Complete source-derived record

Every result. Every reversal. Every reason.

This is the analytical layer between the raw logs and the project narrative. It preserves all 73 dated ledger entries, then expands the failure chains that changed the direction of the work.

Oczy commit
7c9090c297fd
Ledger SHA-256
a29482a4e5901d42…
Integrity classes
63 valid · 10 partial
Source-summary audit
Mismatch: source summary totals 72
The rule—and one live defect

Process success is not scientific success. A null is not a broken run. A blocked gate is not a refutation. An invalid headline does not erase the valid diagnosis that exposed it. The source ledger’s summary currently says 63 valid + 9 partial, while its 73 rows actually contain 64 valid + 9 partial; this publication reports the row audit and preserves the mismatch.

Causal histories

What failed, why, and what changed next

These are the failure chains most likely to be misunderstood when only the final verdict is shown. Each attempt names the layer that failed and the claim that survived.

04 attemptsMeasurement repairThe project first had to learn which apparent gains were produced by the instrument itself.
  1. June aggregateSUPERSEDED
    Tried
    Ranked organs using a saturated aggregate.
    Observed
    NeuralHippocampus scored 1.000 above Oracle 0.844.
    Root cause
    The score rewarded internal bookkeeping rather than learned behavior.
    Response
    Retired the ranking and rebuilt behavior-only metrics.
    Meaning
    The architecture did not beat the oracle; the metric measured the wrong object.
  2. June reranker windowINVALIDATED
    Tried
    Used scope-slot reranking in curriculum and A/B claims.
    Observed
    Three silent bugs prevented the intended reranker behavior.
    Root cause
    Broken control flow, scoring, and update behavior compounded silently.
    Response
    Fixed the reranker and marked affected sections partial rather than deleting them.
    Meaning
    Pre-fix comparisons cannot support claims even where their numbers look plausible.
  3. Stage 2 / Stage 5SUPERSEDED
    Tried
    Measured scope and retention with episode-level teaching helpers.
    Observed
    Headline 1.00 scores fell after repair.
    Root cause
    Expected answers entered teaching through episode IDs and prefix targets.
    Response
    Removed both leakage paths and published the honest baseline.
    Meaning
    Those perfect scores were evidence leakage, not generalization.
  4. 13.5x breakthroughRETRACTED
    Tried
    Changed four parameters together and reported large drift.
    Observed
    Survival ratio 0.354; control logits rose more than target logits.
    Root cause
    Magnitude inflation and common-mode posture were mistaken for targeted learning.
    Response
    Ran single-variable ablations and required trajectories plus seeds for future breakthroughs.
    Meaning
    The loud internal movement was not causal competence.
Read the connected analysis →
03 attemptsResearch 18: distillation evidence lineageA seemingly positive student result remained inadmissible until the teacher, split, and tuning firewall were clean.
  1. Early prototypeINVALID / INCONCLUSIVE
    Tried
    Trained LoRA from corrected responses and measured a 0.333 holdout delta.
    Observed
    A real engineering prototype produced an attractive number.
    Root cause
    Training used corrected_response, holdout signal overlapped training, the holdout was repeatedly tuned against, only two seeds ran, and adapter scope changed.
    Response
    Introduced DEV-only tuning, a one-shot holdout firewall, a fixed teacher gate, and a five-seed diagnostic.
    Meaning
    The early 0.333 cannot be cited as evidence for H-DISTILL.
  2. Five-seed diagnosticBLOCKED
    Tried
    Ran five clean seeds at the registered fallback.
    Observed
    Student deltas {0.3333, 0.3333, 0, 0.3333, 0.3333}; mean 0.2667.
    Root cause
    Teacher DEV delta was identically 0.17647, below the unchanged 0.2 validity gate.
    Response
    Preserved all seeds and refused to lower the threshold.
    Meaning
    The student values are diagnostic only; no accept or refute verdict is permitted.
  3. Mechanism diagnosticsBLOCKED, EXPLAINED
    Tried
    Tested teacher ceiling, prompt contract, fallback template, and training trajectory.
    Observed
    No prompt defects; chat fallback 0; loss 0.70 to 0.16; weak unstable DEV behavior.
    Root cause
    Teacher expressivity and task/prompt ceiling, not malformed prompts or saturated optimization.
    Response
    Retired identical reruns and moved to R19 interface diagnosis.
    Meaning
    The blocker is localized instead of averaged away.
Read the connected analysis →
04 attemptsResearch 19: four attempts to reach one interpretable gateOnly the fourth attempt produced admissible DEV evidence, and that evidence blocked the learned latent interface.
  1. v1INFRASTRUCTURE FAILURE
    Tried
    Load the frozen model in an offline Kaggle kernel.
    Observed
    Model resolution failed before calibration.
    Root cause
    The loader addressed the hub ID instead of the attached local snapshot.
    Response
    Added explicit offline model-path resolution.
    Meaning
    No scientific evidence was produced.
  2. v2INFRASTRUCTURE FAILURE
    Tried
    Run calibration from a commit-addressed source bundle.
    Observed
    Source provenance failed and request features exploded.
    Root cause
    Path/hash handling and unnormalized frozen features.
    Response
    Fixed archive provenance, L2-normalized features, and failed closed on nonfinite inputs.
    Meaning
    No scientific evidence was produced.
  3. v3ARTIFACT COLLECTION FAILURE
    Tried
    Run calibrate-dev after the computation fixes.
    Observed
    Metrics were produced but the campaign collector could not retrieve them.
    Root cause
    Artifacts were written outside /kaggle/working.
    Response
    Rooted all outputs in the collector-visible directory.
    Meaning
    Execution success without collected provenance is not admissible evidence.
  4. v4SCIENTIFICALLY BLOCKED
    Tried
    Run and collect the matched latent-control calibration.
    Observed
    Oracle ceiling 0.357143 passed; Arm B did not beat random-cortex C1.
    Root cause
    The learned latent coupler failed the registered DEV articulation gate.
    Response
    Withheld signoff and holdout access; opened interface and C7 adapter diagnosis.
    Meaning
    The organ can speak when told in text, but the learned latent protocol does not yet control it.
Read the connected analysis →
10 attemptsResearch 20: implementation, cutover, and executor limitsEngineering failures were repaired without converting them into scientific results or relaxing meta-test gates.
  1. DEV smoke v1INFRASTRUCTURE FAILURE
    Tried
    Load frozen Qwen with offline enforcement.
    Observed
    LocalEntryNotFoundError before the run.
    Root cause
    No OCZY_MODEL_DIR resolution under offline mode.
    Response
    Added fail-closed local snapshot resolution.
    Meaning
    Loader repair only; no hypothesis evidence.
  2. DEV smoke v2INFRASTRUCTURE FAILURE
    Tried
    Backpropagate through frozen-organ features into the cortex.
    Observed
    Inference tensors could not be saved for backward.
    Root cause
    torch.inference_mode produced tensors incompatible with trainable downstream layers.
    Response
    Used torch.no_grad so organ parameters stay frozen while feature gradients remain usable.
    Meaning
    Gradient plumbing repair only.
  3. DEV smoke v3INFRASTRUCTURE SUCCESS
    Tried
    Run one outer step, causal controls, deletion audit, and checkpoint.
    Observed
    Exit 0, organ hash unchanged, trace count 0; DEV score 0.0 and most deltas 0.
    Root cause
    The minimum-scale smoke was designed to prove execution, not power a claim.
    Response
    Recorded diagnostics as non-scientific and kept meta-test disabled.
    Meaning
    The mechanism path runs; H-META-CORTEX remains untested.
  4. INT8 first wavePARTIAL EXECUTOR SUCCESS
    Tried
    Train five seeds with 30 tasks per family and two outer steps.
    Observed
    d2 and d4 completed; d0, d1, d3 hit the 12-hour Kaggle limit.
    Root cause
    The registered work exceeded the CPU execution window for three seeds.
    Response
    Preserved failures and retried the same seeds with fewer developmental tasks.
    Meaning
    A provider-limit result, not a null for the learning hypothesis.
  5. INT8 second wavePARTIAL EXECUTOR SUCCESS
    Tried
    Retry the missing seeds with 15 tasks per family.
    Observed
    d0h and d3h completed; d1h again exceeded the limit.
    Root cause
    One canonical seed still ran past the execution envelope.
    Response
    Created uniquely named d1q8 changing only task count to eight.
    Meaning
    Four verified checkpoints exist; the fifth remains a prerequisite.
  6. INT8 third waveEXECUTOR SUCCESS
    Tried
    Run d1q8 under a durable systemd-owned scheduler.
    Observed
    d1q8 completed; checkpoint runtime-manifest verified.
    Root cause
    Eight tasks per family fit the 12-hour window.
    Response
    Published all five checkpoints as an immutable opaque-archive dataset.
    Meaning
    All training prerequisites satisfied; calibration may proceed.
  7. Cal v3 transportINFRASTRUCTURE-INVALID
    Tried
    Dispatch 90 calibration shards with the registered archive.
    Observed
    All 90 jobs failed before experiment code.
    Root cause
    Kaggle auto-extracted the archive; bootstrap required a regular file for SHA-256 verification.
    Response
    Introduced opaque .tar.gz.bin transport and republished datasets under new immutable identities.
    Meaning
    No scientific evidence; a transport defect, not a hypothesis result.
  8. Cal v5 fanoutPARTIAL EXECUTOR SUCCESS
    Tried
    Run 116 four-task jobs covering every registered cell.
    Observed
    29 succeeded, 6 timed out at the 12-hour limit; two failed ranges split into two-task replacements.
    Root cause
    Four-task width has no deterministic margin under 12 hours.
    Response
    Preserved failed attempts and created immutable narrower replacements.
    Meaning
    Executor evidence only; no scientific aggregation yet.
  9. v5 partition auditINVALID FOR SCIENTIFIC AGGREGATION
    Tried
    Aggregate v5 shard metrics into DEV calibration.
    Observed
    Shard width changed the C6 state-swapped donor and the scored state-addressing delta.
    Root cause
    Donor selection was partition-dependent: the next task inside the current shard.
    Response
    Human-authorized version fix; partition-independent donor selection; full width-3 rerun.
    Meaning
    Every v5 shard is valid runtime evidence, but no v5 metric or verdict may be derived.
  10. Cal v6 width-3CANARY PENDING
    Tried
    Run 150 three-task jobs with partition-independent donor selection.
    Observed
    Width-3 canary pending behind five draining legacy kernels.
    Root cause
    Corrected campaign awaits canary promotion before fanout.
    Response
    Durable systemd chain verifies provenance, runtime, CPU-only hardware, and C1–C6 presence before promotion.
    Meaning
    No calibration or scientific verdict yet.
Read the connected analysis →

Authoritative historical classifier

All 73 dated records

The classification is copied from Oczy’s ledger; the outcome filter is derived only to make the record navigable. Open any row for the exact reason and successor.

Showing 73 of 73 records

  1. extended learning evaluationInvalidated or supersededPARTIAL

    What the record says

    Aggregate ranking put NeuralHippocampus at 1.000 (above the Oracle 0.844) — a measurement artifact: it scored internal bookkeeping, not learned behavior. Caught by the 2026-06-21 review (see NOTES.md). Ranking table is SUPERSEDED; the artifact diagnosis is VALID.

    Superseded by

    NOTES.md (2026-06-21 review)

  2. organism curriculum and lm perceptionValid recordVALID

    What the record says

    Organism curriculum + LM perception design; predates the reranker, leakage, and gameable metrics

    Superseded by

    Nothing. This remains the current ledger classification.

  3. cortex kv contractValid recordVALID

    What the record says

    Cortex KV contract design; predates all three invalidation events

    Superseded by

    Nothing. This remains the current ledger classification.

  4. cortexagent raw hidden steeringValid recordVALID

    What the record says

    CortexAgent raw-hidden steering probe; predates all three invalidation events

    Superseded by

    Nothing. This remains the current ledger classification.

  5. prefix steering pocValid recordVALID

    What the record says

    Prefix steering PoC; predates all three invalidation events

    Superseded by

    Nothing. This remains the current ledger classification.

  6. svd init proj c persistenceValid recordVALID

    What the record says

    SVD-init proj_c persistence; predates all three invalidation events

    Superseded by

    Nothing. This remains the current ledger classification.

  7. embedder fork mock foreignValid recordVALID

    What the record says

    Ingestion embedder architecture; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  8. hybrid consolidation architectureValid recordVALID

    What the record says

    Architecture S vs H consolidation; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  9. ingestion pipeline scaffoldValid recordVALID

    What the record says

    Ingestion pipeline scaffold; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  10. multi fact stressorValid recordVALID

    What the record says

    Multi-fact stressor (mock driver); no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  11. multi fact stressor prefixValid recordVALID

    What the record says

    ReservedPosition prefix stressor; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  12. multi fact stressor real driverValid recordVALID

    What the record says

    Real-driver multi-fact stressor; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  13. needle per turn stressorValid recordVALID

    What the record says

    Needle-per-turn stressor tests; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  14. needle sweep scriptValid recordVALID

    What the record says

    Needle sweep benchmark script; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  15. policy head ranking loopValid recordVALID

    What the record says

    Policy head ranking loop (pre-reranker era); curriculum numbers from policy head path, not reranker

    Superseded by

    Nothing. This remains the current ledger classification.

  16. salience threshold ablationValid recordVALID

    What the record says

    Salience filter ablation for ingestion; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  17. auto consolidate shValid recordVALID

    What the record says

    Auto-consolidate S vs H probe; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  18. contrastive cvec discoveryValid recordVALID

    What the record says

    Contrastive cvec discovery and logit biasing; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  19. cortex hippocampus prefixValid recordVALID

    What the record says

    Hippocampus-derived prefix in CortexAgent; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  20. cvec prefix composition tradeoffsValid recordVALID

    What the record says

    Cvec+prefix composition tradeoffs; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  21. deprecate auto prefixValid recordVALID

    What the record says

    Deprecation of stressor auto-prefix; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  22. domain recall metricValid recordVALID

    What the record says

    Domain-level recall metric; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  23. foreign minilm embedderValid recordVALID

    What the record says

    Foreign MiniLM embedder integration; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  24. hippocampus auto prefixValid recordVALID

    What the record says

    Hippocampus-derived auto-prefix (stressor wrapper); no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  25. hybrid cap shValid recordVALID

    What the record says

    Configurable hybrid consolidation cap; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  26. knowledge store prefix targetsValid recordVALID

    What the record says

    KnowledgeStore prefix_targets integration; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  27. length 4096 needle sweepValid recordVALID

    What the record says

    Real-driver needle sweep at length 4096; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  28. memory per byte shValid recordVALID

    What the record says

    Memory-per-byte S vs H probe; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  29. multi fact embedder comparisonValid recordVALID

    What the record says

    Same-LM vs foreign-MiniLM embedder comparison; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  30. paraphrase modeValid recordVALID

    What the record says

    Paraphrased-query multi-fact stressor; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  31. prefix targets paraphraseValid recordVALID

    What the record says

    Query+target-aware hippocampus prefix; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  32. real driver needle sweepValid recordVALID

    What the record says

    Real-driver needle sweep (length 512); no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  33. use agent prefix validationValid recordVALID

    What the record says

    Live CortexAgent hippocampus prefix validation; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  34. lane 01 desaturationInvalidated or supersededPARTIAL

    What the record says

    Desaturation-count acceptance criterion (a metric about metrics) retired by S0.4; sub-metric set now frozen. Lane mechanism itself valid.

    Superseded by

    2026-07-01_remediation_audit.md (Finding 3)

  35. lane 02 kv slot injectionValid recordVALID

    What the record says

    Lane 02 KV-slot fact injection (isolated lane experiment); no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  36. lane 03 layer L extractionRefuted or retractedVALID

    What the record says

    Lane 03 layer-L extraction refutation. Confirmed by S1.4 (2026-07-01_s1_4_hf_layer_probe.md) on two architectures via HF substrate — the refutation is a model property, not a llama.cpp keyhole.

    Superseded by

    Nothing. This remains the current ledger classification.

  37. lane 04 context attractorsValid recordVALID

    What the record says

    Lane 04 context-addressed slot store (isolated lane experiment, separate slot store); no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  38. lane 05 metabolism statusInvalidated or supersededPARTIAL

    What the record says

    Coverage-as-score (1.0) retired by S0.4 — split from the honest lane_05_result (= 0.0). Coverage and result are now separate outputs. See retirement addendum in-file.

    Superseded by

    2026-07-01_remediation_audit.md (Finding 3)

  39. lane 06 bounded growthInvalidated or supersededPARTIAL

    What the record says

    A0b "seed-regenerable" autoencoder met the byte target by regenerating a random matrix from a seed — by construction learned updates can't persist. Compression metric passed; thesis (compress *learned* experience) abandoned in that move. Flagged by the audit.

    Superseded by

    2026-07-01_remediation_audit.md (Finding 3)

  40. lane 07 world model criticInvalidated or supersededPARTIAL

    What the record says

    Headline marker_free_uptake_gap = 1.0 SUPERSEDED — lexical baseline was "0 by construction"; replaced with competitive token-overlap baseline, gap collapsed to 0.0 (S0.4). Mechanism analysis and TD(0) notes remain VALID. See in-file retirement addendum.

    Superseded by

    2026-07-01_remediation_audit.md (Finding 3)

  41. lane 08 cross lane synthesisValid recordVALID

    What the record says

    Lane 08 cross-lane synthesis (composed isolated mechanisms); no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  42. lane orchestration session summaryValid recordVALID

    What the record says

    Session summary for lane orchestration; no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  43. order shuffle stressorValid recordVALID

    What the record says

    Order-shuffle stressor (ingestion pipeline); no reranker, leakage, or gameable metric

    Superseded by

    Nothing. This remains the current ledger classification.

  44. curriculum experiments aggregateInvalidated or supersededPARTIAL

    What the record says

    "Scope-slot reranker (post-aggregate)" and "Verification rerun" sections are INVALIDATED (broken reranker). Pre-reranker aggregate (7/7) and post-fix update sections are VALID.

    Superseded by

    2026-06-30_scope_slot_reranker_fix.md

  45. knowledge core expansion 1mValid recordVALID

    What the record says

    Pre-reranker honest scope=0.0 measurement; post-fix update sections are post-reranker-fix and valid. No leakage-era Stage-2/5 claims.

    Superseded by

    Nothing. This remains the current ledger classification.

  46. reranker ab comparisonInvalidated or supersededPARTIAL

    What the record says

    Original A/B comparison body (lines 1–117) is INVALIDATED (run on broken reranker). "Update: bug fix changes conclusions" section is VALID.

    Superseded by

    2026-06-30_scope_slot_reranker_fix.md

  47. cortex dim benchmarkValid recordVALID

    What the record says

    Cortex dimension benchmark run after reranker fix; no leakage-era claims. Bilinear policy head analysis is valid.

    Superseded by

    Nothing. This remains the current ledger classification.

  48. residual to identity wiringValid recordVALID

    What the record says

    Residual-to-identity wiring report documents both pre- and post-reranker-fix; update sections correctly identify the fix. No leakage-era claims.

    Superseded by

    Nothing. This remains the current ledger classification.

  49. scope slot reranker fixInvalidated or supersededPARTIAL

    What the record says

    Bug diagnosis and fix description are VALID. "Curriculum Impact" table is SUPERSEDED (leakage-era 1.00 claims).

    Superseded by

    2026-07-01_honest_post_leakage_baseline.md

  50. honest post leakage baselineValid recordVALID

    What the record says

    Current reference point. Post-leakage-removal honest baseline. These are the numbers all future work must beat.

    Superseded by

    Nothing. This remains the current ledger classification.

  51. remediation auditValid recordVALID

    What the record says

    Full-repo experiment audit driving SPRINT.md. Meta-document; classifies the three invalidation events and the five strategic findings.

    Superseded by

    Nothing. This remains the current ledger classification.

  52. s1 3 hf kv slot injectionRefuted or retractedVALID

    What the record says

    S1.3 HF-substrate KV-slot fact injection — REFUTE on absolute recall (rank-1 on 1/3 facts); KV-splice ≡ text-prefix parity found. New experiment, leak-free.

    Superseded by

    Nothing. This remains the current ledger classification.

  53. s1 4 hf layer probeRefuted or retractedVALID

    What the record says

    S1.4 HF layer-L probe — REFUTE on Qwen-0.5B (gap −0.083) and LFM2.5 (+0.058 < +0.10). Confirms lane_03; retires Goal 2's mid-layer assumption. New experiment, pre-registered.

    Superseded by

    Nothing. This remains the current ledger classification.

  54. s2 4 breakthrough ablationRefuted or retractedVALID

    What the record says

    S2.4 single-variable ablation of the "13.5x breakthrough" (044cb51) — RETRACTED as magnitude inflation. Survival ratio 0.354 < 0.5; control logits rose more than target. New experiment, pre-registered.

    Superseded by

    Nothing. This remains the current ledger classification.

  55. stage5 scope dsi benchmarksInvalidated or supersededPARTIAL

    What the record says

    Stage 5 scope=1.00 and retention=1.00 claims are SUPERSEDED (leakage-era _SCOPE_TEACHING). DSI Fact Index implementation, external benchmark integration, and papers analysis are VALID.

    Superseded by

    2026-07-01_honest_post_leakage_baseline.md

  56. s1 1 model selectionValid recordVALID

    What the record says

    S1.1 HF substrate model selection — Qwen2.5-0.5B-Instruct (82.8 ms/tok). Decision record, not a claim under any invalidation event.

    Superseded by

    Nothing. This remains the current ledger classification.

  57. s2 1 minimal loopRefuted or retractedVALID

    What the record says

    S2.1 minimal metabolism loop — REFUTE H-LOOP. loop_delta_holdout=0.0000 (5 seeds, 3 holdout probes post-repair), loop_compounding_rho=nan. Validity gate passed (vanilla 0.0 < 0.5). Post-reranker-fix, post-leakage-removal, pre-registered. Mechanism: prefix budget eviction at K=8.

    Superseded by

    Nothing. This remains the current ledger classification.

  58. s2 2 kv content pathBlocked or failedVALID

    What the record says

    S2.2 KV content channel — BLOCKED (degenerate 0-probe holdout + S2.1 REFUTE gate binds). C1/C2 all 0.0000 due to 0 holdout probes. Addendum adjudicates as BLOCKED, not REFUTE. Implementation (minimal_loop_kv.py) merged and valid.

    Superseded by

    Nothing. This remains the current ledger classification.

  59. s2 5 forgetting testBlocked or failedVALID

    What the record says

    S2.5 forgetting test — BLOCKED (0 holdout probes, validity gate failed all 5 seeds). Not a refutation of H-FORGET. Addendum confirms BLOCKED via S2.1 REFUTE gate. Deletion APIs + 2×2 harness merged.

    Superseded by

    Nothing. This remains the current ledger classification.

  60. s3 m1 subtractive ablationValid recordVALID

    What the record says

    S3.M1 subtractive organ ablation (real GGUF driver, dev split, 3 seeds). ScopeSlotReranker +0.0465 all-stage (largest single-organ effect); DSI net-harmful (−0.060). No reranker, leakage, or gameable metric.

    Superseded by

    Nothing. This remains the current ledger classification.

  61. s3 m2 retrieval ablationPositive evidenceVALID

    What the record says

    S3.M2 additive retrieval ablation (HF driver, eval v2 holdout, 3 seeds). Scope-slot reranker zero-variance positive (S0 +0.667, S4 +0.250); hippocampus-at-answer Δ=0.000 exactly; DSI unsupported. Post-reranker-fix, post-leakage-removal.

    Superseded by

    Nothing. This remains the current ledger classification.

  62. s3 organ triage adjudicationValid recordVALID

    What the record says

    S3 organ triage adjudication — combines M1+M2 into KEEP/RETRIEVAL-BASELINE/ARCHIVE verdicts. ScopeSlotReranker=RETRIEVAL-BASELINE; all other organs=ARCHIVE. research/15 declared VACUOUS. No reranker, leakage, or gameable metric.

    Superseded by

    Nothing. This remains the current ledger classification.

  63. eval v2 1 expansionValid recordVALID

    What the record says

    Eval v2→v2.1 curriculum expansion (S0.6 growth path). +12 new ambiguous words across stages 0/1/2; stage-1 holdout 1→9 probes. Existing episodes/probes never modified. Regression locks updated. No reranker, leakage, or gameable metric.

    Superseded by

    Nothing. This remains the current ledger classification.

  64. eval v2 2 protocol repairValid recordVALID

    What the record says

    Human-approved protocol repair: Stage 1 probe-only, Stage 3 episode-interleaved, Stage 4 consolidate-before-post-test, consistent semantic scoring, category-stratified v2.2 split; legacy salt="v2" preserved. New baseline pending.

    Superseded by

    Nothing. This remains the current ledger classification.

  65. campaign 0d48130Blocked or failedVALID

    What the record says

    Campaign 0d48130 curated evidence log. 10 experiment outcomes across 3 commits and 2 providers (kaggle CPU-only, colab). Scientific outcomes: 2 POSITIVE (Exp04, Exp06), 1 POSITIVE+NULL (Exp07), 3 NULL (Exp01, Exp05, R14 M2B metricless), 1 REFUTATION (Exp02), 2 BLOCKED at teacher validity gate / diagnostic only (R18 gate, R18 full), 1 INFRASTRUCTURE BLOCKED (Exp03, original campaign). Exp03 reproducibility closure appended 2026-07-11 (commit ad77e93): real-driver rerun exit 0, layer_l_silhouette_gap=0.10925446726657728 (> +0.10, threshold unchanged) → positive/accept for this single closure; S1.4 not reopened. See the log and 2026-07-11_exp03_real_driver_closure.json.

    Superseded by

    Nothing. This remains the current ledger classification.

  66. exp03 real driver closureValid recordVALID

    What the record says

    Exp03 real-driver reproducibility closure. Durable execution report object: commit ad77e93, --driver real, Colab, exit 0, layer_l_silhouette_gap=0.10925446726657728 (> +0.10 registered threshold, unchanged), all ASI scores, model provenance (LiquidAI/LFM2.5-1.2B-Instruct rev 868df74d…, manifest infrastructure/kaggle/model_manifests/lfm2_5-1_2b-instruct.json), infrastructure fix description. Single run on one architecture; does not reopen S1.4.

    Superseded by

    Nothing. This remains the current ledger classification.

  67. live runner queueBlocked or failedVALID

    What the record says

    Live runner queue launch provenance and completion record. Durable record of the live experiment queue at implementation commit 5b5e93c63d769fea7854073a4e6c359e5d36606f. Records UTC launch date, live local state paths under /tmp/oczy-live-queue/ (batch, state, campaign, campaign_manifest — explicitly labeled as non-tracked live local state), scheduler flags (--watch-batch --watch-interval 30), additive provider capacity contract (10 Kaggle hard-cap + AIMD-learned Colab X, no global cap), source dataset/archive provenance (abdellahkadem/oczy-source-5b5e93c63d76, sha256 bc1ff926…), and the first job r18-distillation-5seed-diagnostic (Kaggle, kernel abdellahkadem/oczy-r18-5seed-5b5e93c63d76, module oczy.experiments.consolidation_distillation, args --seeds 5 --max-steps 10 --stage stage_0_grounding). Job completed (exit 0, state=succeeded, completed 2026-07-11T15:04:52Z, collected 2026-07-11T15:04:54Z). Scientific classification: BLOCKED at teacher validity gate (teacher_dev_delta=0.17647058823529413 < 0.2, identical across all 5 seeds). No positive scientific verdict claimed. Full adjudication in 2026-07-11_r18_five_seed_diagnostic.json.

    Superseded by

    Nothing. This remains the current ledger classification.

  68. r18 five seed diagnosticBlocked or failedVALID

    What the record says

    R18 5-seed diagnostic adjudication. Durable execution/adjudication JSON: commit 5b5e93c63d769fea7854073a4e6c359e5d36606f, Kaggle CPU, kernel abdellahkadem/oczy-r18-5seed-5b5e93c63d76, exit 0, 5 seeds. Per-seed distill_delta_holdout {0.3333, 0.3333, 0.0, 0.3333, 0.3333} (4/5 positive, seed 2 null); teacher_dev_delta=0.17647058823529413 identical across all seeds. Mean distill_delta_holdout=0.2667, mean specificity_delta=0.0261. Gate comparison: 0.1765 < 0.2 → FAILED. Scientific classification: BLOCKED at teacher validity gate / diagnostic only. No H-DISTILL verdict permitted (teacher gate failed after registered fallback). 4/5 conditional signal and seed-2 null both visible. No threshold, metric, or research spec changed.

    Superseded by

    Nothing. This remains the current ledger classification.

  69. r18 mechanism diagnosticsBlocked or failedVALID

    What the record says

    R18 mechanism diagnostics adjudication. Durable execution/adjudication JSON: commit 33169cc0340bf752a67adf63721ec64cb5f3c9f8, Kaggle CPU. Three diagnostic jobs (teacher ceiling, prompt-contract, training trajectory), all exit 0. Teacher ceiling (n=17): vanilla=0, raw_prefix=0.17647058823529413, chat_template=0; neither reaches 0.2 gate; registered chat fallback (0) worse than raw_prefix (0.1765). Prompt-contract audit: all six defect counts (issue/malformed/missing/truncated/answer-leak/mismatch) = 0; teacher_correct_rate=0.17647058823529413; raw/chat prompt accuracies 0; no structural prompt defect. Training trajectory: first submission failed HTTP 400 (long slug, preserved); short-slug retry exit 0 after ~12798s (run of record, preserved). Train loss 0.70→0.16, mean slope -0.0615, second-half -0.0190; underfit=1, instability=1, saturation=0, max final-loss divergence 0.01259. Final DEV student accuracies seeds 0–4: {0.117647, 0, 0, 0, 0.117647}; teacher 0.17647; seed 2 not uniquely divergent (seeds 1, 3 also 0). Adjudication decomposes failed gate into three axes: (1) prompt integrity — NO DEFECT; (2) capability ceiling — teacher expressivity/prompt-task ceiling IS THE BLOCKER; (3) optimization dynamics — token loss fits but DEV behavior unstable/weak, not saturated. Classification: BLOCKED at teacher validity gate / diagnostic only. No H-DISTILL verdict permitted. No threshold/spec/eval changes. All nulls visible.

    Superseded by

    Nothing. This remains the current ledger classification.

  70. r19 dev calibrationBlocked or failedVALID

    What the record says

    R19 DEV calibration adjudication. Durable execution/adjudication JSON: commit bd1ead9a8358b675af5e929c53a01eb505839639, Kaggle CPU. calibrate-dev v4 exit 0, all metrics collected. Manifest SHA-256 77ef4607…, parameter_total 60,388/64,000. DEV articulation gate FAILED (Arm B latent-control DEV accuracy ≤ C1 random-cortex DEV accuracy); oracle ceiling 0.357143 > 0 (PASSED independently). No signoff requested; no holdout accessed. Three prior infrastructure-failed attempts (v1 offline model resolution, v2 source-path/provenance + feature explosion, v3 artifacts not rooted in /kaggle/working); v4 infrastructure-successful but scientifically BLOCKED. C7 adapter discrepancy: manifest c7_available=true but _try_s3m2a_retrieval_adapter() returns None. R20 remains separately blocked on human signoff. No H-LATENT or H-LABEL verdict permitted.

    Superseded by

    Nothing. This remains the current ledger classification.

  71. r20 dev smokeBlocked or failedVALID

    What the record says

    R20 DEV implementation/smoke adjudication. Durable execution/adjudication JSON: commit e26d8291879d078b701f19802f72041e08cfd6a6, Kaggle CPU, kernel abdellahkadem/oczy-r20-dev-v3-e26d8291879d, exit 0, audit_status ok. Infrastructure/mechanism smoke only — no scientific verdict. Three attempts: v1 failed (offline loader failure), v2 failed (inference-tensor/autograd failure), v3 succeeded after fixes. Audit invariants: frozen organ hash identical before/after d8a3a3b…, checkpoint theta hash 8d6c41c5…, trace count 0 after deletion, online optimizer counts unchanged. 207,364 theta params / 829,456 bytes, F/S 64×64, bank 3×896, optimizer steps 1, best DEV validation score 0.0. Causal DEV deltas: trained-vs-update 0, untrained 0, shuffled 0, zeroed 0, swapped 0.0666667 — recorded as observed mechanism smoke. Test suites: focused 262 passed/2 skipped, organ 54 passed/2 skipped. Meta-test remains BLOCKED: no frozen meta_cortex/v1 instrument, distribution checks, power analysis, manifest, or human signoff exists. No ACCEPT/REFUTE verdict permitted. No threshold, metric, baseline, episode, scoring, eval manifest, or research spec changed. No holdout accessed; no signoff requested.

    Superseded by

    Nothing. This remains the current ledger classification.

  72. r18 implementationValid recordVALID

    What the record says

    R18 consolidation-as-distillation implementation and first runs. Initial implementation of consolidation_distillation.py, autoresearch segment 10 wiring (runs #200–#202). Run #202: Qwen2.5-0.5B + LoRA rank 2, ~220s, teacher_dev_delta ~0.176 below 0.2 validity gate — teacher gate FAILED, no H-DISTILL verdict. Concurrent: Numba CPU kernel acceleration (62ab18e), Kaggle research compute workflow (6dee16b), INT8 rescheduling planning. Gate failure confirmed by later mechanism/five-seed diagnostics (2026-07-11).

    Superseded by

    Nothing. This remains the current ledger classification.

  73. campaign 959e114Invalidated or supersededPARTIAL

    What the record says

    R20 INT8 meta_cortex/v2 DEV training and calibration campaign. Training/checkpoint, transport, runtime, and failure-timing evidence remains valid. Scientific aggregation of the v5 calibration shards is invalidated: source 949871b… chose C6 donors from shard-local task membership, making state_addressing_delta partition-dependent, and width-1 shards omitted required C6 entirely. The prior claim that width changes were control-plane-only is withdrawn. Human decision: version fix, let five active legacy jobs finish, then full width-3 DEV rerun. Commit a8c98d638209a8425b14a0f853e9fc46ae7da581 selects canonical next-within-family donors independent of shard membership; 89 focused tests passed. Corrected source dataset abdellahkadem/oczy-source-a8c98d638209, archive 40708cb9…. The v6 campaign has 150 private internet-off CPU jobs covering 5 × 90 DEV cells exactly once at width 3; its canary is queued behind the legacy drain and an automated gate verifies provenance, runtime, hashes, record counts, and C1–C6 before promotion. No scientific verdict; meta-test remains blocked/unaccessed.

    Superseded by

    Nothing. This remains the current ledger classification.