Two comparators before the cortex

The distillation comparator treats consolidation as a weight edit inside the language model. A teacher sees transient experience, a small LoRA adapter is trained, raw traces are deleted, and the student is tested on held-out behavior and specificity.

This is not the final cortex architecture: the language organ’s own effective weights change. It is a useful comparator because it asks whether context can become persistent parametric behavior under a strict deletion and held-out protocol.

The initial prototype was engineering-real but scientifically inconclusive. It trained from episode.corrected_response, allowed the holdout signal to overlap the information used for training and tuning, revisited the holdout during development, ran only two seeds, and changed adapter scope along the way. Its reported 0.333 delta belongs in the history, but not in the evidence column.

The repair was procedural and mechanism-level: a DEV-only tuning firewall, one-shot holdout access, a pre-registered teacher-validity gate, fixed fallbacks, and at least five seeds. This is why the later result can be interpreted even though it remains blocked.

1. PrototypeINVALID / INCONCLUSIVE

   Tried

   LoRA context distillation with corrected responses.

   Observed

   Attractive holdout delta near 0.333.

   Why

   Training/holdout overlap, repeated tuning, two seeds, and changing adapter scope.

   Response

   Rebuilt the evidence firewall and froze the gate.

   What it means

   Implementation progress, not evidence for H-DISTILL.

2. Five-seed runBLOCKED

   Tried

   Run the repaired registered fallback on five seeds.

   Observed

   Four +0.3333 student deltas and one 0; teacher 0.17647 on every seed.

   Why

   The teacher missed the unchanged 0.2 validity gate.

   Response

   Preserved the conditional signal and refused to lower the gate.

   What it means

   Diagnostic student behavior; no accept or refute verdict.

3. DiagnosticsBLOCKER LOCALIZED

   Tried

   Audit prompts, teacher ceiling, fallback template, and training trajectory.

   Observed

   No prompt defects; chat fallback 0; falling loss but weak unstable DEV behavior.

   Why

   Teacher expressivity and task/prompt ceiling.

   Response

   Retired identical reruns and moved to the language-organ interface.

   What it means

   The failure now points to a specific capability ceiling.

On five seeds, four produced a holdout delta of 0.3333 and one produced 0.0, for a mean of 0.2667. Those values are infrastructure-confirmed. They are not admissible evidence for H-DISTILL because every seed measured the teacher DEV delta at 0.17647, below the pre-registered 0.2 gate.

Mechanism diagnostics found no structural prompt-contract defect. The chat-template fallback scored 0, worse than the raw prefix. Training loss fell, but DEV behavior was weak and unstable. Further identical runs were retired instead of lowering the threshold.

Teacher gate

0.1765 < 0.2Failed unchanged threshold on every seed.

Student holdout

mean 0.2667Four positive seeds, one null; diagnostic only.

Specificity

mean +0.0261Recorded, but cannot rescue a failed validity gate.

Verdict

BLOCKEDNo H-DISTILL accept or refute is permitted.

R19 introduced two matched direct-learning arms. One can learn a label-prefix mapping—a parametric retrieval comparator. The other must express query-conditioned cortex state through a latent coupler into a frozen language organ, without transmitting answer text.

The fourth Kaggle attempt completed and collected cleanly. The text oracle ceiling was 0.357143, so the frozen organ can express some of the DEV behavior when directly informed. But the learned latent path did not improve over its no-update baseline, so the DEV articulation gate remained false. No signed evaluation or holdout access followed.

Parameter budget

60,388 / 64,000The implemented diagnostic stayed inside its fixed budget.

Text oracle

0.357143Nonzero upper bound on DEV articulation.

Trace audit

0 tracesDeletion verified; holdout_accessed=false.

Latent gate

FAILEDNo signoff request and no H-LATENT/H-LABEL verdict.

1. v1INFRASTRUCTURE FAILURE

   Tried

   Load the frozen organ offline.

   Observed

   Model resolution failed before calibration.

   Why

   The loader addressed the hub instead of the attached local snapshot.

   Response

   Added explicit offline path resolution.

   What it means

   No scientific evidence.

2. v2INFRASTRUCTURE FAILURE

   Tried

   Run from commit-addressed source.

   Observed

   Provenance failed and frozen request features exploded.

   Why

   Source-path/hash handling plus unnormalized features.

   Response

   Fixed archive provenance, normalization, and nonfinite checks.

   What it means

   No scientific evidence.

3. v3COLLECTION FAILURE

   Tried

   Run the repaired DEV calibration.

   Observed

   Computation finished but artifacts were not collected.

   Why

   Outputs were outside /kaggle/working.

   Response

   Rooted artifacts in the sentinel-visible directory.

   What it means

   Uncollected metrics are not admissible.

4. v4SCIENTIFICALLY BLOCKED

   Tried

   Run the full matched DEV calibration.

   Observed

   Oracle ceiling passed; learned Arm B did not beat random C1.

   Why

   The latent coupler failed the articulation gate.

   Response

   Withheld signoff and holdout access; opened coupler and C7 diagnosis.

   What it means

   The organ can express text instructions, but the learned interface cannot yet control it.

R18 says the task-or-prompt teacher must be strong enough before student persistence is meaningful. R19 says storing a signal in trainable state is not enough if the frozen organ cannot decode the learned latent interface.

Together they motivate Research 20’s developmental outer loop: learn the writer and the articulation protocol across task families before asking either to generalize online to an unseen rule.