SPEC

SPEC — Route C: close the eight open literature audits

Goal (restated)

Eight STRC hypothesis hubs carry an unresolved lit_audit: state. Three are partial (h09, h02, h05), five are deferred (h12, h13, h14, h15, h16). deferred means “nobody ever looked” — this route deletes that state entirely.

For each of the eight I must produce an honest, enumerated parameter-provenance audit per STRC AGENTS.md §0c (literature-first discipline), and land each hub in exactly one of three terminal states:

state	meaning	requirement
fixed	every load-bearing constant maps to a primary source that EXISTS in sources/lit/ with canonical type: lit frontmatter, and the value is grep-verifiable in that file	enumeration of what was checked
partial	irreducible blockers remain	non-empty lit_audit_blockers: YAML list, one entry per blocker
not-modelable	the parameter genuinely is not measured anywhere in the primary literature	non-empty lit_audit_blockers: list with a per-item reason

ZERO hubs may remain deferred. Setting fixed without an enumeration is cheating.

Mandatory order

1. h09-hydrogel — S-tier + partial. The single most dangerous combination in the vault: a top-ranked hypothesis whose numbers were never fully sourced. Gets real effort. Known hard blocker: WH2 × F-actin Kd, previously recorded “CONFIRMED NOT MEASURED”. I must decide whether that is still true and, if so, whether the honest terminal state is not-modelable rather than an eternal partial.
2. h02-piezo, then h05-calcium-oscillation — both partial, both C-tier.
3. h12-sonogenetic, h13-recombinases, h14-protein-replacement, h15-otoa-paralog, h16-zp-prion — the five deferred. C/D-tier and speculative, so a proportionate audit is correct: for a D-tier killed hypothesis, confirming that the kill rests on sourced evidence is enough.

Method per hypothesis

1. Read hypotheses/hNN-*/index.md and log.md; read any existing audit note under hypotheses/hNN-*/phases/.
2. Build the provenance table: every numerical constant appearing in that hypothesis’s scripts (hypotheses/hNN-*/scripts/, models/) and in the load-bearing claims of the hub — value, units, claimed source, and whether that source exists locally in sources/lit/ with canonical type: lit frontmatter.
3. Retrieval for unsourced items: rag search, qmd query, WebSearch/WebFetch, and the paywalled-PDF skills. Any fetched PDF parsed with exactly mineru -p <in> -o <out> -b vlm-engine.
4. Persist every passage I rely on at retrieval time into GROUNDING/hypc-<hNN>-<slug>.md: verbatim passage, source + edition + pages, the exact query run, the index path, and honest evidential weight (STATES the claim vs a VARIANT; a variant carries its bridging derivation in the same file). Unpersisted evidence does not exist.
5. Where a constant is genuinely not measured, that IS the finding: record “CONFIRMED NOT MEASURED” plus the searches establishing it. Never fabricate a value from training data; never present a closest analog as if it were the measurement — analogs are flagged as analogs.

Deliverables (paths)

All under runs/hypc/ in this space; grounding under GROUNDING/.

• runs/hypc/SPEC.md — this file (first).
• runs/hypc/AUDIT-h09.md, -h02, -h05, -h12, -h13, -h14, -h15, -h16 — provenance table + verdict + blockers, one per hypothesis.
• runs/hypc/SUMMARY.md — one table: hypothesis | lit_audit before → after | blocker count | one-line reason.
• runs/hypc/RANKING-DELTA.md — for EACH of the eight, the exact one-line entry to PREPEND to hypotheses/hNN-*/log.md in the format - **{phase}**: {verdict}. {key-number}. -> [[proof note]], plus whether the audit outcome justifies a tier/mech/deliv/misha_fit change and why. Tier moves ARE authorized this run; the chief makes the final call, I recommend.
• runs/hypc/apply.sh — idempotent bash, set -euo pipefail, run BY THE CHIEF, making exactly my recommended edits under /Users/egorlyfar/Missions/STRC: updates lit_audit, lit_audit_date: 2026-08-21, lit_audit_blockers in the eight hub frontmatters; creates the per-hypothesis audit phase-proof notes under hypotheses/hNN-*/phases/ with 2026-08-21 in the filename (h09’s is required by contract) following _templates/fileclasses/ (type: phase-proof, date, tags incl. strc, excerpt ≤180 chars, ## Ranking delta, ## Connections with ≥1 wikilink); prepends the log lines. No wildcards over the whole vault. Never deletes a note.
• runs/hypc/verify.txt — bash -n runs/hypc/apply.sh output, plus for every “sourced” claim the grep command that matches the value inside its sources/lit/ file, and that grep’s output.
• GROUNDING/hypc-<hNN>-<slug>.md — one per hypothesis touched.

Write boundary

• WRITE: only runs/hypc/ and GROUNDING/ inside /Users/egorlyfar/Missions/STRC/labs/spaces/2026-08-21-refactor-rank-and-moderna.
• READ: the STRC vault freely.
• NEVER write anywhere in the vault. NEVER touch ~/Knowledge/Brain.
• The chief applies all vault edits by running my apply.sh.

Done-criteria checklist

• Eight AUDIT-hNN.md files exist, each with a provenance table and a verdict.
• No hub recommendation is deferred.
• Every non-fixed recommendation carries a non-empty lit_audit_blockers: list.
• Every fixed recommendation enumerates what was checked and each value is grep-verifiable in a sources/lit/ file (evidence in verify.txt).
• Every relied-upon passage is persisted in GROUNDING/.
• SUMMARY.md, RANKING-DELTA.md, apply.sh, verify.txt exist.
• bash -n runs/hypc/apply.sh is clean, captured in verify.txt.
• No file written outside runs/hypc/ and GROUNDING/.

Anti-goals

Unpersisted evidence; fixed without enumeration; a non-fixed hub with an empty lit_audit_blockers; fabricated parameters; an analog presented as a measurement; writing outside my two directories; touching Brain; leaving any hub at deferred.