10 — HF-substrate layer-L hidden extraction probe (Sprint 1 / S1.4)

Pre-registered 2026-07-01 (human-approved sprint setup, before implementation). Agents running this experiment MUST NOT edit this spec; deviations are reported as deviations. This re-adjudicates lane_03's refuted H1 on a substrate that can actually see every layer (output_hidden_states=True), with the fallback analyses fixed in advance — lane_03's post-hoc pooling exploration is exactly what this pre-registration prevents.

Hypothesis

H-L: at some mid-depth layer L (25–75% of depth), sense-corrected phrase hiddens cluster by concept better than at the final layer, by silhouette score, gap >= +0.10.

Corpus

The lane_03 phrase corpus (same concepts × paraphrases; reuse its definition verbatim from lanes/lane_03.py). No new phrases may be added after seeing results.

Primary analysis (the ONLY acceptance surface)

• Embedding per phrase per layer: mean-pool over content tokens (stopword-token positions excluded, matching lane_03's mean-pool variant).
• silhouette(L) per layer; gap = max over mid layers silhouette(L) − silhouette(final layer).
• Accept H-L: gap >= +0.10.
• Refute: gap < +0.10. If the HF substrate confirms llama.cpp's refutation, that is a strong, clean negative: the cortex should consume final-layer hiddens and Goal 2's "mid-layer semantic intent" assumption is retired.

Pre-registered secondary analyses (exploratory only — cannot flip acceptance)

1. Last-token pooling per layer (lane_03's post-hoc variant, now registered).
2. Max-pooling per layer.
3. Per-layer table for the chosen model AND (if cached weights permit) the LFM2.5-1.2B HF checkpoint, to separate "substrate keyhole" from "model property".

Reporting

Per-layer silhouette table (all layers, all three poolings, clearly marking the primary), model id(s), corpus hash, and the accept/refute verdict against THIS spec. Log to experiments_logs/.