Summary

Follow-up to PR #249 (Roy-5a analyzer, merged). The user asked me to do the recommended "re-run Roy-5a once or twice to confirm the text-modality silence is stable". Three follow-up runs traced the root cause to a missing env var — not variance, not a regression.

What the investigation found

Run 1 — Roy-5a-confirm (identical spec, no env changes). 0 text-modality EC fires — silence reproduces, not run-to-run variance.

Investigation between runs: traced the gate at integration/memory_hub.py:247:

• MemoryHub._encoder is only wired when MAXIM_SUBSTRATE_PATH=1.
• BioEnrichmentPipeline receives encoder=getattr(memory_hub, "_encoder", None) at runtime/bio_stack.py:378.
• Its text-modality EC fire at bio_enrichment.py:576 short-circuits on encoder=None.

Roy-4's 154 text fires came from BioEnrichmentPipeline queries — requiring MAXIM_SUBSTRATE_PATH=1 to have been inherited from the shell. Roy-5a + Roy-5a-confirm ran from a shell without it; Roy-4 had it.

Run 2 — Roy-5a-substrate-on (same spec, env var explicitly set). 162 text-modality EC fires, 34 NEW text centroids — within run-to-run variance of Roy-4's 154/57. The substrate path was the gate, not a code regression.

The structurally clean H1a verdict

Analyzer re-run on Roy-5a-substrate-on data:

Even with the substrate path active, ZERO of the 14 text-modality centroids are food-bearing. NAc's cluster_reward_bias map keys food clusters exclusively to interoception node IDs. H1a still triggers — same verdict, stronger mechanism.

Bonus structural finding (the dimension-mismatch warning earned its keep)

The dimension-mismatch warning added in PR #249's pre-merge review fold fired on real data: row dims=[384] col dims=[768].

• interoception (SensorEncoder SHA-basis): 384-dim
• text (LinguisticEncoder paraphrase-mpnet-base-v2): 768-dim

Cross-modality cosine (M_dt) is mathematically undefined — different dimensional spaces. _cosine returns 0.0 silently on length mismatch. This is stronger than the plan's H1a framing ("encoder subspaces are far in cosine space"). The real finding: the subspaces are different-dimensional, cosine is undefined, and any cross-modal cosine-based alignment is structurally impossible without a learned projection layer.

The Hebbian binding mechanism cancelled by Roy-4 was always structurally impossible at the cosine level — cross_modal_substrate_binding.md's Stage 4a resurrection conditions would need a learned dimension-reducing projection between encoders, not a Hebbian rule on raw cosine.

Sharpened Stage 3 pass criterion

The original plan said "produce co-firing + Hebbian binding succeeds". The dimension-mismatch finding shows the second clause is structurally impossible at cosine level. The refined criterion is:

At least one text-modality EC node appears in NAc's cluster_reward_bias map keyed to sense_food_source after Stage 3 priming.

Directly observable in aut_nac.json + aut_ec.json — no cosine math required. If Stage 3's narrator-utterance scaffold produces a text-modality food centroid, M_tt becomes non-empty in Roy-5b for the first time across all Roy iterations and H1c / H1b becomes discriminable.

What changed in the repo

• scenarios/roy/roy_5a_confirm_iteration.yaml — variance-check spec.
• scenarios/roy/roy_5a_substrate_on_iteration.yaml — substrate-path-on spec.
• docs/experiments/22_roy_5a.md — outcome doc updated with the three-run findings + dimension-mismatch section + sharpened Stage 3 criterion.
• docs/experiments/protocols/22_roy_5a_reproduction.md — MAXIM_SUBSTRATE_PATH=1 documented as load-bearing in the runner env.
• docs/plans/persona_convergence_crucible.md — iteration log entry updated.

No analyzer code changes needed — the dimension-mismatch warning from PR #249's pre-merge fold did its job and surfaced the architectural finding without changes here.

Test plan

• python -m pytest tests/unit/test_roy_5_cosine_localization.py tests/unit/test_save_aut_state.py -q — 37 passed.
• ruff check clean on touched Python files (the only Python files touched here are the existing scenarios — no Python source changes in this PR).
• Analyzer end-to-end against real Roy-5a-substrate-on session dirs returns rc=0 + correct H1a verdict + dimension-mismatch warning visible in stderr.

Out of scope (deferred)

• Stage 3 implementation (cradle-arc redesign with narrator co-firing). Verdict gates it; this PR doesn't ship it.
• Strategic question for 1.0 / 1.1+ scoping: does the dimension-mismatch finding require a follow-up plan to either (a) align SensorEncoder + LinguisticEncoder on a common embedding dim, or (b) document the learned-projection requirement as 1.2+ research? Belongs in a separate planning thread.

🤖 Generated with Claude Code