AIENOSPERSISTENT MACHINE INTELLIGENCE

The Machine

Turning search into verified understanding

AIENOS is an owned experimental machine for turning search into verified understanding. AIEN proposes structure. Omega makes that structure explicit and checkable. The machine tests it against reality under authority. The Turing measures what was actually learned. Cortex keeps the evidence. Then the next search begins from what survived.

Weights suggest. Programs explain. Verification decides. Evidence teaches. The Turing keeps score.

Suggestions and explanations have different epistemic roles. Neural systems generate hypotheses. Explicit artifacts suit inspection and verification. Physical execution provides contact with reality, measurement determines gain, and evidence provides continuity.

The closed loop

The meaningful artifact of this project is a closed epistemic loop. Each pass runs through eleven steps, and the last step feeds the first.

  1. Observe
  2. Propose
  3. Formalize
  4. Verify
  5. Authorize
  6. Execute
  7. Measure
  8. Score
  9. Preserve evidence
  10. Learn
  11. Search again

In the design, each step has an owner. AIEN suggests. Omega formalizes, synthesizes and verifies. FORGE realizes the verified program, and the physical machine executes it.

The Turing then measures explanatory gain, Cortex preserves the evidence, and AIEN and Omega use that evidence to improve the next search. ATLAS carries the lineage across hardware, and AIENOS provides the owned, persistent substrate underneath.

AEGIS and ARGUS sit around the loop. AEGIS guards the boundary of proof and authority, so nothing executes without permission. ARGUS observes what actually happens. Neither one is a step that work passes through in sequence.

This describes the design. The list under What exists today says which parts run now. See the loop diagram on the homepage.

The Turing keeps the accounts

The Turing is the epistemic accounting layer of AIEN. It measures net held-out explanatory compression: one operational component of machine understanding. It is not a general intelligence score.

For a baseline B, a candidate explanation M, sealed observations D and a scoring profile P, with L meaning description length:

T(M; B, D, P) = [L(B) + L(D | B)] - [L(M) + L(D | M)]

A candidate earns positive Turings only if it shrinks the description of unseen data by more than its own complexity costs.

  • Charged: memorization, unnecessary complexity and bad predictions.
  • Prohibited: post-hoc selection. Test contamination invalidates the measurement.
  • Rewarded: compact structure that generalizes.

The question it answers: what did the machine discover that continued to work on reality it had not seen? Explore the Turing.

Three quantities

  • Turing gain (T)

    Did the machine learn explanatory structure? Gain is net held-out explanatory compression after paying for the explanation.

    PASS Run TY-2 recorded 2,559,679.825 bits of Turing gain against an order-1 baseline on held-out seeds.

    The re-check used the same tool, so it is a self re-derivation and still awaits an independent one. This result certifies the instrument; it does not show that AIEN discovered anything.

  • Turing yield (T/J) NOT STARTED

    What physical resources did that gain cost? Yield will divide Turings by joules, and may separate evaluation yield from discovery yield, the more important of the two.

    Turings per joule cannot be computed until energy is measured, so this has not started.

    PARTIAL Energy attribution. The plan for measuring energy is written down and the code has fixture tests, but no timed run on real hardware has happened.

    No yield figure exists, and none will be claimed until energy attribution and the Turing receipt are joined under the qualifying protocol.

  • Search and verification gap

    How hard was the explanation to discover compared with verifying it once found? The idea to test: discovered explanatory structure reduces the cost of future search, through prediction, compression, abstraction, reusable procedure, reduced search and faster discovery.

    This is a research direction. No result on it exists yet.

Evidence that teaches: Cortex

Cortex exists to preserve the evidence needed to make future search better. That means evidence continuity: verified explanations, failed hypotheses, counterexamples, validity regions, measurement receipts, experimental conditions, search traces, physical costs, uncertainty and provenance.

Today Cortex is an append-only store that records what the World did. Search history, counterexamples and costs are what it is meant to hold next; its current standing is listed under What exists today.

With that evidence in place, AIEN will be able to ask:

  • What did we already try?
  • Why did it fail?
  • Under what regime did it work?
  • What evidence supports this abstraction?
  • What representation changes reduced search before?

The discovery ladder

Each rung is a stronger claim about what the machine has shown, and each rests on the rung below. The ladder is drawn from the research status record, the same record behind the full status page, so this page never states a result by hand.

Current level: L2, Brownian scope only, instrument certification. Outside Brownian scope the ceiling is L1. TY-2 does not change the ceiling.

The scorer behaves as specified on known stochastic cases. It does not mean AIEN discovered anything.

Read from the bottom up. Each step rests on the one below it. Filled steps are achieved, half-filled steps are partial, hollow steps are not achieved.

  1. L0Definition

    achieved

    Achieved

    Evidence: QS:101 "met"; definition only

  2. L1Compression bridge, EXP-001R

    achieved

    Achieved (tested regime, amended overlap gate, unsigned)

    Evidence: QS:102; EXP-001R final_receipt.json

    FAILEXP-001failed attempt, kept on the record

    First attempt at the compression bridge. Failed its overlap audit (S2), so no measurement was produced. Kept on the record, not deleted. Led to EXP-001R.

  3. L2Known stochastic structure, EXP-002A-B-CCurrent level

    achieved

    Achieved for Brownian scope only, as instrument certification; verdicts checked against the private sealed records on 2026-10-01

    Evidence: QS:103, QS:114

  4. L3Invariant or unifying structure

    not achieved

    Not achieved (interventional replication PASS, EXP-002D INCOMPLETE)

    Evidence: QS:104

  5. L4Active experimental choice, EXP-003

    not achieved

    Not achieved (blocked, not frozen)

    Evidence: QS:105

  6. L5Energy attribution

    partial

    Partial (protocol and fixture tests; no timed run)

    Evidence: QS:106

  7. L6Turing yield T/J

    not achieved

    Not achieved (blocked on L5)

    Evidence: QS:107

  8. L7+Sensitivity, replication, multiple domains

    not attempted

    Not attempted (L7, L8, L9)

    Evidence: QS:108-110

Every experiment, its receipt and its limits

Self-improvement, measured

The interesting version of recursive self-improvement is narrower than a machine that modifies itself. Experience leads to a measured explanation, then a reusable abstraction, reduced search, better candidate selection, a verified improvement and the next generation.

A self-modification will count only if it survives verification and measurement.

The keystone question: can a persistent machine accumulate verified explanatory structure in a way that measurably improves its ability to discover further structure?

  • Does T increase?
  • Does T/J improve?
  • Does search cost fall?
  • Does verification remain trustworthy?
  • Do representation changes transfer?
  • Do interventions become more informative?
  • Does learned structure survive new domains?
  • Does the system preserve its failures?
  • Can independent scorers reproduce the evidence?

Connections to other research

Each of these is a connection or an opportunity. Where the research status record has a matching entry, its badge sits beside the name.

Petri nets and reaction systems
The World lifecycle work already has a state model, reachable-state exploration, invariants, shortest counterexample traces, mutants and differential checking against the real World code. Formal concurrency theory may offer reachability, liveness and deadlock analysis, partial-order reduction and invariant proofs. The World keeps its own name; it is not being relabelled a Petri net.
Counterexample-guided synthesis (CEGIS) Partially implemented
Omega's search already runs candidate, transform, verify, accept or reject, and records the trace and cost. The opportunity: turn a failed candidate into a semantic counterexample, generalize it into a constraint, eliminate a whole family of candidates, and measure how much search that saves.
Bayesian adaptive search Experimental
A Bayesian cost model for choosing between verified realizations is being tried against recorded signals, and its failed attempts stay on the record. A future direction is safe contextual exploration. Exploration must stay explicitly authorized, and production authority is never inferred from prediction confidence.
Active causal discovery BLOCKED
EXP-003 is designed as the move from passive prediction to active inquiry: hold competing explanations, predict their consequences, choose the intervention that best separates them, run one authorized experiment and update belief. Not run. Its contract is still a draft and must wait for the Brownian records to be sealed.
E-graphs and equality saturation Research hypothesis
A research reference for treating rewritten expressions, basis changes, coordinate systems and equivalent implementations as one discovery. Omega has no e-graph code.

What exists today

This list is drawn from the research status record, built from a check of the code on the main branches of Omega and AIENOS. Built work and open research questions are kept apart, and each entry opens to show its limits and evidence.

Implemented (14)

Code and tests exist. Each entry says how far that goes.

  • Resident World / reactionsImplemented (host reference)

    one shared object world; reactions wake on object and field-mask change; budgets, quarantine, causal crumbs

    Limits and evidence
    Claim level
    host reference only
    Limitations
    R16 merged with evidence-immutable CI failing (inventory.json edited in place); R15 G8 attempt 1 FAIL 0.832, attempt 2 PASS file exists (value not verified); whole-product orchestrator retirement deferred
    Next dependency
    runtime program lanes
    Evidence
    src/runtime/rx_world.h:2-8,73
    Receipt
    tests/runtime/rx_r12_resident.c, rx_r13_living.c, rx_r14_recovery.c; evidence/R16/22d7a79a...json (R16_ORCHESTRATOR_RETIRED=PASS); evidence/R15/
    Commit
    last touch bc60f86; R16 omega#112 3dd5eaa
    Date
    2026-09/10
  • Lifecycle modelingImplemented (host-only)

    explicit state machine of wake/admit/run/commit/invalidate/fail/reclaim/cancel; SPEC, AS_BUILT, mutant profiles

    Limits and evidence
    Claim level
    host only
    Limitations
    4 reactions in fixed tree; CPU only; no CANCEL against rx_world
    Next dependency
    store LG receipt
    Evidence
    tests/model/world_model.h:1-25; mk/proof.mk:9,26
    Receipt
    tests/model/model_check.c, receipt.sh via make test-proof; no LG receipt committed under evidence/
    Commit
    PR #174 0d8b057
    Date
    n/a
  • State-space explorationImplemented (bounded)

    BFS over reachable model states, shortest violation trace

    Limits and evidence
    Claim level
    test code, not shipped runtime
    Limitations
    bounded by WM_EXT_MAX and WM_TICK_MAX; no stored run result
    Next dependency
    none stated
    Evidence
    tests/model/world_explore.c:1-6
    Receipt
    via make test-proof; counterexample replays I2, I4 in world_diff.c
    Commit
    0d8b057
    Date
    n/a
  • Differential checkingImplemented (host-only)

    AS_BUILT model vs real rx_world, all op sequences to depth N; TSan target

    Limits and evidence
    Claim level
    host only
    Limitations
    depth bounded; settled states only; visor/search differential unchecked
    Next dependency
    none stated
    Evidence
    tests/model/world_diff.c:1-30,701-702
    Receipt
    make test-proof logs; no committed receipt
    Commit
    0d8b057
    Date
    n/a
  • Semantic / program artifacts: program idImplemented, narrow

    program identity v2 = canonical body + contract hash; unary u64 ADD/SUB/MUL/AND/OR chains

    Limits and evidence
    Claim level
    narrow
    Limitations
    cannot name matvec or Omega-X; omega_program_realize declared (omega_program.h:117) with no definition per 2026-09-29 audit (not rechecked); M6 self-host is a fixed output
    Next dependency
    general compiler
    Evidence
    src/omega_program.h:41,44,117
    Receipt
    tests/program/test_program_id.c; evidence/PROGRAM_ID/ (2 receipts)
    Commit
    omega#75; 9d17fcc
    Date
    2026-09-29
  • Semantic / program artifacts: action graph IRImplemented

    goal to typed graph to resident reactions; workflow fusion into MetaSkills

    Limits and evidence
    Claim level
    host
    Limitations
    fusion means fewer reactions and CPU, not faster
    Next dependency
    none stated
    Evidence
    src/runtime/rx_graph.h:181-196
    Receipt
    tests/runtime/rx_action_graph.c, rx_workflow_fusion.c; evidence/ACTION_GRAPH/, WORKFLOW_FUSION/
    Commit
    omega#51 df0ce76, #61 ceb68d6; d9ed2e8
    Date
    2026-09
  • Verification infrastructureImplemented

    per-milestone gate scripts, content-addressed receipts, evidence immutability CI

    Limits and evidence
    Claim level
    host
    Limitations
    no shared receipt schema across repos; immutability check bypassed once (R16); no common harness
    Next dependency
    shared schema
    Evidence
    src/omega_evidence.h:48-74
    Receipt
    evidence/omega_verify_qualification_receipt.json; evidence/SHA256SUMS; .github/workflows/evidence-immutable.yml
    Commit
    5c23ec9
    Date
    n/a
  • Search traces (M23)Implemented (corpus half)

    records every candidate the program search touches; byte-identical replay; sealed holdout commitment

    Limits and evidence
    Claim level
    corpus only
    Limitations
    training half waits on M22; hooks off by default; nothing consumes the corpus
    Next dependency
    M22
    Evidence
    spec/searchtrace/M23_SEARCH_TRACE_CORPUS_V1.md:1-30
    Receipt
    evidence/M23/receipts/m23-corpus-ee498211...json PASS (3072 steps, commit 12b4d6e)
    Commit
    a0e59f4
    Date
    n/a
  • Capabilities / authorityImplemented (with conflict)

    capability roots, 64-bit generations, capability graph, skill router; aienos C authority with containment

    Limits and evidence
    Claim level
    host
    Limitations
    Rust and C authorities coexist (CONFLICTING per audit); Effect Broker partial (Rust only); machine identity has 3 shapes
    Next dependency
    pick one authority
    Evidence
    src/runtime/rx_capq.h:227-263; aienos_cap.h:70-91
    Receipt
    tests/runtime/rx_capability_graph.c, rx_capability_query.c, tests/effect/test_effect_cap64.c; evidence/CAPABILITY_QUERY, EFFECT_CAP64; aienos native/capability tests; 7 Rust tests
    Commit
    omega#53 670bf98, #71 8e7a445
    Date
    2026-09
  • Turing measurement machineryImplemented (instrument); science partial

    field, selector, replay, compression stream (tc_*), Turing Yield units (ty_*) including energy attribution code

    Limits and evidence
    Claim level
    instrument
    Limitations
    science state is in Table 1; chooser failed kill test, now record-keeping only; TURING not wired into live runtime (QS:31 NOT STARTED)
    Next dependency
    live wiring, H4
    Evidence
    QS:19-29; TY2:1-17
    Receipt
    tests/turing/test_turing.c, test_tc.c, test_ty_energy.c, test_tc_eval.sh; ty2_heldout_receipt.txt
    Commit
    a0e59f4
    Date
    n/a
  • Cortex componentsImplemented (host reference, omega C) plus Rust twin

    omega: append-only digest-chained store, journal, single writer, promote, recall, World recording (single process). aienos: Rust store, record, status, verification

    Limits and evidence
    Claim level
    host, single process; "implemented, not qualified" per architecture
    Limitations
    two implementations (C vs Rust, CONFLICTING); thin Rust tests; no Visor consumer; predictive memory (EST-8) PLANNED
    Next dependency
    resolve duplicate, qualification
    Evidence
    rx_cortex.h:1-14,170-225; ROADMAP.md:313
    Receipt
    tests/runtime/rx_cortex_canon.c; 2 Rust tests in aienos-cortex store.rs
    Commit
    omega 00160d9 (M20)
    Date
    n/a
  • AEGISImplemented (host, resident authority faculty)

    verifier and policy faculty: omega R8 resident AEGIS (request, decide, root mint through native AIENOS authority) plus omega aegis_verification and a Rust aienos-aegis crate (capability graph, effect broker, evaluator)

    Limits and evidence
    Claim level
    host (covers only the R8 resident AEGIS receipt; the Rust aienos-aegis crate and omega src/aegis_verification.c are present in the repos but no receipt is cited here for them)
    Limitations
    receipt says host processor, authority runs as a host library, not an AIENOS kernel; policy is not learned at run time; receipt predates later runtime edits (omega rx_aegis.c last touched fed4f07, R16 C6, 2026-09-30) and was not rerun; the Rust aienos-aegis crate and omega src/aegis_verification.c were not read or run by scout
    Next dependency
    native AIENOS kernel authority; shared receipt schema
    Evidence
    omega spec/r8-aegis-resident.md; src/runtime/rx_aegis.h:1-40; tests/runtime/rx_r8_aegis.c; aien-architecture doctrine/ROADMAP.md row R8; docs/adr/0013-machine-physics-lowering-and-aegis-invariant-checker.md (Accepted)
    Receipt
    evidence/R8/a666db9f3d2fd9f6b01be6035935cdfa5b3d233b03100014fedabb31fcf04181.json (R8_CONSTITUTIONAL_AEGIS PASS, 116 checks, 0 failures, clean tree, candidate fd57ed5)
    Commit
    omega#48 fe4924a
    Date
    2026-09-28
  • ATLAS (M1 ATLAS_BOOT and roadmap)Implemented, narrow (QEMU qualified; native hardware pending)

    1,480-byte bootstrap seed atlas.bin, dual-seam verified in QEMU virt AArch64 (301 instructions word-reconciled, 268 of 268 corruptions refused); first rung of the 41-row roadmap (M0 to M40)

    Limits and evidence
    Claim level
    QEMU only
    Limitations
    ATLAS_BOOT_NATIVE_PASS pending and decoupled from QEMU; atlas master still carries .py gate scripts (run_milestone1_gates.py, seam2_execution_harness.py, generate_audit_ledger.py) beside the C/shell suite; roadmap table counts at read time: 20 rows COMPLETE variants, 21 PLANNED (M20 to M40 all PLANNED); roadmap status is doc-level, not rerun
    Next dependency
    native hardware boot (paused until TRUST-1 owner-key chain); M20 OMEGA_TENSOR opens the planned half
    Evidence
    aien-architecture doctrine/ROADMAP.md rows M0-M40 and lines 87-95; atlas qualification_receipt.json
    Receipt
    atlas qualification_receipt.json (qualification_status COMPLETE_QEMU_QUALIFIED, sha256 f7802501...b9a5)
    Commit
    atlas fcc3874 (seed, 2026-09-26); atlas 0ec17d4 (C11 and shell gate suite); roadmap table aien-architecture f81e4ce (doctrine/ROADMAP.md, 2026-10-01)
    Date
    2026-09-26
  • Typed result contractsImplemented, narrow (synthetic backend)

    9 typed contracts, constrained generation versus free generation, and a publish gate that is the sole writer of the published object (rx_contract)

    Limits and evidence
    Claim level
    host, synthetic cognitive backend
    Limitations
    receipt not_claimed: neural or LLM backend, live R11 reactions through the gate, valid-but-wrong answers, lookahead, GPU execution, AIENOS kernel; constrained invalid rate 1.26% vs free path (measured 32% per memory note, not read from the JSON by scout); semantic failure rate similar on both paths; not rerun
    Next dependency
    real cognitive backend; R11 routed through the gate; efficiency golden path capstone
    Evidence
    omega src/runtime/rx_contract.c, rx_contract.h; tests/runtime/rx_typed_results.c; spec/typed-result-constraints.md:1-12; receipt tail gates block
    Receipt
    evidence/TYPED_RESULTS/b75d295fa4007e13bcb443d7dd4e3c3246a58e7acf4d553d99712e98937b0bbb.json (OMEGA_TYPED_RESULT_CONSTRAINTS_PASS, 47,062 checks, 0 failures, clean tree)
    Commit
    omega#59 1b63b20 (candidate d8da311)
    Date
    2026-09-28

Partially implemented (4)

Some of it exists. The missing part is named.

  • CEGISPartially implemented, unverified (treat as not established)

    only the word appears in synthesis path; no dedicated module or test found

    Limits and evidence
    Claim level
    none
    Limitations
    scout did not read the matching lines
    Next dependency
    read src/omega_synthesis.c, omega_realize_synth.c
    Evidence
    grep hits in those two files
    Receipt
    none
    Commit
    none
    Date
    n/a
  • Evidence receiptsPartially implemented

    digest-named JSON receipts, run commit and dirty flag; aienos Rust evidence tool (capture, verify, rollback)

    Limits and evidence
    Claim level
    partial
    Limitations
    no shared schema; BuildIdentity planned (TRUST-1), not code
    Next dependency
    shared schema, BuildIdentity
    Evidence
    src/omega_evidence.h:48-74; TURING_CURRENT_STATE.md rows Provenance, BuildIdentity
    Receipt
    evidence/*.json; aienos ck_gates_*.json; 36 Rust tests in aienos-evidence
    Commit
    5c23ec9
    Date
    n/a
  • ARGUS (ARGUS-0, ARGUS-1)Partially implemented (ARGUS-0 complete under documented limits; ARGUS-1 containment partial)

    defensive plane in C: 128-byte event ABI, rings, resident core, hard-invariant detectors, omega producer; ARGUS-1 adds ABI v1.2, bridge to AEGIS and one narrow revoke (everything else escalates)

    Limits and evidence
    Claim level
    host
    Limitations
    performance passes only with the consumer on a spare core (+3.87% R8; unpinned +5.31% fails the 5% bar); ADR 0017 2% clock target unmet, amendment pending Drake; hostile review 63 tests, 58 defended, 3 expected-fail, 2 N/A; ARGUS-1 G6 offline checker not delivered so G13 stays open; Mac G12 run not done; freeze is an explicit stand-in; no synchronous blocking; ARGUS never authorizes (AEGIS does)
    Next dependency
    ARGUS-1 G6 offline evidence checker; Mac regression run; owner decision on 2% target
    Evidence
    aienos native/argus/docs/ARGUS0_GATES.md:3; ARGUS1_GATES.md:6,29,53,62,70; aien-architecture docs/adr/0017-argus-defensive-plane.md:3
    Receipt
    aienos native/argus/docs/ARGUS0_GATES.md (9 of 10 gates PASS, hostile and performance gates PASS-WITH-DOCUMENTED-LIMITS); native/argus/docs/ARGUS1_GATES.md (G8 PASS, G10 PASS 47,283 events 0 proposals, G11 implementation check PASS, G12 Linux PASS, Mac pending, G13 partial); omega evidence/ARGUS/h2-verify/VERIFY.md and evidence/ARGUS/speed2
    Commit
    aienos#160 7523933; omega#70 470dcae; aienos#161 d39dd5b; aienos#163 e954b0b; aienos#170 efadc17; aienos#172 603c91d; ADR 0017 aien-architecture#48 10cd118 and #53 b19b4ac
    Date
    2026-09-29
  • FORGE (v1 and V2)Partially implemented (v1 qualified on host and GB10; V2 and analog simulation are experimental, no physical device)

    machine realization subsystem: v1 FORGE-0 realize/verify seam and FORGE-HWID observed GB10 descriptor; V2 substrate-neutral descriptor and evidence encodings with known-answer tests; AR2 simulated analog provider with digital oracle parity

    Limits and evidence
    Claim level
    host and GB10 for v1; V2 host-only simulation, no physical analog claim
    Limitations
    physics docs/FORGE_SUBSTRATE_V2_SPEC.md still says Status DRAFT; golden vectors are synthetic fixtures and state no physical run; no analog device exists (AR4 needs an operator device decision); the 47 KAT + 6 gates count in memory notes was not reproduced by scout
    Next dependency
    AR3 calibration and uncertainty; AR4 physical device decision; AR5 omega multi-substrate selection
    Evidence
    aien-architecture docs/adr/0018-substrate-neutral-physical-realization.md:3,195-206 (Accepted 2026-09-29); doctrine/FORGE.md; doctrine/ROADMAP.md Gates 3-4 and 14 rows; physics docs/FORGE_SUBSTRATE_V2_SPEC.md:3-9; docs/forge-v2-golden-vectors.md:5-7
    Receipt
    omega evidence/FORGE-GATES/702cf450534cf2727e4846243d00e7523b8252a95601d159ca8dd5d392ccafb7.json (Gate 3 FORGE-0 PASS, Gate 4 FORGE-HWID PASS, sm_121, clean trees); omega evidence/GATE14-FOUNDATION/50dd611b...json; V2: no committed receipt, only tests/test_forge_v2_kat.c, tests/run_forge_v2_gates.sh and regenerated docs/forge-v2-golden-vectors.md in physics
    Commit
    physics#13 5969159 (v1, 2026-09-29); physics#16 1f7c321 (V2, 2026-09-29); physics#21 d52759d (AR2, 2026-10-01); omega#111 5517d22 (Gate 14)
    Date
    2026-09-29/10-01

Experimental (1)

Being tried. Results so far, including failures, are on the record.

  • Bayesian cost prediction (belief/estimation)Experimental

    Kalman filter, mixtures, predictors, v4 families vs recorded signals

    Limits and evidence
    Claim level
    experimental
    Limitations
    v1 and v2 FAIL on record (heavy tails, idle regime); v4 selects among screen-passers on one-step log score; EST-5 uncertainty-aware cost model not found as code; rx_costmodel is matvec only. Architecture: EST-3 v1-v3 FAIL, v4 no verdict, EST-4..10 blocked
    Next dependency
    declared loaded period; EST-5
    Evidence
    docs/estimation/EST3C_FORENSICS.md:3,97-119; est-v4/RESULT.md
    Receipt
    docs/estimation/receipts/est23-v1 FAIL, est23-v2 FAIL, est3c-v3, est-v4/RESULT.md (family G1 lag PASS, 3 void attempts kept)
    Commit
    omega#105; est-v4 PR #153 6ff7c30
    Date
    2026-09-30/10-01

Planned (3)

Written down as a plan or specification. No code yet.

  • Full Physics ZeroPlanned

    specification and plan only; M35 PHYSICS_ZERO_REAL_LAB PLANNED; no AIEN-P0 clean lineage

    Limits and evidence
    Claim level
    none
    Limitations
    README.md:30 lists it as spec/plan only
    Next dependency
    M35
    Evidence
    README.md:30; QS:92; architecture ROADMAP.md:74
    Receipt
    none
    Commit
    none
    Date
    n/a
  • Self-accelerating RSI via measured abstractionPlanned (doctrine term)

    no code symbol

    Limits and evidence
    Claim level
    none
    Limitations
    RSI promotion discipline is spec only; full RSI integration out of scope for R16
    Next dependency
    later phases
    Evidence
    TURING_CURRENT_STATE.md:56; spec/path-semantic-object.md:462-463; docs/r16-operator-brief.md:532
    Receipt
    none
    Commit
    none
    Date
    n/a
  • Dirac (DIRAC-0 and sealed exam ladder)Planned (spec, ADRs and oracle design only; all gates NOT_RUN)

    integration workstream that would test generic Omega algebra, CPU and GB10 realization, and blind discovery against a hidden Dirac evaluator; ADR 0026 places Dirac as first rung of a sealed exam ladder (Zeta, symmetry, SUSY)

    Limits and evidence
    Claim level
    none
    Limitations
    ADR 0023 and ADR 0026 are PROPOSED, not accepted; oracle C file exists but was never compiled or run; M20 OMEGA_TENSOR semantic layer and CPU realization merged via omega#136 (c7f9fcd, 2026-10-01) but is not qualified (GB10 NOT_RUN, M20 not qualified); no GB10 tensor realization; G3 sealing BLOCKED_OPERATOR so no sealed dataset; no T/J (TURING H5 NOT STARTED); ESTIMATION v4 is recorded INCONCLUSIVE in omega#153 (merged 2026-10-01), v5 in progress as omega#170 (open)
    Next dependency
    operator acceptance of the spec (D-01), then D-02 oracle build; M20 OMEGA_TENSOR qualification; G3 sealing; TURING H5
    Evidence
    aien-architecture docs/plans/dirac/DIRAC_0_CURRENT_STATE.md (gates NOT_RUN); docs/adr/0023-dirac-0-program.md:3; docs/adr/0026-sealed-discovery-examinations.md:3; omega research/dirac-oracle/README-NOT-BUILT.md
    Receipt
    none; no evidence/DIRAC-0/ receipt exists; gates DIRAC_PREP_FROZEN, DIRAC_D0_ALGEBRA_PASS, DIRAC_CPU_REALIZATION_PASS, DIRAC_D1_REALIZATION_PASS, DIRAC_D2_DISCOVERY_PASS all NOT_RUN
    Commit
    aien-architecture#83 6ebd2fe (spec + ADR 0023 PROPOSED); aien-architecture#97 5910dff (ADR 0026 PROPOSED); omega#158 458c57f (oracle spec, NOT BUILT)
    Date
    2026-10-01

Research hypothesis (5)

Not built. An idea to test, not a feature.

  • General causal discoveryResearch hypothesis

    none

    Limits and evidence
    Claim level
    none
    Limitations
    grep "causal discovery" in omega: 0 hits; only "causal crumbs" (provenance)
    Next dependency
    n/a
    Evidence
    omega grep
    Receipt
    none
    Commit
    none
    Date
    n/a
  • Full active experimentationResearch hypothesis (blocked)

    appears only as EXP-003 spec

    Limits and evidence
    Claim level
    none
    Limitations
    policy interface in protocol .tex:20,45,1281 only
    Next dependency
    EXP-003
    Evidence
    QS:29
    Receipt
    none
    Commit
    none
    Date
    n/a
  • Equality saturationResearch hypothesis

    none

    Limits and evidence
    Claim level
    none
    Limitations
    only prior-art text, no e-graph code
    Next dependency
    n/a
    Evidence
    docs/turing/TURING_PRIOR_ART_MATRIX.md:25,55,147
    Receipt
    none
    Commit
    none
    Date
    n/a
  • General representation discoveryResearch hypothesis

    none

    Limits and evidence
    Claim level
    none
    Limitations
    0 hits; related only fixed realization families src/algebra/realize_*.c and narrow omega_discovery.c (not read in detail)
    Next dependency
    n/a
    Evidence
    omega grep
    Receipt
    none
    Commit
    none
    Date
    n/a
  • SUSY / Zeta rediscoveryResearch hypothesis

    none

    Limits and evidence
    Claim level
    none
    Limitations
    0 hits in omega
    Next dependency
    n/a
    Evidence
    omega grep
    Receipt
    none
    Commit
    none
    Date
    n/a

The same list on the research status page, beside every experiment.

Nature invented the loop

Evolution runs variation, selection and retention. The immune system runs variation, binding and amplification. Science runs hypothesis, experiment and theory. AIEN runs proposal, verification, evidence and memory.

These are analogies and not equivalences. What they share is a pattern: expensive generation paired with decisive filtering, where retained structure changes what gets generated next.

Nature invented the loop. This work proposes a unit for one measurable thing the loop can produce: explanatory structure that survives unseen reality. The aim is to make the loop explicit, persistent, owned, auditable, metered, reproducible and bounded by authority.

Next: Scientific Discovery, the examinations ahead.