk: land the three unlanded speedups found by the 2026-10-22 audit (parent items L445 CPU top-K sc#646, L446 GPU coalesced stores omega#515, L447 omega#490 kt1). Rebase check done read-only today: sc#646 head e75e240d conflicts with main 1e6ae85c in .crumb ONLY; omega#515 head cfb7b5fd conflicts with main a8ffd2b9 in .crumb ONLY; omega#490 head 2048fc71 merges CLEAN with main a8ffd2b9 and needs no rebase. #646 was dropped from the desk block sequence (carried by #701 head e1d0a469, absent from #703 02ff6bca) rather than judged. Audit ~/handoffs/2026-10-22-speedup-audit.mdPR deskdone when the three PRs carry a PASS judge verdict on their rebased heads and a merge sha is quoted for eachLedger line 454 on GitHub
- L449Serving context for local coding work: 18091 runs --max-context 49152 (start-35b-batch.sh, Q36_MAX_CONTEXT default), which is below the 64000 that Hermes Agent requires of an auxiliary compression model. A local Hermes session on the motor therefore dies with ValueError as soon as it needs compression, which is any session that reads a real file. Level is env-settable: Q36_MAX_CONTEXT. Raising it costs KV memory and needs one live-server restart under the documented swap procedure (pid file ~/.local/state/aien-q36/serve.pid, TERM only, wait for 'shutdown complete'). 18091 was serving another agent at 12:10Z, so this was NOT restarted now.Coordinatordone when 18091 reports a context >= 65536 and a local Hermes session on it completes a turn that reads a 500-line file, log path quotedLedger line 455 on GitHub
Settled
Decisions made, with their evidence
A line closes on evidence (a commit, a receipt, a file) or is dropped for a stated reason. Newest first.
- 39 min agodoneL437
Host Omega slices for the next CUDA toolkit gaps: cuRAND XORWOW step, CUB scan, cuBLAS GEMV, n=4 Cholesky, 2x2 singular vectors
feat/fft-rng@37b0c09d00f9b38aba8f19ccf90e9b4c40c69540 evidence/fft-rng/more.log
Ledger line 443 on GitHub - 58 min agodoneL435
R5b+birth: seed-born Omega is the language (Drake 2026-10-22, this chat). Names pass in om4 and omc, then make omega-from-seed leaves seed/out/omc. Parent L408. Codex L408 worker dead (usage limit, log frozen 02:06 CDT, pgrep empty). Do not relaunch L408 onto ~/workspace/omega-in-omega.
seed/omega-in-omega@29b12760f4c962606411d6bf25032aedef063301 gate4 RC=0 gate5 RC=0 make omega-from-seed RC=0 seed/out/omc sha256 191b776ef98821597c3e30655da6d680a9b66edbc9ceaca3c39c61ea6df098e9 ~/handoffs/2026-10-22-omega-seed-birth.md
Ledger line 441 on GitHub - 7 hours agodoneL400
CUDA++ developer experience (Drake 2026-10-22): research CUDA pain points (Reddit, GitHub, forums; why people cannot leave), docs/PAIN_POINTS.md in aien-dev/cudaplusplus, a D-bar that beats CUDA (one command, day-one chips, faults named at the cause, host sim, determinism, work per answer, no licence trap), build top rows, CONTRIBUTING + issue drafts; framing: less computation away from big data centers. Coordinates with L395 (repo owner), L386, L394, L332. Prompt Mac ~/handoffs-mac/briefs/cudapp-dx.md
CHIP-RESERVED now admits L400. The host halves are already on CUDApp main via #1. Chip halves can take a value window; no new worker started here.
Ledger line 405 on GitHub - 7 hours agodoneL424
CUDA++ unblock: consolidate finished illustrated docs and DX onto current CUDApp main, run host gates once, identify desk enrollment gap and request bounded non-disruptive chip slot from existing coordinator; no duplicate workers or driver reboot
CUDApp#1 merged 7da2853030e4bf6f45703c4c92cd66a502093794 from 780961cdfe68ba8834bcf535fdf81fe4bd638bae; hello PASS window 20261022T044350Z-0d494e
Ledger line 430 on GitHub - 8 hours agodoneL421
Fix: flaky host test qwen36_batch::tests::row_swap_mutant_is_caught (sc ship/motor-l387 dae01e2e = live bin-19173f9e source): fails ~1-3 in 40 runs ('the swap changed N conversations', N < 2)
43a4f95b on fix/row-swap-flake; 0/200 fail without features, 0/200 with aien-test-switches (was 2/200); crate tests green both ways
Ledger line 426 on GitHub - 8 hours agodroppedL387
Motor ship build (Drake 2026-10-21 hurry: motor -> Allen -> alpha): every READY speed lever (L369 prefill, L312 strokes 3-7, L314 exact head, L313 bf16 state, L375 prefix) in one merge-train build, train PASS, timing receipts, swap 18091; brief ~/handoffs/briefs/L386-motor-ship-build.md
from L376: next train please take serve/q36-board-ship a77d6798 (fast-forward from ship/motor-l387 dae01e2e = live), bin-a77d6798 pinned; suite arm board enabled + case board-gate on the branch
Ledger line 392 on GitHub - 8 hours agodoneL369
35B prompt reading (prefill) at chip speed on the batch door: today prompts are fed one token per beat through the decode graph (Hermes: 261-token prompt = 5.9 s, 22.7 ms per prompt token); target chunked prefill (prompt chunk as a GEMM, DeltaNet chunked scan, attention over the chunk)
shipped live 2026-10-22T03:17:50Z: 18091 = bin-19173f9e (sc ship/motor-l387 dae01e2e, omega ship/motor-l387 6de999da). Merge train 20261022T0244Z-bfc41659 L369 PASS (goldens, concurrent-32, turn2, l369-long-prompts 37 identical); live gate2 PASS 32/32 + turn2; prompt read on live 22.696->211.23 tok/s (~2.5k), 19.126->156.09 (~8k), 27.124->108.85 (~21k), shared machine, receipts evidence/l387-motor-ship/README.md
Ledger line 374 on GitHub - 8 hours agodoneL405
Stop the lock pile-ups (Drake 2026-10-22: stop chip locking): correctness work never waits behind timing locks; exclusive quiet/chip leases capped at 20 min; test servers use their own GPU lock name; idle-aware timing windows; watchdog on overruns. Incident: L371 window held quiet exclusive 3 h, 30+ builds/tests queued 1-3 h at load 3.5; L371 test server held gb10 exclusive 6.9 h
tools installed 20261022T011917Z (quiet-build v3 01:25:31Z), backups *.bak-*; harness old FAIL 8 / new PASS 11; docs/lock-policy-l405@cc8e6f0a READY; 2 h: 455 builds median 0 s max 16 s
Ledger line 410 on GitHub - 8 hours agodoneL413
Fix om4 (L412 F1+F2): ulong numbers 2^63..2^64-1 refused (ulong 9223372036854775808 -> error 8, 18446744073709551615 -> error 6; OMEGESE s9.3 makes ulong 64-bit); and 'view x' passed to an edit parameter gives error 7, s9.4 says 20. Repro: tools/omega_ref/probe_ref.sh cases conv_sign cmp_unsigned write_wrap view_pfx_to_edit on seed/omega-ref@e6b27a2c; details tools/omega_ref/SPEC_GAPS.md. seed/ is L408's lane: L408 fixes or proposes.
dedfd93c257b2e9935f5da2e416c00fe1fec0fa0
Ledger line 418 on GitHub - 9 hours agodoneL409
Languages study + compatibility design (Drake 2026-10-22): every major language (Fortran..Ruby, Python, Rust, Go, GPU languages) what it got right/wrong from designers' own words, decisions 25+ for Omega; design for Omega callable from/calling all languages via the C ABI (rules, not C code), rung list with gates
study/languages@b7859ab081f098f68536759c96bb086b1c1621ea, ~/handoffs/2026-10-22-languages.md; LANGUAGES.md 407 sources, COMPATIBILITY.md 109 sources, 472/472 quotes re-fetched PASS; verdicts 26-45 in L409 notes (19 ACCEPTED, 34 -> L414)
Ledger line 414 on GitHub - 10 hours agodoneL407
Fix: gpu_attention_test --hd 128 --host-only fails 16 checks on main (paged_bf16 + gqa_f32 'kernel generated' + nvdisasm listing lacks MUFU.RCP)
no new fix needed: open PR #496 (4e9cfbe7) verified red->green (107 checks, 0 failed at hd128); handoff ~/handoffs/fixer-attn-hd128.md
Ledger line 412 on GitHub - 10 hours agodoneL089
Omega runtime R1 API spec v0 + manifest + version check (omega-mv#594), branch runtime/r1-api
runtime/r1-api@164fb76f READY; test-runtime-api PASS 9/9 negative controls; finished under L386
Ledger line 90 on GitHub - 11 hours agodoneL406
Fix: omega test-gpu-attention FAIL on main a8ffd2b9 (--hd 128 --host-only: paged_bf16 84 gprs > 64 cap; nvdisasm listing lacks MUFU.EX2/RCP/FMNMX)
no new fix needed: already on open PR #496 (fix/470-attention-hd128-host @ 020f059a), awaiting merge
Ledger line 411 on GitHub - 11 hours agodoneL404
Literature-to-code audit for CUDA++ performance, with Wolfram calculations: MoE chunk/tile reuse, Gated DeltaNet, attention IO and honest profiling bounds; feeds L403/L369
Report + inputs + baseline pushed 8e98a9e1002c4537fa9302e39a75d87db2405200, remote ref read-back matched; four primary papers with verified quotations, citation gate PASS; handoff to L403/L369 and READY read back.
Ledger line 409 on GitHub - 12 hours agodoneL397
Fix: quiet-build --help (and -h) does not print usage: it queues behind the quiet lock and then runs '--help' as a command (flock -s /tmp/aien-quiet.lock --help); agents calling it to read usage hang (L396 hit it, pid 4139762 waited 55 min)
quiet-build --help/-h/help/no-arg prints usage, exit 0 in 2 ms, before any wait or lock; normal run still works; backup ~/.local/bin/quiet-build.bak-20261021T235537Z
Ledger line 402 on GitHub - 12 hours agodoneL393
aienos.com /omega/ page: what Omega is (language for all computing; CUDA replacement), CUDA layer map, live no-CUDA receipt, bar C1-C6 checklist; other Omega pages agree (prompt Mac ~/handoffs-mac/briefs/site-omega-page.md)
aienos.com site/omega-what-it-is@d274e4a2 pushed, READY in desk log (merge after omega docs/what-omega-is@237c2095); test 129 pass, build PASS; shots ~/handoffs/shots/L393/
Ledger line 398 on GitHub - 14 hours agodoneL391
Fix: omega host test CMDLIST_NEAR_CAP_SUBMITS_WITH_L2_INVAL_PENDING fails on ship/motor-l387 6de999da (st.submits == 2 expect, test_omega_gpu_platform_linux.c:234)
fix/cmdlist-near-cap 20f412df pushed; make test-gpu-session-platform all PASS (test expectation was stale after MAX_LAUNCHES 2048->4096)
Ledger line 396 on GitHub - 14 hours agodoneL385
SWAP 18091 to bin-16b0e6f8 (L354 batch strokes v2): Q36_BATCH_BIN=bin-16b0e6f8 AIEN_Q36_STROKE=1 ~/.local/state/aien-q36/start-35b-batch.sh (8 slots x 131072 as today; the script's AIEN_Q36_STROKE default is 0, so pass 1). Same tools + AIEN_SERVE_GPU_LOCK as bin-61cad330 (16b0e6f8 contains q36-tools 61cad330). Gate: real8 32/32, real32 32/32, swap control red (evidence sc serve/q36-batch-v2 9c62690a evidence/q36-batch-20261021/README.md). Timing 32 slots: N=8 57.1 vs 31.8 tok/s rows in turn (1.80x calc:3658d8e7), N=16 1.98x, N=32 1.88x, N=1 equal. Back: AIEN_Q36_STROKE=0 (same binary, rows in turn) or Q36_BATCH_BIN=bin-61cad330.
18091 serves bin-16b0e6f8 AIEN_Q36_STROKE=1 since 21:46Z (Hermes, Drake chat). gate2 on live: gate-batch.sh 18091 18091 -> 32/32 identical, turn2 identical, prompt0 == host golden, PASS (sc-q36-batch evidence/q36-batch-20261021/live-16b0e6f8-1021T2146Z/verdict.json); get_weather tool call parsed. Markers /tmp/aien-q36-live-restart + live-down-for-tests removed: q36-extra back to 1 slot, queued tests of the stopped agents now admitted one at a time.
Ledger line 390 on GitHub - 22 hours agodoneL381
Fix: tool-call XML parser (serve/q36-tools tools.rs) truncates parameter values containing closing tags, coerces string params by JSON shape, drops malformed calls silently
fix/tool-call-parse 5924342d pushed (parser keeps closing tags in values, types by schema, reports dropped calls); lib 48 ok + loopback 16 ok, clippy/fmt clean, verified by L376; handoff ~/handoffs/2026-10-21-fixer-tool-call-parse.md; READY on BOARD 14:14Z
Ledger line 386 on GitHub - 22 hours agodoneL382
Fix: omega tests/gpu_cmdlist_test.c does not compile (-Werror missing-field-initializers: OmegaDecSpec cases list 5 values, struct has 6 incl. groups) on origin/motor/serve, 0bef2e59 and serve/q36-batch-stroke
332bee05 + test PASS
Ledger line 387 on GitHub - 22 hours agodoneL378
Drake-voice dataset: join all Wispr dictations (13,206, ~/handoffs/wispr/history.jsonl) to agent traces (Mac+Spark Claude Code, Hermes, agent-to-agent), refreshed catalog to today; per dictation: what agent was doing before, Drake words (raw asr + formatted), agent reply, downstream agent coordination; labels anger/praise/correction/intent; PII scrub; private, never published
~/handoffs/wispr/traces/ (traces.jsonl, pairs/, STATS.md, README.md), ~/handoffs/wispr/HANDOFF.md, ~/handoffs/wispr/TRAINING_PLAN.md, code agent-archive@wispr-traces c5957a5
Ledger line 383 on GitHub - 23 hours agodoneL283
aien-train math-training (Drake structure-constants / let-math-deal-with-it idea): E1 frozen features + direct ridge head solve vs gradient training, E2 structured adapter basis vs LoRA, E3 closure of a weight algebra (structure constants) on real layers, E4 invariant activation subspace across nonlinearities; each PROVEN/PARTLY/RULED OUT with numbers
13:19Z coordinator: CONSOLIDATED (Drake: fewer Claude workers). Agent ended after 11 h paused; item stays open in the research backlog; resume with: claude -p --resume 9f228758-7e50-4ae2-8bcb-32a0d2e2fcab (coordinator only).
Ledger line 288 on GitHub - 23 hours agodoneL380
Fix: aien-omega-q38 q36.rs fails to build without AIEN_OMEGA_GPU_LIB (unresolved crate::session_err, motor/serve)
14c89f6b + both checks green (stub and with lib, 0 warnings)
Ledger line 385 on GitHub - 23 hours agodoneL377
Fix: omega Makefile skips check-physics-lock for explicit targets (make OUT_DIR=x x/libomega_gpu.a) and never rebuilds nvrm.o/m16_native.o when PHYSICS_DIR changes: L313 linked a stale nvrm.o from ~/workspace/physics e95e3ed (cached GPFIFO) -> 'gdn alloc: marker wait timed out (latched)' on the chip
ddcbd8a2 (q36/serve) + 1456883b (main), tests/physics_lock_make_test.sh
Ledger line 382 on GitHub - 24 hours agodoneL366
URGENT 18091 DOWN since 03:44:32Z: kernel OOM killer took the live 35B (bin-0915000b pid 1497185). Restart with start-35b.sh (or start-35b-batch.sh = swap L363). Root cause: test servers and gate tools each check MemAvailable >= 45 GB then load a full 35B (~37 GB); two passed the check at once (L354's 18094 server at 03:43:53Z + another 35B load) and the kernel killed the biggest process, 18091. Proposed fix (cross-lane, coordinator's call): one admission lock for every extra 35B load beside 18091, e.g. flock /tmp/aien-q36-extra.lock held for the life of a test server or gate tool, check MemAvailable inside the lock; and oom_score_adj -500 for 18091 in start-35b*.sh so the kernel never picks the live server first.
L355: gen-p4 near-tie check done without a full-35B load (windowed qwen38_omega_full, chip lock): base and my branch identical, flip native to chip FP8 at margin 0.010. My q36-extra slot request is withdrawn. Note for the rule: q36-extra holds the slot while it waits for memory, so a job whose own guard needs more than the 48 GB check (q36_levers needs 37.3+12 GB) can block the line; I stepped out of it.
Ledger line 371 on GitHub - 24 hours agodoneL370
q36-extra admission: since 18091 runs the batch door (bin-99230310, 8 rows x 131072 = 20.8 GiB on paper, calc:252e20b5) MemAvailable sits at 46.7 GiB (11:40Z, RSS 18091 57 GB), below Q36_EXTRA_NEED_GB 48: no extra 35B can load (L354 stroke gate bin-13f06473 waiting in q36-extra since 11:38Z; L354 extra needs ~38 GB: weights + 8 rows at ctx 2048 + exact-head screen). Decide: lower the need (e.g. 44) for extras that declare a small ctx, or other.
need 42 GB, oom ordering protects 18091
Ledger line 375 on GitHub - 24 hours agodoneL365
SWAP 18091 to bin-d839a8ca (L356 levers 1+2: a|b overlap ON + greedy replies take the chip token, AIEN_SERVE_CHIP_GREEDY=0 off) gate2 PASS on 18096, mutant 4 FAIL, 200-token reply identical to live; evidence sc levers/q36-quick 6b5c14ab evidence/levers-q36-20261021; start: Q36_BIN=bin-d839a8ca start-35b.sh; rollback bin-0915000b
handled 11:25Z: 18091 restarted (07:25Z) then swapped to the batch door bin-99230310 (L363); OOM rule q36-extra + oom -900 for 18091; exact head + levers to ship inside the batch build (L354)
Ledger line 370 on GitHub - 24 hours agodoneL367
SWAP 18091 to bin-409c8ffa (L314 exact two-pass head + L356 levers a|b overlap and greedy chip token, all ON; supersedes L365's bin-d839a8ca, which it contains) gate2 PASS on 18093, mutant 5 (MASK inverted) FAIL, ON == OFF ids 5/5 (4 greedy x128 + 1 sampled); greedy chip 21.8 vs 24.0 ms/token. 18091 is down (L366): restart it on this bin with start-35b.sh (Q36_BIN=bin-409c8ffa or BIN file). Evidence sc hum/exact-head-levers@28d04d87 evidence/hum-exact-head/serve-combined/. Off switches: AIEN_Q36_EXACT_HEAD=0, AIEN_OMEGA_Q36_OVERLAP_AB=0, AIEN_SERVE_CHIP_GREEDY=0.
handled 11:25Z: 18091 restarted (07:25Z) then swapped to the batch door bin-99230310 (L363); OOM rule q36-extra + oom -900 for 18091; exact head + levers to ship inside the batch build (L354)
Ledger line 372 on GitHub - 24 hours agodoneL363
SWAP REQUEST (L354, Drake #1 many conversations): put the 35B batch door on 18091: stop the live server (bin-0915000b), run ~/.local/state/aien-q36/start-35b-batch.sh (bin-99230310 sha 8dd25b2d, --batch-slots 8, context 131072 per row, same lib 678df626, keeps L356's a|b overlap and chip-greedy); then gate2-http.sh 18091 must PASS. Back: start-35b.sh. Gate: 32/32 concurrent replies identical to alone, turn 2 identical, golden identical, row-swap control red (sc serve/q36-batch 532cd940 evidence/q36-batch-20261021). Stage 1 = no server_busy up to 8 conversations; total tok/s same as today (rows run in turn) until L312's stroke plugs in.
swapped 18091 to bin-99230310, gate2 PASS live
Ledger line 368 on GitHub - 26 hours agodoneL368
Fix: mindmap build.sh reverted body-only note edits on no-change rebuilds (data loss since 10-03)
09:22:15Z CORRECTION: the 09:19Z done note named the wrong commit (2468f04 is the note). Real fix = ~/mindmap commit a693e43 (hash includes notes/ideas/hubs/bin/README/build.sh; checkout restricted to data page mirror). First test at 09:20Z lost the edit: race with the cron run of the OLD script. Re-test 09:21:55Z PASS: body-only append survived ./build.sh, committed 82acca4.
Ledger line 373 on GitHub - 32 hours agodoneL362
Fix: SC-AGENT-PREFLIGHT browser_mirror stub_page_screenshot_and_eval null element 'h' (13 FAIL PR heads in census A/D/E/F, flake on main)
fix/browser-mirror-page-target c692ebfc pushed (code 743aa79a); 20/20 local passes (10 desktop env, 10 bare judge env), aien-cli suite green (118 tests); READY line in ~/handoffs/2026-10-18-pr-desk.md 03:55:42Z; desk opens/judges/merges
Ledger line 367 on GitHub - 32 hours agodoneL364
Fix: q36_levers (sc levers/q36-quick, crates/aien-inference-abi/src/bin/q36_levers.rs ~line 106) writes every AIEN_* env var into its output JSON/log, including the secret AIEN_COORD_KEY
08ca2508 + test (rustc --test on the filter; cargo test blocked by omega pin, see handoff)
Ledger line 369 on GitHub - 32 hours agodoneL307
Judge once, reuse everywhere: gate verdict cache keyed by inputs across PR heads, shared gates first then unique, host gates in parallel (Drake 2026-10-20: test A once for nine PRs; do it in blocks)
From L335: sc fix/gate-reuse-inputs (ad950d7f) is stacked on judge/once 73565567 and must land after it. It adds manifest keys outside_inputs, listed, target_lock and run --shared-cache, and in cache::cross_head treats an input whose first component has * as the whole tree. Omega fix/gate-reuse-inputs (1aebbd4c) uses cache: inputs_complete + listed + outside_inputs, so it needs both judges installed first. Store logic otherwise untouched.
Ledger line 312 on GitHub - 32 hours agodoneL335
Efficiency fixer F1: gate reuse inputs (sc + omega) (findings G3-ci-judge-1, G3-ci-judge-10, G3-ci-judge-2, G3-ci-judge-9; review L325)
sc ad950d7f + omega 1aebbd4c pushed, READY 03:32Z. G3-1 M1 212719->7798 ms reused; G3-2 47338->10134 ms; G3-9 2.54/merge calc:c468f09f (L307 + declarations); G3-10 TRAIN 22979 ran/255 reused/19064 reran, 1994 s/70 heads calc:e2c60bae. Audits PASS, mutant FAIL; handoff ~/handoffs/2026-10-21-eff-gate-reuse-inputs.md
Ledger line 340 on GitHub - 33 hours agodoneL357
SWAP 18091 to bin-0915000b (L356 lever: a|b beside mix overlap ON, AIEN_OMEGA_Q36_OVERLAP_AB=1 default) gate2 PASS on 18093, mutant 4 FAIL, 200-token reply identical to live bin-5b726818; evidence sc levers/q36-quick 0408f8a6 evidence/levers-q36-20261021; start: Q36_BIN=bin-0915000b ~/.local/state/aien-q36/start-35b.sh (or write bin-0915000b to BIN); rollback: BIN back to bin-5b726818
swapped 18091 to bin-0915000b, gate2 PASS live
Ledger line 362 on GitHub - 33 hours agodoneL353
Merge census (read-only): every open PR and READY branch, verdict on exact head, merge state, superseded/duplicate/parked, landing order and batches, for the post-motor merge phase (L352)
census written: ~/handoffs/2026-10-21-merge-census.md + .json; counts in ledger note
Ledger line 358 on GitHub - 34 hours agodoneL339
Efficiency fixer F6: Aienux crypto speed (findings G4-os-3, G4-os-4, G4-os-5, G4-os-6, G4-os-7, G4-os-8; review L325)
01:40Z quiet-measure receipt committed d7c22e9 (core 5): GCM-SIV 4KiB 1194.9->4.43us, SHA256 4MiB 20711.6->1719.9us, X25519 457.6->43.6us, verify 168.3->97.3us, outputs identical; READY line moved to d7c22e9
Ledger line 344 on GitHub - 34 hours agodoneL336
Efficiency fixer F2: OMEGA-R11-OWN-HOLD fails under load (findings G3-ci-judge-8; review L325)
01:21Z coordinator review: HELD in the desk log; the owner case passes on a skipped living run under load (a skip). Follow-up L345.
Ledger line 341 on GitHub - 35 hours agodoneL326
Efficiency review G1-omega: find where the stack does more work than the math requires; findings with inputs + Wolfram formulas (integration L325)
/home/drakestapleton/handoffs/2026-10-21-efficiency-review/G1-omega.json + G1-omega.md; formulas checked in Wolfram by the coordinator (RANKED.md)
Ledger line 331 on GitHub - 35 hours agodoneL327
Efficiency review G2-sc-inference: find where the stack does more work than the math requires; findings with inputs + Wolfram formulas (integration L325)
/home/drakestapleton/handoffs/2026-10-21-efficiency-review/G2-sc-inference.json + G2-sc-inference.md; formulas checked in Wolfram by the coordinator (RANKED.md)
Ledger line 332 on GitHub - 35 hours agodoneL328
Efficiency review G3-ci-judge: find where the stack does more work than the math requires; findings with inputs + Wolfram formulas (integration L325)
/home/drakestapleton/handoffs/2026-10-21-efficiency-review/G3-ci-judge.json + G3-ci-judge.md; formulas checked in Wolfram by the coordinator (RANKED.md)
Ledger line 333 on GitHub - 35 hours agodoneL329
Efficiency review G4-os: find where the stack does more work than the math requires; findings with inputs + Wolfram formulas (integration L325)
/home/drakestapleton/handoffs/2026-10-21-efficiency-review/G4-os.json + G4-os.md; formulas checked in Wolfram by the coordinator (RANKED.md)
Ledger line 334 on GitHub - 35 hours agodoneL330
Efficiency review G5-physics-train: find where the stack does more work than the math requires; findings with inputs + Wolfram formulas (integration L325)
/home/drakestapleton/handoffs/2026-10-21-efficiency-review/G5-physics-train.json + G5-physics-train.md; formulas checked in Wolfram by the coordinator (RANKED.md)
Ledger line 335 on GitHub - 35 hours agodoneL331
Efficiency review G6-protocols-data: find where the stack does more work than the math requires; findings with inputs + Wolfram formulas (integration L325)
/home/drakestapleton/handoffs/2026-10-21-efficiency-review/G6-protocols-data.json + G6-protocols-data.md; formulas checked in Wolfram by the coordinator (RANKED.md)
Ledger line 336 on GitHub - 35 hours agodoneL333
Fix: crumb-doctor records FAIL AGREE '<repo> PR list' when the gh guard refuses reads (hourly reserve), raising a false CRUMB-DRIFT watchdog alarm
~/.local/bin/crumb-doctor patched; evidence ~/handoffs/fixer-crumb-doctor-guard-skip.md
Ledger line 338 on GitHub - 35 hours agodoneL324
Fix: aien-cli tool_batch two_readers_of_one_path_run_together fails at random under load (PR #670 own new test, 150 ms sleep as overlap proof)
fix/tool-batch-overlap-flake@4f6e840288c07b97f5aec9633015e7862f034650
Ledger line 329 on GitHub - 35 hours agodoneL322
Fix: spark-window starves the front of the line. _hold takes the chip, finds the quiet lock held SHARED (quiet-build/quiet-measure users arrive continuously), lets the chip go, and raw-flock users (L283 training queue, aien-test judge) take the chip; l306-router (top of the chip order) waited 40+ min (window.log ~/workspace/investigations/2026-10-20-q36-gate1-cosine/window.log). Needs: writer preference (an exclusive quiet waiter at the front blocks NEW shared takers), a correctness-only mode (chip lock, no quiet) for non-timing chip checks, and raw gb10 flock users routed through the line or made to yield to the front ticket. Related coordinator item L041
00:54Z L274 old-loop window now won quiet, waits for the chip (L283 turnover); I raised the gate once more for it, 15 min max, ends when its held file appears. Last extension.
Ledger line 327 on GitHub - 35 hours agodoneL318
Spark native boot 1: find the cause of the 3 M2 gate FAILs (pre-exit BOOTREPORT.TXT missing while native report present; nvram write after ExitBootServices marked FAIL; BootOrder gained 0004 = firmware-added recovery stick) - real fault or wrong check, fix whichever is wrong
feat/spark-boot-package@5bc20076 (= fix/boot1-gate): AienOS never ran on boot 1 (stale Sept report, firmware gave Boot0004 to stick); collector+preflight+stager fixed; verify_all STRICT PASS, M2 replay test, dry run READY; handoff ~/handoffs/2026-10-20-boot1-gate.md; READY line
Ledger line 323 on GitHub - 1 day agodoneL320
Fix: aien-run-exec/aien-test record a killed or never-started sandbox run (systemd-run 'Failed to start transient scope', exit 143 on shutdown) as FAIL rc_nonzero, and the cache then reuses that FAIL for every later run of the same experiment (aienos 77e346e6 block #440: 8 gates 'cached FAIL' from the 22:57Z reboot; 9 heads in 4 repos carry the signature)
d122afdea928326321ca376553aaabd4dfa19ab2 + 3 tests (verdict::an_infra_stop_is_not_run_never_fail_or_pass, runner sandbox_start/killed_by_shutdown)
Ledger line 325 on GitHub - 1 day agodoneL319
Fix: renda-harness restarts the operator judge on the same exact head every pass after it already has a verdict (aien-architecture #293 29dfeab4 NOT_RUN UNMAPPED_PATH started 33 times, aienos #440 77e346e6 FAIL 11 times since the boot); operator_run() checks only for a running judge, never for an existing operator bundle on that sha. Also find the cause of each verdict: aienos block #440 8 gates FAIL (CK-B7B-PROBE, CK-GATE-SELFTEST, CK-HOST-TESTS, CK-QEMU-GATES, DISK-QEMU, INFER-CRATE, NATIVE-NET, VERIFY-ALL), arch #293 UNMAPPED_PATH
harness installed (bak ~/.local/bin/renda-harness.bak-20261020T2348Z-L319), test ~/workspace/fix-harness-operator-once/test-operator-once.sh red 5/8 old, green 8/8 new; loops restarted 23:52Z, pass 2 at 23:54Z = operator runs 0, 4 desk lines (one per head), judge-run lines stay at 61; causes: aienos 77e346e6 = reboot-killed runs cached as FAIL (L320 fixer + L307 note), arch #293 = evidence/ unmapped, fix fix/arch-290-evidence-path d62d07f1 + READY
Ledger line 324 on GitHub - 1 day agodoneL182
Public proof: one-command 27B demo on a clean Spark, write-up with receipts, org front door, X thread, Anthropic/OpenAI letters (Drake 2026-10-20: be noticed, OpenClaw model)
clean demo PASS 14.15 tok/s wall 0.708 of 273 (calc:62fd33ec, calc:8fe79f5c), receipt aien-receipts demos/l182-one-command@611ba22b6e; site c9ea65fdc0, .github b032000309, sc d6942e21d0; READY lines 2026-10-20T22:28Z; handoff ~/handoffs/2026-10-20-public-proof.md
Ledger line 183 on GitHub - 1 day agodoneL266
The AienOS boot experience everywhere: the iMac's two circling black holes, galaxy and sound, ported to every target (Spark, Pi, PC live/VM images, and the aienos-run window via the VM framebuffer + audio), measured frame rate, then the first-run UX design document
L282 dogfood round 1: hello screen first picture at 8.0 s shows sky + two holes and no words (/tmp/l282-df/round1/games-local/01-hello.png); a person may not know to wait for the merge. Hint text drawn 150/256 and faint 110/256 (allen_ui.c 271/300) is below 4.5:1 as drawn. Report docs/design/DOGFOOD-2026-10-20.md on feat/allen-design-system@8286892a.
Ledger line 271 on GitHub - 1 day agodoneL267
Allen's setup wizard: asks intent (faster games / run an AI model / other), detects every device on the computer and on the network (Interplane discovery, L263) and sets them up with nothing manual, then the LLM pick: auto-detect installed local models (Ollama, LM Studio, llama.cpp files, Omega serve), or connect Claude/OpenAI/any provider with a key, guided inside Allen's window
L282 dogfood round 1 (wizard 660106f3, QEMU, 6 runs all reached done), details docs/design/DOGFOOD-2026-10-20.md on feat/allen-design-system@8286892a. Wizard copy: (1) games screen 'Frames a second: I couldn't measure this yet' is a dead end line (setup_wizard.c ~300, picture /tmp/l282-df/round1/games-local/05-games.png); (2) 'Paste your key' (lines 554, 851) but the kernel has no paste: person types ~100 chars; (3) 'OpenAI-compatible apps' (547) is jargon. Round 2 against 36cc66cf follows.
Ledger line 272 on GitHub - 1 day agodoneL203
Decide: where build/judge/git/agents live after Linux is off (gap map sec 6); approve moving the Phase 3 judge; drop Wi-Fi/BT/audio (D13)?
Drake answered: everything lives on AienOS
Ledger line 204 on GitHub - 1 day agodoneL204
Look at the Spark unit: ConnectX ports and serial header physically present? Both root ports 0000:00 and 0002:00 show empty buses
answered from software; build proceeds
Ledger line 205 on GitHub - 1 day agodoneL287
Fix: AienOS console session busy-polls one host core at ~94% CPU while idle (VZ and QEMU-HVF on M2)
2230f7f9 on fix/console-session-idle: QEMU idle CPU 98.8% -> 1.2% (1 vCPU); console gate rows 146-155 PASS; -smp>1 firmware spin remains, see handoff
Ledger line 292 on GitHub - 1 day agodoneL199
Aienux NVMe: queue depth above 1 and direct DMA, 512-byte disk, measured in QEMU (gap D4, path item 3)
feat/spark-nvme-queue-depth@dfbacaa; QEMU gates PASS; strict verify_all PASS (QEMU only, hardware NOT_RUN); QEMU read 135->7877 MB/s; handoff ~/handoffs/2026-10-20-spark-nvme-queue-depth.md; READY line in pr-desk log
Ledger line 200 on GitHub - 1 day agodoneL198
Aienux RTL8127 10GbE driver for the Spark against a chip model + UDP boot report (gap D5, path item 2)
aienos feat/spark-rtl8127-net@e9ca5a0, READY line in pr-desk log, handoff DONE; attended boot #2 remains as hardware test
Ledger line 199 on GitHub - 1 day agodoneL263
Every AienOS instance (native on Spark/iMac/Pi, or inside aienos-run on Windows/macOS/Linux) joins Interplane: identity, discovery, roster, reachability through the signed transport and the gateway; one design, built and proven in QEMU with three instances
aienos feat/aienos-interplane-join 5af5ac7d (gate AIENOS_CK_INTERPLANE PASS, 3 instances + revoked + TEST), interplane feat/aienos-join a51bd73, READY line in pr-desk log
Ledger line 268 on GitHub - 1 day agodoneL265
aienos-run for macOS: Virtualization.framework backend on the shared core (L262), runs the aarch64 AienOS image natively virtualized on Apple Silicon, measured boot-to-console on Drake's MacBook for real, one-command try (curl | sh), signed/notarization steps documented
From L267: the wizard's host-description reader is only in the launcher image: boot IMACKERN-HOST.EFI (make imac-kernel CK_X86_HOSTDESC=1 in feat/allen-setup-wizard@9c8551e7) and pass fw_cfg opt/aienos/host (format: native/kernel/svc/tests/fixtures/setup/*.txt). The default IMACKERN.EFI carries no fw_cfg code on purpose (doctrine check in the Makefile). x86 only today.
Ledger line 270 on GitHub - 1 day agodoneL188
Fix: aien-serve-http re-takes /tmp/aien-gb10.lock between back-to-back requests (flock unfair), starving every other waiter 18h on 2026-10-20; server must yield to queued waiters before re-locking
8fcea131 pushed fix/serve-gpu-lock-yield; red on main, green 6/6 host tests; READY line added
Ledger line 189 on GitHub - 1 day agodoneL194
omega #598 integration: fold fix/r13-intervals-under-load (L172) into fix/host-gate-flakes, recompile crumbs with the pinned tool if the repo still lists CRUMB-VERIFY, push with lease, READY line with the new head so Renda runs the operator judge
8d591af9 already pushed by desk; local R13 host test PASS; no READY line needed
Ledger line 195 on GitHub - 1 day agodroppedL292
aien-train data: training corpus on the Spark for the first run (FineWeb-Edu sample 10BT + a held-out slice; sha256 manifest, SOURCE files, licence notes; token count; tokenizer choice with reason; stored under ~/datasets/aien-train, never in git)
DROPPED: Drake 2026-10-20 19:13Z: our training only (AIEN_0 on Omega synthesis traces), no web-text pretraining; agent stopped by the coordinator; 27 GB partial download left under ~/datasets/aien-train for spark-cleanup
Ledger line 297 on GitHub - 1 day agodoneL189
Fix: spark-window _hold starves for hours when the quiet lock is held shared back-to-back (it wins quiet, lets it go, waits for gb10, then loses quiet again; L131 window 20261020T134016Z-713c9c: 12 tries in 2 h)
aien-architecture fix/spark-window-starve d612f4e; test tools/ops/test-spark-window-starve.sh (old fails, fixed 4/4); installed ~/.local/bin/spark-window; handoff ~/handoffs/2026-10-20-fixer-spark-window-starve.md
Ledger line 190 on GitHub - 1 day agodoneL298
Fix: aien-watch CRUMB-DRIFT alarm: AGREE check fails on aien-sovereign-core PR #698 head 380b3f6e (since 2026-10-20T18:53:48Z)
False positive in crumb-doctor: #698 (Am2) has no gate/hosted crumb step (n/a) but doctor still compares the bare crumb shim ('14 of 14 stale'); no crumb file wrong. Proposal: treat gate=n/a+hosted=n/a as PASS in check_agree; else clears on #698 merge, which also deletes the doctor so merger runs aien-watch clear. Evidence /tmp/L298-doctor.out, handoff ~/handoffs/2026-10-20-fixer-crumb698.md. Alarm NOT cleared.
Ledger line 303 on GitHub - 1 day agodoneL196
All Spark work paused for Linux replacement; lift MERGE-PAUSE and resume stopped items (see /home/drakestapleton/handoffs/2026-10-20-shutdown-for-linux-replacement.md) when Drake says
Drake said resume
Ledger line 197 on GitHub - 1 day agodoneL285
aien-train history-nvidia (Drake: scrape NVIDIA training docs, learn their hard lessons, find the good stuff): developer blog, docs.nvidia.com performance and mixed-precision guides, Megatron/NeMo/Transformer Engine docs, Nemotron reports, FP8/NVFP4 papers, DGX Spark training playbooks; cited lessons table
study/nvidia@edcc611113425eb22a4233672c412be5919d9813 ~/handoffs/2026-10-20-aien-train-history-nvidia.md
Ledger line 290 on GitHub - 1 day agodoneL278
aien-train efficiency: engineered small-compute training frontier (speedruns, Chinchilla, data quality, FP8/FP4, megakernels), cited
study/efficiency@1c03dc13f3edc3e097b252093bf76f7896c1b9a5 ~/handoffs/2026-10-20-aien-train-efficiency.md
Ledger line 283 on GitHub - 1 day agodoneL277
aien-train history-others: Claude, Grok, Google, Meta, DeepSeek, Mistral, phi, right and wrong, cited
study/others@e18fdb0e4818c66c56bbf52eabb2ed650048325a ~/handoffs/2026-10-20-aien-train-history-others.md
Ledger line 282 on GitHub - 1 day agodoneL276
aien-train history-openai: GPT-1..GPT-4/o-series, right and wrong, cited
study/openai@b0cf41bbd6ea4014f626cc7c4e8c84a6eefe0b74 ~/handoffs/2026-10-20-aien-train-history-openai.md
Ledger line 281 on GitHub - 1 day agodoneL214
Reverse engineer the MacBook Air Apple GPU (AGX) for AienOS, on the Mac: prior art + licences, ISA asm/disasm from compiler output, firmware interface from ioreg, m1n1 trace kit for the attended step, simulated AGX + vector add (Drake: reverse engineer it yourself)
feat/macbook-agx-reverse-engineering@4b11bb9, 6 kernels bit-identical real vs simulated, 177 checks; attended trace step pending (docs/MACBOOK_AGX_TRACE_PROCEDURE.md)
Ledger line 215 on GitHub - 1 day agodoneL275
Fix: wall-power-log.sh not running since 2026-10-19T10:13 CDT (last row ~/yardsticks/wall-power/2026-10-19.csv); no 2026-10-20.csv; nothing restarts it
Cause: old logger replaced on purpose by spark-sensors (running pid 738092, @reboot cron), not killed; no restart needed. Today's log: ~/yardsticks/sensors/2026-10-20.csv, 21076 rows, current, no gaps >10 s; wall W in column 3. Handoff ~/handoffs/2026-10-20-fixer-wall-power.md
Ledger line 280 on GitHub - 1 day agodoneL192
GB10 cut C host side: GSP firmware boot protocol encoder + RPC subset, host tests
aienos feat/gb10-gsp-host, READY line written by the agent; GSP boot package built in memory with host tests
Ledger line 193 on GitHub - 1 day agodoneL213
Fix: iMac GPU code/docs/tests name the chip Radeon Pro 555/560 Polaris 11 id 1002:67ef; real iMac18,3 evidence says 1002:67df rev c5 (Polaris 10, Pro 570/575/580)
a8c274c
Ledger line 214 on GitHub - 1 day agodoneL195
Fix: physics nvrm_gp_entry/nvrm_enqueue accept pushbuffer VA >= 2^40 (GPFIFO GET_HI is 8 bits) and silently corrupt LEVEL/LENGTH
From L193: also nvrm.c comment above nvrm_gp_entry says 'LEVEL 9:9 = MAIN' but (1ull<<41) is GP_ENTRY1 bit 9 = LEVEL_SUBROUTINE (clc96f.h NVC96F_GP_ENTRY1_LEVEL 9:9, SUBROUTINE=1). Behaviour is chip-proven: do not change the bit without a chip run; correct the comment to say what the code does and mark the MAIN/SUBROUTINE meaning as needing a chip check.
Ledger line 196 on GitHub - 1 day agodroppedL205
D13 driver ports: a documented, tested driver interface for Wi-Fi, Bluetooth and audio in Aienux with NO driver behind it (Drake: make the port, not the driver); dispatch after L198 lands so it follows the RTL8127 driver shape
DROPPED: Drake 2026-10-20: make the drivers, not just the ports (superseded by the Wi-Fi/BT/audio driver streams)
Ledger line 206 on GitHub - 1 day agodoneL191
Replace Linux on the Spark: complete gap map + tracking plan
aienos feat/spark-replace-linux-map@bb2399f docs/SPARK_REPLACE_LINUX.md, handoff ~/handoffs/2026-10-20-spark-replace-linux-map.md
Ledger line 192 on GitHub - 1 day agodoneL154
Chip lock held while nobody uses the chip: judge holds gb10 lock across CPU gates (97 min today); spark-window takes quiet lock before it has the chip. Fix lock order in aien-test and spark-window; log long holds
judge binary switched to L154 build at 16:01Z (backup /home/drakestapleton/.local/state/aien-local-ci/aien-test.bak-20261020-pre-L154); omega fix/gb10-gate-prepare may now merge
Ledger line 155 on GitHub - 1 day agodroppedL020
Agent: flashnext-port
DROPPED: Drake 2026-10-20 (desktop chat): drop Flash-Next, free the disk. Removed: model qwen3.8-flash-next-nvfp4-fc694b54 (124 GB; .SOURCE/.sha256 kept), worktrees omega-flashnext + sc-flashnext, ~/flashnext-scratch, targets/flashnext-omega. Code kept on GitHub: omega flashnext/port a6daa5c0 + wip snapshot of the 7 uncommitted files, sc flashnext/port 3ca1a497, handoffs ~/handoffs/2026-10-20-flashnext*.md
Ledger line 21 on GitHub - 1 day agodoneL173
Fix: omega R13 'plan was not for the goal' FAIL under load (AIEN's plan not for the goal, no drift note) seen 1/40 loaded base runs on fix/host-gate-flakes c2dd2a0d
ce2be092 pushed on omega fix/r13-plan-for-goal (ef350ccd fix + crumb); loaded loops base 13/150 -> fixed 0/150 of the 'plan why core class, goal 0' shape; test-r11 691/0; evidence r13p-tools/load/{diag,fix5}.log; handoff 2026-10-20-omega-r13-plan-for-goal.md
Ledger line 174 on GitHub - 1 day agodoneL185
Real sky follow-up: Large and Small Magellanic Clouds from real catalog data in night sky (feat/real-sky, 5 uncommitted files from dead agent)
feat/real-sky@662b86d3 pushed; sky 15/15, Allen UI 91+26, memory 189, imac-kernel ok; READY line in ~/handoffs/2026-10-18-pr-desk.md; handoff ~/handoffs/2026-10-19-real-sky.md
Ledger line 186 on GitHub - 1 day agodoneL187
Fix: feat/imac-gpu-core amd_gpu_regs.h register numbers do not match Linux AMD headers (GRPH 0x5045 vs 0x1a04, SMC_MSG 0x178 vs mmSMC_MESSAGE_0 0x94)
7642c70a fix/imac-gpu-regs-verified pushed: 87/88 register numbers corrected from AMD headers, scripts/gpu_bench/check_regs.sh red 113 problems -> green 164 checks, host tests 21/21+24/24, kernel builds GPU=0/1 ok; handoff ~/handoffs/fixer-imac-gpu-regs.md
Ledger line 188 on GitHub - 1 day agodoneL026
Rig: TARGETS.md, what others get on 5070 + ceilings + FPS levers
docs/targets pushed 2026-10-19T17:10Z (aien-rig branch docs/targets exists, RTX 5070 targets + ceilings)
Ledger line 27 on GitHub - 1 day agodoneL027
Rig: Windows FPS tuner in Rust (rig-tune.exe)
aien-rig tuner/v0 pushed 2026-10-19T17:20Z, Windows CI rig-tune (x86_64-pc-windows-msvc) success
Ledger line 28 on GitHub - 1 day agodoneL032
Measure whether a 2nd 27B server shares the model page cache with the 1st (36.7 GiB peak, ~35.5 GB is model files)
measured 2026-10-19 17:22Z: 2nd server does NOT share page cache; loader demands 41,305 MiB + 8,192 MiB free; refused at MemAvailable 46,249 MiB (serve-27b.log)
Ledger line 33 on GitHub - 1 day agodoneL037
Horizon: black-hole dusk animation for goodbye (ui), worktree aienos-goodbye-horizon
feat/goodbye-event-horizon 903277b7 pushed, ui host test PASS 98 checks with frames, READY line appended. Follow-up (real astrophysics disk) tracked as a new item
Ledger line 38 on GitHub - 1 day agodoneL044
Merge note: sa-0 (old horizon agent) died, work continued as L037 (1c05936 pushed); panel-off branch feat/goodbye-panel-off 2954f5e is READY and done
L037 READY line seen (903277b7); panel-off 2954f5e READY
Ledger line 45 on GitHub - 1 day agodoneL109
TURING Yield rerun with energy connected (Drake): TY-7 pre-registered then measured = bits of T per wall joule for the 27B on our engine decoding held-out text (spark-receipt wall J + hwmon J), 3 rounds; close TY-4/TY-5 if the code runs; single measure owner
READY research/turing-energy-rerun@a1e7f349, TY-7 protocol pushed before run, TY7_RESULT.md + evidence/TURING_YIELD/ty7-*, handoff ~/handoffs/2026-10-19-turing-energy.md
Ledger line 110 on GitHub - 1 day agodoneL118
Reimagine build+measure (Drake: too many jams): spark-broker owns chip + quiet state + queue; waiting happens in the queue not inside locks; builds run while timings are queued; measurements one at a time with L113 windows; quiet-build/quiet-measure become shims; chosen by replayed wait-time numbers vs the lock wrappers
READY ops/core-split@3da7e400 with replay comparison; verdict INCONCLUSIVE by the pre-registered rule, so not installed; negative verdict is the deliverable
Ledger line 119 on GitHub - 1 day agodoneL124
Fix: Python file evidence/cascade-summary.py on sc cascade/mtp-chain (and branches built on it) breaks the no-Python rule
READY fix/cascade-summary-no-python@8c4017cc; mtp-chain already carries 38997e29
Ledger line 125 on GitHub - 1 day agodoneL126
SPEED-2 speculative decode end to end: wire tree-verify GEMM (cascade/tree-verify-gemm, 32 rows ~ 1 row) + MTP chain k8 into the decode graph and aien-serve-http; output identical to plain greedy (L123 audit)
READY omega speed2/specdecode@31db8ecf + sc speed2/specdecode@68fa1683: tokens identical to k0 on every run (rows-mode verify)
Ledger line 127 on GitHub - 1 day agodoneL127
SPEED-3 4-bit GEMV to the streaming ceiling: next-block prefetch (second wait slot in codegen_w4), down/mixer tile split so all 48 SMs work (L123 audit)
READY omega speed3/w4-gemv@8655007f (base dd5e89ff): load-ahead prefetch measured, see note
Ledger line 128 on GitHub - 1 day agodoneL134
SPEED-10 below 4 bits: CPU-only drift study of a 3-bit trellis / codebook code on the 27B before any kernel (ON PAPER 27.7 tok/s ceiling, moonshot B row 6) (L123 audit)
READY sub4bit/drift-study@4451fec3 (CPU drift study, 3-bit trellis, on forgejo drake/aien-sovereign-core)
Ledger line 135 on GitHub - 1 day agodoneL152
Fix flaky heartbeat test writer_waits_for_lease_release: depends on a 10 ms sleep, failed at load 37 (seen in L080); make it wait on the event, not on time
READY fix/heartbeat-lease-test@df2cfceb: channel-based wait, old test failed 5/20 under load, new 20/20
Ledger line 153 on GitHub - 1 day agodoneL158
Fix: nvidia-modeset/kthread_q kernel thread burns one full CPU core since boot (41h52m CPU in 42.6h up)
root cause documented + proposal in ~/handoffs/fixer-modeset-kthread.md (nvidia_modeset kthread spin; blacklist at next reboot; interim taskset); no live change
Ledger line 159 on GitHub - 1 day agodoneL164
Merge train + gate memo in aien-test (Drake: faster merges, same rigor, intuitive): gates declare inputs and are reused by content key; host gates first, silicon queued; READY branches judged as one candidate; aien-local-ci queue page
READY feat/merge-train@d78c38f1 (stacked on #695), evidence/MERGE-TRAIN, dry run 18 in / 34 out; desk PRIORITY written: #695 -> merge-train -> install judge -> trains
Ledger line 165 on GitHub - 1 day agodoneL166
Entity pilot G5 made decidable: horizons 3/10/40, warm and cold interleaved per objective, lazy template rebuild arm, run at load under 15
READY research/entity-pilot-276-h40@70307c5d, evidence/ENTITY-276/h40, agent reported done
Ledger line 167 on GitHub - 1 day agodoneL167
#584 re-measurement: A/B/C0 interleaved per objective, reuse opportunities per key 1/3/6, new seeds, load under 15
READY research/learning-search-584-r2@ce47b8b8, evidence/LEARNING-SEARCH-584/r2
Ledger line 168 on GitHub - 1 day agodoneL168
Runtime fix: rx_plan adaptation stores and reuses a 12-move plan where search finds 9 (negative transfer, E03 finding); red test first, smallest fix, payback floor rerun
READY fix/rx-plan-adaptation-quality@4c9fae6c: red test cc3261eb, fix 9562508d, payback heldout F2 bad rows 51 -> 0
Ledger line 169 on GitHub - 1 day agodoneL171
Fix: omega Makefile test recipes run ./$(OUT_DIR)/... so an absolute OUT_DIR fails (Error 127)
fixer done: Makefile absolute OUT_DIR fix pushed (see ~/handoffs/fixer-makefile-abs-outdir.log)
Ledger line 172 on GitHub - 1 day agodoneL183
Fix: CRUMB-DRIFT 5 AGREE fails (sc #698 380b3f6e, sc #674 1e2a8000, aien-protocols main 4529f393, aegis-runtime main f85dbe39, spark-rsi main 3d5dc2f1) since 13:27Z
2026-10-20T15:24:52Z /usr/local/bin/crumb reinstalled from crumb-tools/12ad0141 (sha 2fb09401, old kept as crumb.bak-*); crumb-doctor --repos aien-protocols,aegis-runtime,spark-rsi: 7 pass 0 FAIL; sc#698/#674 lines clear when they merge
Ledger line 184 on GitHub - 1 day agodoneL165
Fix: spark-adapters test_challenge1_concurrency_50_threads_sub_millisecond_reads fails under load (wall-clock avg 1724us vs 1000us) on sc main 5208aa2e judge
PR #699 opened
Ledger line 166 on GitHub - 1 day agodoneL177
guards: allen-guard Rust crate: programme close guard (reopen tracker when phase/prereq open or checklist incomplete; fixtures open / closed-missing-evidence / fully-accepted) + source-control guard (raw logs, weights, secrets) + CI workflows
guards/v0@ec97384701c6e86aa2be745c1336b70645a549d4, ~/handoffs/2026-10-20-allen-p0-guards.md
Ledger line 178 on GitHub - 1 day agodoneL175
inventory: Allen P0 conversation log inventory on Spark + MacBook (agents, hosts, date spans, counts, raw mirror coverage, tool evidence, gaps) + access-controlled source manifest pointer; counts and dates only
inventory/v0@82c8425e7e5d3121c0e473a19344bbf065c7351f, ~/handoffs/2026-10-20-allen-p0-inventory.md
Ledger line 176 on GitHub - 1 day agodoneL176
contract: Allen B0 acceptance contract (first bounded task frozen, human-effort + outcome gates, budgets, access rules, role/permission boundary, devices, deletion/export ownership)
contract/v0@2bc61abd56d7afea98346159d8a49efd5b56051a,
Ledger line 177 on GitHub - 1 day agodoneL174
map: Allen P0 source map (SubjectState/ADR0035, Cortex, agent-archive, runtime, protocols, device code) with pinned commits, extend/reuse/missing, ownership + import boundaries; plus arch repo-map branch
map/v0@de430139a0fea033638c4af92e489b87a8ae0af1, arch docs/allen-repo-map@46aa535822942d95c7eb0bcfd885dfc1a764a4ae, ~/handoffs/2026-10-20-allen-p0-map.md
Ledger line 175 on GitHub - 1 day agodoneL172
Fix: omega R13 'production intervals' FAIL under load (AIEN=0 Omega=0) in R16-AUTHPATH on #598 c2dd2a0d
8d591af9 pushed on fix/r13-intervals-under-load (fix ef4f5656 + crumb); loaded loop base 7/40 interval FAILs -> fix 0/100; mutant killed; handoff ~/handoffs/2026-10-20-omega-r13-intervals.md; READY line in pr-desk
Ledger line 173 on GitHub - 1 day agodoneL159
Fix: tree-verify GEMM (omega_gpu_tv_f32) runs ~3.5x slower inside the qwen38 batch door than in gpu_tcgemm_test (503 vs ~150 ms chip per 64-layer step at 1 row)
omega fix/tv-inserver b81195e85655773b3cc461d394dfd0b3843729dc (fix 3208be4b + 78c93b35 + evidence): tv GEMM waits with the platform sleep-poll explicitly; in-server gemm launches 170.0/177.8 ms/step at nice 10 (calc:0f5e9f0f calc:a23e8f35) = 1.19x of the 150 ms isolated reference (calc:48e27c70), same-window base 175.4/214.3; decode-thread camera shows sleep-poll (22 slices/launch) vs spin (0.8); host tests 14/0 8/0; evidence/l159-tv-inserver-20261020 on the branch
Ledger line 160 on GitHub - 2 days agodoneL078
Manifesto page in Drake voice: intelligence should be owned; LLMs came from the corpus of all human knowledge so every human who can run it has the right to; compute as a service is fine, paywalls that keep people from running it themselves are not; it is our responsibility
coordinator: closed early by another session; page is READY (site/manifesto@5927b50f) but not merged and Drake has not yet said it sounds like him; tracked on the desk item with /allen/
Ledger line 79 on GitHub - 2 days agodoneL093
Op list carries OSC-2 structs: struct table, struct ref types, fload/fstore, version-2 encoding in tools/omega_emit (emit + lift); schema frozen first for the Omegese agent
READY feat/emit-structs@54114817 RT 11/11 C2 11/11 NEG red/green
Ledger line 94 on GitHub - 2 days agodoneL094
Omegese struct words (struct/new/get/put on fields), omegese next mask, emit to the extended op list; both candidate spellings measured, fewest tokens wins
READY feat/omegese-structs@a4935b61 (fixer L156 follows)
Ledger line 95 on GitHub - 2 days agodoneL098
Bytes tool: hand-written GPU instruction words run on the GB10 through the existing launcher with a receipt (simplest kernel: add two vectors); same for AArch64 words; then measure ceiling fraction of hand bytes vs the C-emitted kernel for the same op
READY feat/omega-bytes@b4c7c6ef, vadd PASS 3/3 on chip
Ledger line 99 on GitHub - 2 days agodoneL103
Seed-0 birth: hand-written AArch64 hex seed (reads hex, writes bytes) -> hex1 (labels) -> text-table assembler -> Omegese-subset-to-words translator; existing C compiler and GNU tools as reference checks only; rung-3 word-count table seed chain vs C path on the 10 dev tasks
READY seed/birth@9de9f45b, 4 rungs byte-identical
Ledger line 104 on GitHub - 2 days agodoneL105
Atlas reborn from nothing, scoped into omega seed/ (Drake 2026-10-19): rebuild the 1,480-byte boot seed through the seed chain (L103 rung 2 assembler), no GNU as/ld anywhere; the audited atlas.bin is the reference oracle; then the boot image/ELF header as hand hex words; QEMU bare-metal run through the existing Atlas M1 gates
READY seed/rungs56@217b7cc8 rung 5 atlas
Ledger line 106 on GitHub - 2 days agodoneL117
Seed rungs 5+6 follow-on (rung 2 READY @34468c19): Atlas reborn through asm2 byte-identical to the audited atlas.bin + QEMU M1 gates; sm_121 target table in asm2 bit-identical to C codegen words + one kernel run on chip via the bytes tool
READY seed/rungs56@217b7cc8 GATE6 6091 programs bit-identical
Ledger line 118 on GitHub - 2 days agodoneL135
R2 floats in Omega (OSC-4 slice): f32/f16/bf16/fp8 types, ops, one rounding rule each, source-order reductions; encoder+decoder forms with manual citations; G1 10k bit-exact vs C reference per op, G2 dot+softmax bit-identical, G4 version bump
READY feat/osc4-floats@b18757f0
Ledger line 136 on GitHub - 2 days agodoneL136
R3 more than one GPU channel / overlap transfer and compute (decision 12 accepted): N-channel session, decode graph split with Fences; G1 N=1 bit-identical, G2 N=2 bit-identical, G3 timing receipt N=1 vs N=2 with J/token in a window; fallback = reorder on one channel
READY feat/gpu-channels, N=2 no gain 0.999, stays opt-in
Ledger line 137 on GitHub - 2 days agodoneL137
R4 436 launches per token: launch-tax receipt (real vs empty-kernel command list vs one launch), fusion map by bytes saved, top two fusions bit-identical, timing receipt after
READY feat/fewer-launches 436->340 jobs -2.22%/token, launch tax 0.0295
Ledger line 138 on GitHub - 2 days agodoneL151
Machine ceiling map (Drake: the project is scored against the machine limit, not one stream): every data path and arithmetic unit of the GB10, spec vs measured with receipts, concurrent whole-machine bytes/s, joules per byte, every lane placed against its binding wall
READY doc/machine-ceiling@3c8ce239, 222.3 GB/s = 0.814 of 273
Ledger line 152 on GitHub - 2 days agodoneL163
Fix: omega R14 scenario E seat kill under load: 'the seat did not hold the lane claim (claimed 0 held 0)', 'lost 3-4 ... cause 0 want 1653' (3/46 runs, cause 4 of L160)
2e1455fe pushed on fix/r14-e-seat-claim: test-side race fixed (driver_pause handshake); knob red 5/5 -> green 5/5; load base 2/175 -> fix 0/175; mutant caught 2/2; r14-host r13-host r16-authpath r10 r11 PASS; handoff ~/handoffs/2026-10-20-omega-r14-e-seat-claim.md
Ledger line 164 on GitHub - 2 days agodoneL161
Fix: omega speed checks report 'machine too busy to time' as FAIL instead of NOT_RUN cpu_busy (RX-HOST-REFERENCE on #598 4adbb6a6, load ~60)
pushed fix/speed-check-cpu-busy-not-run ccae45c2: busy -> NOT_RUN cpu_busy (step exit 77, full step run); M2 busy+real failure FAIL (full step exit 2); M1 slowdown FAIL; quiet PASS UNVERIFIED; run.sh changed (pipeline config); handoff ~/handoffs/2026-10-20-omega-speed-check-cpu-busy.md
Ledger line 162 on GitHub - 2 days agodoneL162
Fix: spark-window close without an ID closes another lane's open or queued window (L129's bare close at 07:31:05Z killed L159's window 20261020T073002Z-19bfd7; L159's close then closed L127's queued 073105Z-0c7237)
fdabc61c6e3b4216bdcac155ecbee9b7ac64e362
Ledger line 163 on GitHub - 2 days agodoneL160
Fix: omega R14-HOST intermittent FAIL (adapt -10 no promotion in 90 s, refusal AUTHOR; also adapt -5, production stopped while candidate refused)
origin/fix/r14-recovery-flake e5404a05; 3 runtime commits; red 20/20 -> green 19/20, stale-author 3/44 -> 0/46, t_drift_moved red 4/green 0; all 8 host targets PASS; handoff ~/handoffs/2026-10-20-omega-r14-recovery-flake.md
Ledger line 161 on GitHub - 2 days agodoneL157
Fix: sc aien-kv-store test flake 'store directory ... is in use by another open store (LOCK is held)' in a_legacy_object_others_can_read_is_tightened_on_recovery
PR sc#696 MERGED 05:25:41Z 07228020 (desk)
Ledger line 158 on GitHub - 2 days agodoneL156
Fix: omegese next says 'space' legal at the 96-words-per-line boundary, then refuses prefix+space (complete's own witness starts with that space and emit accepts it)
af31a025328c6a80b875422af835348ed4bb12ae
Ledger line 157 on GitHub - 2 days agodoneL080
Wormhole: 27B (resident) and Flash-Next MoE (experts streamed by the heartbeat ring) work together through their shared vocabulary as draft + block verify; race both directions and DFlash, fastest wins; output must stay identical to the verifier
auto: full receipt committed wormhole/27b-flashnext@232b4ba4 evidence/WORMHOLE-ACCEPT-69825c92 (ring/pair results in README tail)
Ledger line 81 on GitHub - 2 days agodoneL022
Agent: omega#580 rolls/chords
feat/580-chords b6fbc8f4 pushed; evidence/OMEGA-580-CHORDS/{06-chip-r4,07-joules,RESULT.md}; READY line in pr-desk; open: clean rerun, ceiling, voices scaling
Ledger line 23 on GitHub - 2 days agodoneL155
Fix: omega tests/r15_energy_preflight_test.sh 'valid' case flakes under load (package counter did not advance)
3cb00b68 fix/r15-preflight-flake; 20/20 under load ~60, stuck case still fails
Ledger line 156 on GitHub - 2 days agodoneL067
Real sky: real star catalog, real Milky Way, real galaxies in night.c; science audit of black holes, lensing, chirp; worktree aienos-real-sky
feat/real-sky rebased onto feat/goodbye-panel-off, pushed 2f14900 (force-with-lease from 49ab287). Level 7 Gaia Milky Way (196608 cells, sha256 pinned); Milky Way at physical brightness (V 20.7 band over V 22.0 sky), colour at fixed luminance; gain 1.2 calibrated (Galactic-centre luminance ratio 3.23 vs physical 3.31). Coordinate entry (35 north 90 west, 35.1, -90.0): 26/26. Science doc section 2 rewritten for Gaia with queries and sha256. Flight plan docs/SKY_FLIGHT_PLAN.md (no code). Tests: sky 13/13, Allen UI 91/91 + place 26/26, memory 189/189; IMACKERN.EFI 405504 B at 4221f12. OPEN: Magellanic Clouds not visible (no light model since the photograph was removed; Gaia grid at the LMC shows no cloud). NOT RUN: full make test suite, QEMU. NOT MEASURED on iMac.
Ledger line 68 on GitHub - 2 days agodoneL146
Session-limit cutoff 2026-10-19 ~19:3x CDT killed 9 Claude agents (L094 L097 L098 L112 L113 L117 L135 L136 L137); restart scheduled 21:03 CDT via systemd-run (~/handoffs/restart-2103/restart.sh); L093 L103 L108 L109 still alive
ledger: done/drop refused on closed items, every write keeps LEDGER.tsv.prev, list prints counts; 22 met items closed 2026-10-20 02:5xZ
Ledger line 147 on GitHub - 2 days agodoneL014
Agent: mtp-chain(+tree, +n-gram)
READY 230574e0, suites finished, summary written (met; JSON pass flags for desk)
Ledger line 15 on GitHub - 2 days agodoneL059
#249 E06 checked elimination: versioned rewrite contracts + independent derivation checker, proven vs probe-agreement, manual vs greedy vs bounded cost-guided search ranked by rx_costmodel (unknown cost never 0), negative cases, frozen budgets/held-out, full search+check+exec cost
research/rewrite-elim-249@79c436c2 READY (met)
Ledger line 60 on GitHub - 2 days agodoneL060
#276 acceptance freeze doc in aien-architecture: child interfaces, budgets, held-out gates, numeric thresholds, pinned child freezes (OEV1@5d864920, LS584 FREEZE, E03 protocol@293b49a7, crossover@1cd04c8e), entity-pilot + #249 gates, completion checklist status
research/276-acceptance-freeze@21913232 G1-G3 PASS G4 NOT_RUN G5 INCONCLUSIVE, READY (met)
Ledger line 61 on GitHub - 2 days agodoneL084
sc crumb/advisory f7646e38 conflicts with main (verify.rs, LOCAL_VERIFICATION.md after #680/#681): rebase, resolve, test, push
READY 380b3f6e (met)
Ledger line 85 on GitHub - 2 days agodoneL097
Map the chain Omegese -> GPU bytes as it is today: every stage, its language (C/Omega/bytes), line count, which stages are CPU-only baggage; propose the shortest C-free chain and what the language needs for it
doc/chain-to-gpu@d88b8097 CHAIN_TO_GPU.md, READY line (met)
Ledger line 98 on GitHub - 2 days agodoneL099
Study: how compilers were birthed without a compiler (gcc 3-stage, first C, hand-bytes chains); extract only the stage structure and gates that apply to Omega self-host (Sonnet reading)
note ~/handoffs/2026-10-19-bootstrap-study.md (met)
Ledger line 100 on GitHub - 2 days agodoneL100
Fix: omega R9 flip holds world callers_mu across fsync (global stop of production)
fix sha pushed, E8 red on main green with fix (met)
Ledger line 101 on GitHub - 2 days agodoneL101
Driver dig: trace each of the 8 ioctls the M16 native channel uses (physics/m16, physics/nvrm) through the vendored open kernel module source to where it becomes a GSP firmware call; size what a minimal module would have to implement; name what is signed firmware and cannot be replaced
doc/driver-floor@30cf10a3 DRIVER_FLOOR.md, READY line (met)
Ledger line 102 on GitHub - 2 days agodoneL102
Fix: omega RX-HOST-REFERENCE test-workflow-fusion parallel-fragment check flakes under host load
8d521ab8 (code 8afbf236) + test: tests/runtime/rx_workflow_fusion.c judges parallel latency only when original showed overlap
Ledger line 103 on GitHub - 2 days agodoneL107
Lineage inventory: every artifact that runs on the Spark, the iMac and the Pi (Atlas, physics, Omega compiler, GPU kernels, AIENOS kernel, NVMe/storage, aien-serve-http, Rust crates, shell tools): what builds it today (gcc/rustc/as/ld/our encoder), size, and which seed-chain rung could birth it; honest gaps in plain words
docs/SEED_LINEAGE.md doc/seed-lineage@fa57eacb (commit verified; PR being opened)
Ledger line 108 on GitHub - 2 days agodoneL114
Founding review of OSC-0 (Drake: was Omega founded correctly? anything better?): each decision KEEP/AMEND/ADD-MISSING vs the 2026-10-19 mission with code citations; what OSC-0 never decided (throughput concurrency, AI as writer, LLM-runtime subset, energy); plain summary
doc/osc0-review@85d3bf90, READY line; proposals 12-15 accepted, decision 25 added (met)
Ledger line 115 on GitHub - 2 days agodoneL116
Fix: QWEN38 gate holds /tmp/aien-gb10.lock around aien-serve-http, whose qwen38 engine takes the same lock for upload + each generation -> server load waits on its own parent (not_ready 900 s); blocks the whole GPU queue incl. the resident 27B on 18091
same fix landed as L068 c3dfb705; proof run belongs to L068 (met by L068)
Ledger line 117 on GitHub - 2 days agodoneL119
spark-connector stream connector: Rust node (join/run/status), sandboxed judge run, fork-only publish to aien-receipts, install.sh + systemd unit; evidence = two receipts on one real head
connector/v0@0426afca READY (met)
Ledger line 120 on GitHub - 2 days agodoneL120
spark-connector stream trust: NODES.md registry, VERDICT_POLICY.md (two registered READY + verify-bundle = merge-grade, spot re-verify), THREATS.md, COMMUNITY.md; evidence = verify-bundle pass + tampered fail
trust/v0@c8235688 READY (met)
Ledger line 121 on GitHub - 2 days agodoneL123
Speed audit (Drake: faster, whole goal): ceilings of the box with sources; where the 27B decode is today from receipts; per-token decomposition (weight bytes, KV bytes, 436 launches, one channel, host C, doorbell); every lever with its arithmetic gain, what exists, and the measurement that decides; foundations and mathematics judged by per-token cost; ranked table; one unassigned ledger line per top-ten lever
doc/speed-audit@ddc00525 SPEED_AUDIT, top-ten lanes L125-L134 opened (met)
Ledger line 124 on GitHub - 2 days agodoneL141
Archives study (Drake): B to C, what Ritchie and Thompson were fixing and how, the bandages that became permanent, where programming came from, and how Omega must differ; roadblocks to decide now as yes/no
note ~/handoffs/2026-10-19-b-to-c-archives.md; decisions 16-24 in PROGRAM.md (met)
Ledger line 142 on GitHub - 2 days agodoneL144
iMac GPU shaders: our own GCN 4 instruction encoder + PM4 packets + first compute kernels (fill, blit, sky lens); worktree aienos-gpu-shader
feat/imac-gpu-shader@5b56323 283/0 vs llvm-mc, 4 kernels PASS, READY (met)
Ledger line 145 on GitHub - 2 days agodoneL145
GPU shader encoder + PM4 + interpreter + kernels feat/imac-gpu-shader
feat/imac-gpu-shader@5b56323 READY in pr-desk (met)
Ledger line 146 on GitHub - 2 days agodoneL148
Fix: sampled decode spends ~28 ms/token on the host (17% of wall, serve.log receipt omega evidence/CHAIN-TO-GPU-1): top_k_mask sorts all 248320 logits, then sample_with_params_traced exps + sorts all 248320 again for top_p
fix pushed, 0.165 ms per call, same tokens, tie mutant red (met)
Ledger line 149 on GitHub - 2 days agodoneL149
Fix: omega tcgemm codegen overflows gx[8]/sa[8]/stg[8] when 4*rt/warps > 8 (rt=4, warps=1 accepted by omega_tcgemm_spec_ok) -> stack smash / segfault
35eecb3c + test (met)
Ledger line 150 on GitHub - 2 days agodoneL153
Fix: quiet-build/quiet-measure --help queue on the quiet lock instead of printing help
live quiet-build+quiet-measure guard; repo aien-architecture-mv fix/quiet-help@0b650b1; handoff ~/handoffs/fixer-quiet-help.md
Ledger line 154 on GitHub - 2 days agodoneL046
Allen canvas: the model decides what is painted (scene lines rendered in Allen style), worktree aienos-allen-canvas
feat/allen-canvas@e57ef1e pushed; canvas test PASS 83, real 27B replies 13/14 clean (r1 10/10 prompt v1, r2 3/4 prompt v2), crumb verify OK 91, READY in desk log; evidence/allen-canvas/
Ledger line 47 on GitHub - 2 days agodoneL066
27B server on 18091 is down since 14:41 (its keepalive cannot reload it: other 27B servers on 18191 and 18291 hold the memory); not touched by the allen-model agent
coordinator 2026-10-19 23:0xZ: 18091 /health says ok, engine omega-qwen38, uptime 7658 s (curl by coordinator)
Ledger line 67 on GitHub - 2 days agodoneL016
Agent: continuous-batching
cascade/continuous-batching@7e69aeda pushed; evidence/qwen38-batch-20261019T1630Z (README, summary.txt/json, compare-all.txt, repeat-*, mem-*); handoff ~/handoffs/2026-10-19-cascade-continuous-batching.md; READY line in pr-desk. Open: prompt-3 path difference cause unknown; speeds taken on loaded box
Ledger line 17 on GitHub - 2 days agodoneL104
Fix: omega RX-HOST-REFERENCE test-action-graph speedup>=2 check flakes under host load (1.99 < 2)
ca8ac474 (code 2f426e9c) + test: fastest run vs serial time with plain-thread host reference; negative control forced to 1 worker fails 8 of 8
Ledger line 105 on GitHub - 2 days agodoneL086
Fix: Cortex World record keeps only first output's field values (rx_cortex_record.c:70); arrangement unit moves never reach the journal
82ad834b canon 100/0 journal 34/0 composition 519/0+85/0 gate PASS; plan-reuse G1 fail pre-existing; r13 episode flaky on base too
Ledger line 87 on GitHub - 2 days agodoneL061
#249 independent review by a different model than the builder (Fable): re-derive 3 accepted rewrites and 3 rejections from the contracts alone; report agree/disagree
evidence/REWRITE-249/REVIEW.md on research/rewrite-elim-249@79c436c2dcdccecefa3d0fdc2e212300738467ab: 6 of 6 agree (Fable reviewer, Opus builder)
Ledger line 62 on GitHub - 2 days agodoneL018
Agent: document-memory
cascade/document-memory@8348affd evidence/qwen38-docmem-20261019/README.md
Ledger line 19 on GitHub - 2 days agodoneL091
Omega runtime R3 portability note + table (omega-mv#596), branch runtime/r3-portability
92cd9145766ef9ef93baa15ad4e7a044b7865f67 + evidence/RUNTIME-R3-PORT/ (test-runtime-portability.log, real-tree-dummy-control.log, gpu-engine-freestanding.log)
Ledger line 92 on GitHub - 2 days agodoneL025
Rig: engine serves best model in 12 GiB budget on Omega
rig/engine-12gb@8eb1b389 pushed, aien-rig engine/core@19b0032 pushed; MXFP4 http gate PASS 3/3 27-29 tok/s 9.1 GiB; FP8 http PASS 3/3 21-23 tok/s 11.6 GiB; evidence docs/inference/evidence/rig-engine-20261019; READY lines in pr-desk
Ledger line 26 on GitHub - 2 days agodoneL087
Fix: dagsim test-dagsim target runs ./$(DS_UNIT), fails with absolute OUT_DIR (branch research/cost-sim-583)
proposal recorded for #583 owner
Ledger line 88 on GitHub - 2 days agodoneL055
Fix install.sh --with-server exit 101: cargo build -p spark-crumbs is ambiguous (path crate + git dep), install.sh line 567; server never built
fix/install-crumbs-ambiguous@4ee467fe, ~/handoffs/2026-10-19-fix-l055.md
Ledger line 56 on GitHub - 2 days agodoneL058
#276 entity pilot in omega: view = projection of committed state (RxWorld+Cortex journal+visor snapshot), admit objective, detach view, work continues, reattach same digest, correction via same commit path, restart via cx_open; unchanged-state control; dagsim prediction before the measured run
research/entity-pilot-276@2191323242f7af2a67fec288729c6b84a3ba2e50 evidence/ENTITY-276/README.md receipt ~/yardsticks/receipts/2026-10-19-entity-276-pilot.json
Ledger line 59 on GitHub - 2 days agodoneL056
serve-27b/qwen38 loader: memory check happens before the 30-min GPU lock wait; re-check memory after the lock is acquired or hold the check until then
fix/serve27b-mem-after-lock@6a4a86fa, handoff ~/handoffs/2026-10-19-fix-l056.md
Ledger line 57 on GitHub - 2 days agodoneL070
aienos.com discoverability: SEO (titles, descriptions, canonical, hreflang, sitemap complete, structured data), AI-agent surface (llms.txt, llms-full.txt, JSON index, robots allows all assistants), social cards (OG/Twitter per page, og image), feeds (Atom for blog and /now/), 404, Lighthouse SEO 100; launch texts drafted
aienos.com#46 merged 1e8e52aa: www canonical, JSON-LD, generated sitemap, robots for AI crawlers, feeds, llms.txt, og cards with @aienscore; Lighthouse 100/100/100
Ledger line 71 on GitHub - 2 days agodoneL015
Agent: w4-gemv
omega cascade/w4-gemv origin evidence/cascade-w4-wm-* (F 119.35,120.57; W 72.60,75.09,76.16; WW 73.46 ms median chip) + host-w4-codelong2; W 13.32 tok/s vs F 8.34 (1.60x), WW 13.61; chip vs host-same-weights cosine mean 0.99977 top pick 303/305; fewer than 4 valid F/WW rounds (shared-memory load failures published); ~/handoffs/2026-10-19-cascade-w4-gemv.md Result section
Ledger line 16 on GitHub - 2 days agodoneL069
aienos.com: every card/box on every page links to its GitHub evidence; whole box clickable with visible Evidence on GitHub affordance; /now/ cards link to their ledger line and shas/issues in notes become links; builder test fails on any box without an href
aienos.com#45 merged 7bc21813: 247 boxes across 11 pages link to GitHub evidence; builder fails on any unlinked box
Ledger line 70 on GitHub - 2 days agodoneL050
Symposium crossover check for #276: point research agent at symposium-dev/symposium + agents-in-tree BRIEF/ADR 0038 + borrowed-ideas PLAN before #583/#584 implementation starts
docs/research/276-crossover-matrix.md@1cd04c8e + arch#276 comment: no overlap with Symposium or agents-in-tree
Ledger line 51 on GitHub - 2 days agodoneL006
Decide: register QWEN38 serve gate + speed gate under tests/gates so signed bundles carry token/speed results (operator decision)
Drake approved 2026-10-19 (desktop chat): register the QWEN38 serve gate and speed gate under tests/gates so signed bundles carry token/speed results
Ledger line 7 on GitHub - 2 days agodoneL039
Allen through the model: UDP courier + relay to 18091 + learning log, worktree aienos-allen-model
aienos-mv feat/allen-through-the-model@f49854f pushed; QEMU e2e PASS 22/22 with real 27B on 18091; handoff ~/handoffs/2026-10-19-allen-through-the-model.md; READY line in pr-desk; NOT tested on real iMac
Ledger line 40 on GitHub - 2 days agodoneL054
www.aienos.com homepage: same inclusivity fix (all machines, speed test, all streams)
aienos.com#44 merged 60ab3e7a, pages deployed; live: Five machines + 7.63 tok/s on www.aienos.com, /machines.json
Ledger line 55 on GitHub - 2 days agodoneL065
Open + merge PRs for org README branch and site/all-machines when READY; desk runs site tests on Mac since no Spark lane for aienos.com
aienos.com#44 (60ab3e7a) + .github#15 (11732d66) merged
Ledger line 66 on GitHub - 2 days agodoneL053
aien-dev GitHub org profile README: cover Spark + iMac 2017 + Pi 5 + QEMU, speed numbers with receipts, everything in flight
aien-dev/.github#15 merged; https://github.com/aien-dev
Ledger line 54 on GitHub - 2 days agodoneL023
Agent: skipped ideas lab
sc research/skipped-ideas@1dc1202c, evidence/skipped-ideas-20261019/README.md (verdicts), handoff ~/handoffs/2026-10-19-skipped-ideas-lab.md; spark-cleanup --only own dirs queued behind quiet-build (pid 3344172), will run when the lock frees
Ledger line 24 on GitHub - 2 days agodoneL052
Public status dashboard at www.aienos.com/status: live agent work, next, decisions made, decisions pending; Spark publisher + site-builder page
live: https://www.aienos.com/now/ and /now.json, built from data/now 16556467; publisher cron */5 on Spark
Ledger line 53 on GitHub - 2 days agodoneL057
Inventory for site + org README: every machine (Spark, iMac 2017, Pi 5, QEMU), every stream in flight, every published speed number with receipt link, every repo with one-line purpose
~/handoffs/2026-10-19-site-inventory.md + .json: 5 machines, 61 streams, 17 speeds, 31 repos, 26 audit findings, 12 gaps
Ledger line 58 on GitHub - 2 days agodoneL029
Rig: speech-to-text on Omega/AIEN
aien-rig speech/v0@af71f46 (omega rig/speech@3aebf9ee); evidence/speech/2026-10-19-gb10/README.md; handoff ~/handoffs/2026-10-19-rig-speech.md; WER 0.83/0.93/1.56%, Omega 2.6-3.2x RT; gpu rerun open as new item
Ledger line 30 on GitHub - 2 days agodoneL009
Fresh-account alpha test (user aienalpha)
VERDICT FAIL as written (install.sh exit 101), steps 3-6 PASS with workaround; handoff ~/handoffs/2026-10-19-alpha-fresh-account-test.md; receipts alpha/fresh-account-evidence@997d22bd evidence/alpha-fresh-account-20261019/
Ledger line 10 on GitHub - 2 days agodoneL028
Rig: YouTube helper (rig-create)
creator/v0@87f98af, CI green at 7aef4de, evidence/creator/2026-10-19-e2e-18091 (exit 0, 49m43s), handoff ~/handoffs/2026-10-19-rig-creator.md
Ledger line 29 on GitHub - 2 days agodoneL012
Flash-Next full download once disk is clean (NVFP4 132.7 GB or FP8 185.6 GB)
NVFP4 nvidia/Qwen3.8-Flash-Next-NVFP4@fc694b54 132.7GB in ~/models/qwen3.8-flash-next-nvfp4-fc694b54; 13/13 LFS sha256 match HF listing, sizes match 25/25, sha256sum -c PASS; SOURCE+sha256 written; FP8 not downloaded
Ledger line 13 on GitHub - 2 days agodoneL024
Rig: Omega backend targets RTX 5070 sm_120 (omega rig/gb20x)
omega rig/gb20x@12e17891 pushed; M17 gates 18/18 (139 cum) evidence/rig/gb10_m17_gates.txt; handoff ~/handoffs/2026-10-19-rig-omega-gb20x.md
Ledger line 25 on GitHub - 2 days agodoneL021
Agent: symbolic-rewrite
omega symbolic/rewrite-layer@75f0adb1, evidence/symbolic-rewrite/VERDICT.md, handoff ~/handoffs/2026-10-19-symbolic-rewrite.md
Ledger line 22 on GitHub - 2 days agodoneL038
Panel off + no restart after goodbye (x86), worktree aienos-panel-off
feat/goodbye-panel-off@2954f5e pushed; handoff ~/handoffs/2026-10-19-goodbye-event-horizon.md (panel-off section), READY in 2026-10-18-pr-desk.md; NOT-TESTED real iMac
Ledger line 39 on GitHub - 2 days agodoneL010
Disk cleanup now + automatic cleanup routine (timer + end-of-session step in skills)
df 518G free (87% used) -> 1138G free (28%, 71% used), freed 617G; timers: spark-cleanup.timer daily 04:00 + spark-cleanup-lowspace.timer every 15 min if under 20% free; handoff ~/handoffs/2026-10-19-disk-cleanup.md
Ledger line 11 on GitHub - 2 days agodoneL001
CEILING.md and alpha plan state BOTH ceilings: one-word memory 10.15 tok/s AND compute ~3,800 (207.7 TFLOPS) / ~4,600 (250) total words/s; promised 2026-10-19 ~16:00Z, never done
sc alpha/ceiling-and-model@56d42869 commit 39e79fad docs/qwen38/CEILING.md (memory 10.15/8.14, compute ~4,050 at 207.7 FP8 to ~4,880 at 250 spec, ~2x FP4, 115.36 bf16 measured; 11 calc receipts) + ~/handoffs/2026-10-19-alpha-readiness.md cascade section updated (old 3,800-4,600 used bytes as weights)
Ledger line 2 on GitHub - 2 days agodoneL002
CEILING.md tells readers to run aien-calc verify but aien-calc is only on unmerged calc/aien-calc: merge it or reword
sc alpha/ceiling-and-model@56d42869 commit 39e79fad docs/qwen38/CEILING.md: every expression written out by hand, calc ids kept as references, aien-calc named as on unmerged calc/aien-calc
Ledger line 3 on GitHub - 2 days agodoneL003
Model download is a manual hf/curl command: give install.sh a --with-model step that downloads + sha256 checks
sc alpha/ceiling-and-model@56d42869 commit ec11d1d9 install.sh --with-model + scripts/test-install.sh (9/9, 4 mutants red) + docs/qwen38/TRY_IT_ON_YOUR_SPARK.md; real run evidence/qwen38-install-with-model-20261019/realrun.log (2 fetched, 1 resumed, 81/81 sha256 OK; banner fix rerun OK)
Ledger line 4 on GitHub - 2 days agodoneL004
Re-measure 2nd-server memory need (38.5 GiB is from the old binary) on the current build
sc alpha/ceiling-and-model@56d42869 commit 6b8b2848 evidence/qwen38-serve-memory-20261019/proc-status.txt: VmHWM 38499160 kB = 36.7 GiB on current binary df936011 (pid 2473716, read only; fresh-account receipts not present); ALPHA_STATUS.md, guide, serve-27b.sh, README.md updated
Ledger line 5 on GitHub - 2 days agodoneL005
Put omega test fix 4afd7597 onto qwen38/decode-speed if that branch stays in use
omega fix/decode-speed-gpr-test@585df2f5 (cherry-pick -x of 4afd7597 onto origin/qwen38/decode-speed 0010f7d3; branch stays in use: live server binary, 6 cascade/flashnext branches based on 0010f7d3). make test-gpu-gpr-boundary at nice 19: 298 checks 0 failed PASS; control at base fails -Werror missing sched. Receipt ~/yardsticks/receipts/2026-10-19-l005-gpr-boundary-585df2f5.log. No PR (desk).
Ledger line 6 on GitHub
How this page is made
One file, no hands
The coordinator and every agent append to one tab separated file on the Spark. A publisher pushes that file to the data/now branch whenever it changes. The site builder reads it at build time and writes this page and /now.json. A line that is not in the file is not here. Every card, queue row and settled entry links to its own line of that file, and the commits, pull requests and evidence folders named in a note link to GitHub. Owners shown as "agent" are autonomous coding agents; the operator is one person.