experiment log

S3 — Organ triage adjudication (research/14 verdicts)

File
2026-07-03_s3_organ_triage_adjudication.md
Size
3.9 KB
SHA-256
9b49b7a9efee973e…
Primary source. This is the verbatim Oczy document. The analytical field notes on the research page interpret and summarize these sources.

S3 — Organ triage adjudication (research/14 verdicts)

Date: 2026-07-03 Pre-registered spec: research/14-s3-organ-ablation-matrix.md; the KEEP / RETRIEVAL-BASELINE / ARCHIVE rule is applied mechanically below. Inputs:

  • M1 subtractive (full OrganismAgent − organ, real GGUF driver, dev split, 3 seeds): 2026-07-02_s3_m1_subtractive_ablation.md
  • M2 additive (MinimalOrganism + component, HF driver, holdout, eval v2, 3 seeds): 2026-07-02_s3_m2_retrieval_ablation.md
  • Code audit predictions recorded in the spec before either run.

Verdicts (spec rule: KEEP requires M2 Δ>0 with CI excluding 0 on ≥1 stage AND all-stage Δ≥0; retrieval paths that meet it are RETRIEVAL-BASELINE; everything else ARCHIVE)

Component M2 evidence M1 evidence Verdict
Scope-slot reranker S0 +0.667 ±0.00, S4 +0.250 ±0.00 (zero seed variance), all-stage +0.283 S2 +0.205 (largest single-organ effect) RETRIEVAL-BASELINE — kept, labeled retrieval in every future table, never counted as metabolism
Hippocampus at answer time Δ = 0.0000 on every stage and seed (bit-identical to BASE) all-stage +0.007 ±0.036 (noise) ARCHIVE (the answer-time retrieval path only; the hippocampus as consolidation-time replay buffer is part of the minimal organism and not under test here)
DSI fact index no stage with CI excluding 0 (S1 +0.667 ±1.43 on v2's single stage-1 probe — unsupported) net harmful in full stack (−0.060; removing it improves S2/S3/S4) ARCHIVE, with a named appeal: re-test S1 transfer on eval v2.1's 40-probe battery under a new pre-registered spec
WorldModelCritic not run (M2b run stopped; harness merged) −0.001 ±0.046 (noise) ARCHIVE (cannot be KEEP without M2 positive; M1 noise + audit: untrained MLP)
IdentityHypernetwork not run −0.012 ±0.021 (noise/harmful) ARCHIVE
SkillImmuneCortex not run +0.001 ±0.033 (noise) ARCHIVE (audit: keyword matcher, no learned params)
ExperienceAutoencoder not run −0.014 ±0.023 (noise/harmful) ARCHIVE (audit: no decoder exists)

Per the spec: "No middle category. A component that 'almost' helps is archived; it can return by winning a future pre-registered experiment." The M2b harness (organ_additive_organs.py, merged, 6 tests) is the standing appeal instrument for the four organs.

Deviations

  • M2b (additive arm for the four audited organs) was not executed: two agent attempts died silently and the orchestrator's run was stopped by the operator. Their ARCHIVE verdicts therefore rest on M1 (all CIs include 0) plus the spec's KEEP precondition (M2 positive required), which no unexecuted run can satisfy. This is the weakest link in the adjudication and is recorded as such; running M2b would only be needed to promote an organ, and every prediction and M1 datum says none would be.
  • M1 ran on the dev split at 3 seeds; M2 on eval v2 (pre-expansion) at 3 seeds. Verdict-relevant CIs are split-relative, so the rule applies unchanged.

Consequences

  • research/15 (S3.3 tensor wiring): VACUOUS — no component earned KEEP. Per that spec, this outcome "closes Goal 3's question honestly": there is no organ output worth wiring to tensors; the retrieval baseline stays a reranker by honest label.
  • S3.4 (archive under attic/ with post-mortems): now unblocked for critic, identity, immune, autoencoder, DSI, and the answer-time hippocampal path. Code task, pending.
  • Exploratory observation for the plasticity bets (18/19): BASE itself lifts stage 4 (0.250 ±0.00 vs vanilla 0) and stage 5 (0.556 vs 0.333) — the content channel is not uniformly dead outside stage 0; and the reranker's zero-variance wins mark the exact bar research/19's trained head must clear on transfer, where exemplar rerank structurally cannot generalize.