experiment log
Organ Ablation Matrix — 2026-07-02_s3_m1_subtractive_ablation
- File
2026-07-02_s3_m1_subtractive_ablation.md- Size
- 5.1 KB
- SHA-256
36a1fac3311fbabd…
Primary source. This is the verbatim Oczy document. The analytical field notes on the research page interpret and summarize these sources.
Organ Ablation Matrix — 2026-07-02_s3_m1_subtractive_ablation
Path: real-driver (LlamaCVecDriver GGUF).
Generated: 2026-07-02 22:22:17 UTC Seeds: 3 | Split: dev | Backend: real GGUF + CortexAgent
Accuracy Matrix
| Config | Stage 0: Sense grounding | Stage 1: Transfer within domain | Stage 2: Scope control | Stage 3: Dialog | Stage 4: Consolidation stress | Stage 5: Cross-domain disambiguation |
|---|---|---|---|---|---|---|
| FULL | 0.8000 ± 0.0000 | 0.8095 ± 0.2049 | 0.7179 ± 0.1103 | 0.2000 ± 0.0000 | 0.8333 ± 0.0000 | 0.8519 ± 0.1594 |
| MINIMAL | 0.4000 ± 0.0000 | 0.9048 ± 0.2049 | 0.6410 ± 0.1103 | 0.2000 ± 0.0000 | 0.5556 ± 0.2391 | 0.5926 ± 0.4217 |
| FULL-hippocampus | 0.8000 ± 0.0000 | 0.8571 ± 0.0000 | 0.6667 ± 0.1103 | 0.2000 ± 0.0000 | 0.8333 ± 0.0000 | 0.8148 ± 0.1594 |
| FULL-critic | 0.8000 ± 0.0000 | 0.8571 ± 0.0000 | 0.6410 ± 0.1103 | 0.2000 ± 0.0000 | 0.8333 ± 0.0000 | 0.8889 ± 0.0000 |
| FULL-identity | 0.8000 ± 0.0000 | 0.8571 ± 0.0000 | 0.7436 ± 0.2207 | 0.2000 ± 0.0000 | 0.8333 ± 0.0000 | 0.8519 ± 0.1594 |
| FULL-immune | 0.8000 ± 0.0000 | 0.8571 ± 0.0000 | 0.6667 ± 0.1103 | 0.2000 ± 0.0000 | 0.8333 ± 0.0000 | 0.8519 ± 0.1594 |
| FULL-autoencoder | 0.8000 ± 0.0000 | 0.8571 ± 0.0000 | 0.7179 ± 0.1103 | 0.2000 ± 0.0000 | 0.8333 ± 0.0000 | 0.8889 ± 0.0000 |
| FULL-dsi | 0.8000 ± 0.0000 | 0.8571 ± 0.0000 | 0.8462 ± 0.0000 | 0.4000 ± 0.0000 | 1.0000 ± 0.0000 | 0.6667 ± 0.0000 |
| FULL-scope_slot_reranker | 0.8000 ± 0.0000 | 0.8095 ± 0.2049 | 0.5128 ± 0.2207 | 0.2000 ± 0.0000 | 0.8333 ± 0.0000 | 0.7778 ± 0.0000 |
Per-Organ Δ (FULL minus FULL-organ)
Positive Δ → organ contributes; negative → ablation improves accuracy.
| Organ | Stage 0: Sense grounding | Stage 1: Transfer within domain | Stage 2: Scope control | Stage 3: Dialog | Stage 4: Consolidation stress | Stage 5: Cross-domain disambiguation | All-stage mean ± CI |
|---|---|---|---|---|---|---|---|
| NeuralHippocampus | +0.0000 | -0.0476 | +0.0513 | +0.0000 | +0.0000 | +0.0370 | +0.0068 ± 0.0363 |
| WorldModelCritic | +0.0000 | -0.0476 | +0.0769 | +0.0000 | +0.0000 | -0.0370 | -0.0013 ± 0.0459 |
| IdentityHypernetwork | +0.0000 | -0.0476 | -0.0256 | +0.0000 | +0.0000 | +0.0000 | -0.0122 ± 0.0212 |
| SkillImmuneCortex | +0.0000 | -0.0476 | +0.0513 | +0.0000 | +0.0000 | +0.0000 | +0.0006 ± 0.0328 |
| ExperienceAutoencoder | +0.0000 | -0.0476 | +0.0000 | +0.0000 | +0.0000 | -0.0370 | -0.0141 ± 0.0232 |
| DifferentiableFactIndex | +0.0000 | -0.0476 | -0.1282 | -0.2000 | -0.1667 | +0.1852 | -0.0596 ± 0.1481 |
| ScopeSlotReranker | +0.0000 | +0.0000 | +0.2051 | +0.0000 | +0.0000 | +0.0741 | +0.0465 ± 0.0873 |
Per-Organ Verdicts
Each verdict compares FULL-organ against FULL (real-driver (LlamaCVecDriver GGUF)).
- NeuralHippocampus (
hippocampus): contributing (organ saves +0.0068 on average) (per-stage Δ: Stage 0: Sense grounding: +0.0000, Stage 1: Transfer within domain: -0.0476, Stage 2: Scope control: +0.0513, Stage 3: Dialog: +0.0000, Stage 4: Consolidation stress: +0.0000, Stage 5: Cross-domain disambiguation: +0.0370) - WorldModelCritic (
critic): dead weight (delta indistinguishable from noise) (per-stage Δ: Stage 0: Sense grounding: +0.0000, Stage 1: Transfer within domain: -0.0476, Stage 2: Scope control: +0.0769, Stage 3: Dialog: +0.0000, Stage 4: Consolidation stress: +0.0000, Stage 5: Cross-domain disambiguation: -0.0370) - IdentityHypernetwork (
identity): harmful (ablation improves accuracy by +0.0122 on average) (per-stage Δ: Stage 0: Sense grounding: +0.0000, Stage 1: Transfer within domain: -0.0476, Stage 2: Scope control: -0.0256, Stage 3: Dialog: +0.0000, Stage 4: Consolidation stress: +0.0000, Stage 5: Cross-domain disambiguation: +0.0000) - SkillImmuneCortex (
immune): dead weight (delta indistinguishable from noise) (per-stage Δ: Stage 0: Sense grounding: +0.0000, Stage 1: Transfer within domain: -0.0476, Stage 2: Scope control: +0.0513, Stage 3: Dialog: +0.0000, Stage 4: Consolidation stress: +0.0000, Stage 5: Cross-domain disambiguation: +0.0000) - ExperienceAutoencoder (
autoencoder): harmful (ablation improves accuracy by +0.0141 on average) (per-stage Δ: Stage 0: Sense grounding: +0.0000, Stage 1: Transfer within domain: -0.0476, Stage 2: Scope control: +0.0000, Stage 3: Dialog: +0.0000, Stage 4: Consolidation stress: +0.0000, Stage 5: Cross-domain disambiguation: -0.0370) - DifferentiableFactIndex (
dsi): harmful (ablation improves accuracy by +0.0596 on average) (per-stage Δ: Stage 0: Sense grounding: +0.0000, Stage 1: Transfer within domain: -0.0476, Stage 2: Scope control: -0.1282, Stage 3: Dialog: -0.2000, Stage 4: Consolidation stress: -0.1667, Stage 5: Cross-domain disambiguation: +0.1852) - ScopeSlotReranker (
scope_slot_reranker): contributing (organ saves +0.0465 on average) (per-stage Δ: Stage 0: Sense grounding: +0.0000, Stage 1: Transfer within domain: +0.0000, Stage 2: Scope control: +0.2051, Stage 3: Dialog: +0.0000, Stage 4: Consolidation stress: +0.0000, Stage 5: Cross-domain disambiguation: +0.0741)