04-lit-audit-summary

SUMMARY — Route C: eight literature audits closed

Run date 2026-08-21 · executor EXECUTOR · order mandated: h09 → h02 → h05 → h12–h16.

The table

#	Hypothesis	Tier	lit_audit before → after	Blockers	One-line reason
h09	hydrogel	S	partial → not-modelable	4	All remaining gaps are unmeasured-anywhere, not un-retrieved; 9 searches over two campaigns confirm the WH2 × F-actin Kd does not exist.
h02	piezo	C	partial → partial	3	V_saturation is unmeasurable in principle (prestin NLC is Boltzmann), but OHC membrane resistance is retrievable and the 210 kPa phantom is still live in code.
h05	calcium-oscillation	C	partial → partial	6	Three STRC kinetics blockers are genuinely unmeasured; three are retrievable — and CaMKII k_auto, previously logged “not measured,” was located this run.
h12	sonogenetic	C	deferred → partial	1	Swept parameters correctly need no source; the one external anchor, “Wu et al. 2023” (62×), has no PMID/DOI and no file in sources/lit/.
h13	recombinases	C	deferred → fixed	0	Technology-watch stub: zero computational steps, zero model constants, scripts/ empty; the one checkable quantity (STRC CDS 5,325 bp) verifies.
h14	protein-replacement	C	deferred → partial	2	The hub’s kill (“no delivery route”) contradicts the note’s own “feasible in hours”; the RWM step is quantifiable from a Salt 2001 note the vault already owns.
h15	otoa-paralog	D	deferred → partial	2	The 13.78 Å kill number is sound in-house compute, but the RMSD threshold bands that turn it into a verdict are uncited, and TM-align was never run.
h16	zp-prion	D	deferred → partial	2	Genotype is sourced, but the inference to “wrong patient” is unrecorded and contradicted by the note’s own text; the ZP mechanism has no sources/lit/ citation.

deferred count: 5 → 0. Every non-fixed hub carries a non-empty lit_audit_blockers: list. Total blockers enumerated: 20.

What the states now mean, and why the split came out this way

The three-state scheme did real work: it forced apart two things that partial had been silently merging.

• not-modelable (h09 only) — the parameter does not exist in the primary literature and further searching cannot produce it. partial on h09 had been carrying an implicit promise that another retrieval cycle would close the WH2 × F-actin gap. After 9 distinct attempts across two independent campaigns four months apart, that promise is false and actively wasteful — it invites the next agent to re-search a settled absence. not-modelable redirects that effort at the wet lab, which is the only thing that can help.

• partial (six hubs) — blockers that further work will reduce. h05’s CaMKII blocker is now a PMID away; h02’s membrane resistance needs one PDF; h14’s kill needs one calculation from a paper already on disk; h15’s needs one citation and one cheap TM-align rerun.

• fixed (h13) — nothing to source, and that was verified rather than assumed.

Cross-cutting findings

1. One wet-lab measurement would close blockers in two hypotheses. h05’s “STRC copy number per OHC” and h09’s STRC_NORMAL_OHC_M = 1 µM are the same missing number, reached independently through two different Barr-Gillespie datasets (Wilmarth 2015 mouse vestibular; Krey 2019 chick utricle) and independently confirmed absent in both. A single sorted-OHC single-cell proteomics run closes both. Worth knowing when wet-lab budget is allocated.

2. Audits that fix the tables and not the code leave the fabrication running. h02’s “210 kPa per Masaki 2009 PLoS One 4:e4877” was identified as a phantom on 2026-04-23. parameters/tectorial-membrane.md was corrected on 2026-04-25. The phantom string is still at models/piezo_voltage_budget.py:262 today, with "E_TM_basal_kPa": 210 at line 271 — against a measured 24.3 kPa. A future run that greps the parameter tables sees ✅ and never opens the script. Recommend a vault-wide sweep: for every parameter row marked ❌ or corrected in parameters/, grep the scripts for the old value and confirm it was removed.

3. “Phantom citation destroyed” is not the same as “value does not exist.” h05’s CaMKII k_auto was logged “not measured — Chao 2011 is structural only.” The measurement exists: Bradshaw, Hudmon & Schulman 2002 JBC 277:20991–20998 (PMID 11925447), ~12 s⁻¹ at 30 °C by chemical quenched-flow. The prior pass searched the structural CaMKII literature after killing a bad citation and concluded the quantity was unmeasured. Killing a citation should trigger a fresh search on the quantity, not a conclusion about it.

4. Two kills do not rest on what their hubs say they rest on. h15 (uncited RMSD threshold bands + an unrun TM-align that the note itself calls cheap) and h16 (a “wrong patient” verdict contradicted by the note’s own genotype reasoning). Neither finding argues for revival. Both mean the D tier is currently asserted rather than shown — which is what deferred was concealing.

5. Sweeping beats sourcing for a genuinely unmeasured parameter. h12 handles Kd_CaN and NFAT_threshold by sweeping 245 combinations across 10–50× ranges, so its conclusion needs no literature point value for either. On this narrow point a C-tier speculative hypothesis is better disciplined than the S-tier one, which had to reach for the same technique retroactively in Phase 4n.

Honest limits of this run

• h09’s two carried absences (STRC per-OHC quantification; the HTC 19-link derivation) were cited from the 2026-04-25 audit, not re-searched. Both were established with specific enumerated evidence and neither is the kind of claim new literature silently overturns in four months. Stated so the not-modelable verdict is not read as four fresh confirmations.
• h02’s OHC membrane resistance got 2 attempts, not 5. It is recorded as open and retrievable, not as a confirmed absence. Calling it confirmed would have been the easy way to a cleaner-looking table and would have been false.
• h05’s CRE K_half was not attempted. Budget went to CaMKII, which was the higher-value correction. Recorded as such rather than quietly folded in.
• No sources/lit/ note was created for the Bradshaw 2002 lead, because the value came from a search-engine restatement and the PubMed verification fetch returned a cookie interstitial. §0c rule 4 is explicit; a search snippet is not a locally-parsed paper. The lead is recorded with full precision instead.
• Nothing was written to the vault. All eight hub edits, eight phase-proof notes and eight log lines are staged in runs/hypc/apply.sh for the chief.