research(srt): adjudicate the SRT interpretive-signal proposal as study-only

[aurascoper]

Docs-only adjudication of an externally-supplied proposal to adopt SRT (Semiotic-Reflexive Transformer) for dialectical fork detection, discourse-aware retrieval, and cross-modal provenance.

Tracks #49. Filing does not authorize work — no milestone identifier is assigned, nc-queue.yaml and every goal file are unamended, and m7-b remains the only active milestone.

Contents

• docs/research/srt-interpretive-signal-evaluation_v1.md — the adjudication record
• docs/scoping/srt-interpretive-signal-register-v0.json — 4 register entries, all validating against research-decision-register.schema.json, all study_only or deferred with runtime_dependency_authorized: false

Why this base branch

Based on feat/local-dialectic-review, not main. The research-decision-register apparatus (docs/scoping/research-decision-register.schema.json) arrives in 2a8869e and is absent from main — a register entry targeted at main would reference a schema that does not exist there.

Findings that changed the disposition

• SRT's headline TruthfulQA-MC2 AUCs are attributed to a separate frozen-backbone probe, explicitly not an adapter capability. The adapter's own novel channels are disclaimed by its authors as not separating hallucinated from faithful free-form output above chance.
• The proposal omits Semantic Entropy Probes — a linear, single-forward-pass baseline with public code. Without that comparator a positive SRT result cannot show SRT was needed, so the reproduction is reordered behind it.
• Hidden-state methods are documented to degrade under genuine ambiguity by conflating it with model confusion, so the safety question is whether a signal separates interpretive plurality from model ignorance. Existing turn records carry no labels and cannot answer it.
• The loop already computes post-generation candidate tension but not branch preservation. The genuine gap is timing, not semantics.

Also recorded: a blocking prerequisite

Embedding-space comparability fails open. Embedding.cosineSimilarity(to:) compares on dimension alone, so two distinct same-dimension spaces mix silently — contradicting the documented rule that vectors from different space identities never share one index. Live correctness defect independent of SRT; blocks the retrieval question.

Verification

• 4/4 register entries validate against the schema as it exists on the base branch
• No code: 2 files, both under docs/
• Built in a clean worktree from the base branch's committed state, so none of the working tree's in-flight changes are swept in
• nc-queue.yaml untouched (operator-owned, outside any git repo)

🤖 Generated with Claude Code