verdict-B

Reviewer B verdict (first-principles lens)

Overall: ACCEPT-WITH-FIXES

The non-cell-autonomous question (does extracellular stereocilin create a rescue route the packet missed?)

Start from what is actually true about the molecule. Two independent web-verifiable facts and two vault-internal facts matter here:

1. Stereocilin has no transmembrane domain and is secreted extracellularly (confirmed by OMIM 606440 and the PNAS/J Comp Neurol structural papers on stereocilin’s horizontal-top-connector and tectorial-membrane-attachment role). It is detectable in serum as a circulating biomarker (PubMed 39033359), which at minimum proves the protein can leave a cell and travel through extracellular fluid at some scale.
2. PANEL-PACKET/02-moderna-assessment.md:68 states, without qualification: “There is no cross-correction, no bystander effect, no clonal expansion of hair cells… Every rescued cell must be individually reached.” This is presented as a settled mechanistic fact.
3. The vault’s own record disagrees with that certainty, and the packet does not cite or rebut it. hypotheses/h04-mrna-lnp/phases/STRC mRNA-LNP Strategy B Audiogram Rescue.md:124 states verbatim: “Unmodelled Strategy B advantage: non-cell-autonomous secretion. Stereocilin is a secreted extracellular protein — if it diffuses from transfected to adjacent untransfected OHCs (within 10 μm radius = immediate neighbors), the effective functional fraction doubles or triples.” The same file’s action item at line 145 proposes exactly the diffusion model this reviewer was asked to check for. hypotheses/h06-mrna-therapy/phases/STRC mRNA Therapy Hypothesis.md:85 lists as a stated advantage: “Extracellular protein: secreted, can act non-cell-autonomously (neighboring untransfected OHCs may benefit).”
4. This is not a hypothetical I am inventing — it is a claim already sitting, unresolved, in the two hubs the Moderna assessment is scoring against. §0c (literature-first) and ordinary diligence require engaging with existing vault content before asserting the opposite as fact. The packet did not do this.

However — and this is the Feynman-relevant part — when I go looking for an actual measurement to settle the question, the vault’s own best real rescue data argues for the packet’s conclusion, not against it, and the packet should have used that argument instead of assertion. Iranfar et al. 2026 (GROUNDING/chief-h03-deliv-rederivation.md) reports dual-AAV9-PHP.eB rescue in the DFNB16 mouse: 55–59% of OHCs STRC-positive, and separately, 49 ± 5% preservation of stereocilia bundle cohesion by SEM. If non-cell-autonomous secretion were rescuing untransduced neighbors at the 2–3x multiplier the h04 note speculates, bundle preservation should exceed the transduced fraction by a wide margin. It does not — it tracks slightly below it. That is a real, in-model-organism, in-target-cell-type data point that argues against a large bystander effect, and it is a far stronger argument than “there is no such thing as a bystander effect for a structural protein,” which is asserted by definition rather than shown.

Verdict on this point: the packet’s directional conclusion (no large non-cell-autonomous rescue) is probably right, but it is argued by fiat rather than evidence, and it contradicts material already in the vault without addressing it. This is the single most important physics question in the packet and it is the weakest-argued claim in the packet. Grade the argument as unsound even though the conclusion likely survives a proper argument.

Physical soundness of the tonotopic argument

Correct, and not new to this run — it is inherited from a prior audit. reviews/STRC Cure-Route Audit mRNA vs De-Novo Binder 2026-08-20.md:92 already states: “a scalar functional-OHC fraction cannot map onto a frequency-specific audiogram at all, however well calibrated. The cochlea is a tonotopic map; a dead region at one place is not compensated by a healthy region elsewhere… The scalar mean remains valid as a ceiling on protein level… but it is structurally incapable of being a ceiling on hearing outcome.” This is physically airtight: a place-coded organ cannot be summarized by a whole-organ mean when the output metric (an audiogram) is itself frequency-resolved. 01-outcome-chain-disposition.md:39-40 implicitly relies on the same geometry argument to kill the Schuknecht & Gacek anchor (“heterogeneous, localized… pathology. That geometry cannot identify one frequency-agnostic curve”).

One caveat: the specific sub-claim I was primed to check — that round-window delivery is basal-to-apical non-uniform — is not actually well supported by the mouse data this packet retrieved. GROUNDING/chief-h03-deliv-rederivation.md reports Landegger 2017 at “60 and 100% in apical and basal regions” and Iranfar 2026 at 58.7±3% basal vs ”[..]%” apical/55.4±2% medial — roughly flat across cochlear place in the mouse round-window data actually in hand. The 2026-08-20 audit’s “eff(x) over cochlear place, higher at base” claim is physically plausible in general (perilymph pharmacokinetics literature typically does show base-favored gradients for round-window dosing) but is not the specific evidence this packet retrieved, and this packet does not re-derive or re-check it. Not a defect in this run — it correctly inherits a sound prior argument — but the chain of custody on that specific numeric claim (basal > apical) is thinner than the certainty with which it is stated in the 2026-08-20 note.

The h03 delivery physics

Sound, and this is the best piece of physical reasoning in the packet. The chain is: (1) the vault’s cited deliv: 5 justification bundled two extrapolations — mouse→primate and neonate(P1)→four-year-old child — inside one number; (2) the mouse→primate half was directly tested with the same vector (Anc80L65) and the same route (round window) in Andres-Mateos 2022, and it failed for OHCs specifically: four transduced outer hair cells total, against 90–30% of inner hair cells in the same animals; (3) a different single-species, correct-cell-type, correct-disease data point exists (Iranfar 2026, dual-AAV9-PHP.eB, mouse DFNB16, ~55–59% OHC-positive with correctly trafficked stereocilin and restored tectorial-membrane imprints), which is why the packet does not collapse deliv to 1–2.

Is 3/5 the right number? Defensible, with one clarity gap worth naming. Four OHCs out of an NHP cochlea’s several-thousand-OHC population is ~0.03–0.05% transduction — that is not “reduced” delivery, it is a null result for that specific vector+route combination. A score of 3/5 is only defensible as a rating of the underlying concept (OHC-directed AAV gene therapy is achievable, demonstrated in the correct disease model with a different vector) — not as a rating of the specific single-vector Anc80L65/round-window plan h03 actually proposes, for which the only primate data point is close to zero. hypotheses/h03-mini-strc/index.md’s next_step field does distinguish these (“does ANY vector/route transduce primate OHCs?” as the P0), so the distinction exists, but the top-line score doesn’t visibly carry it. I would not have gone below 3 either, but I would flag in the hub, explicitly, that 3/5 scores the strategy space, not the plan currently on the bench.

One physical mechanism the packet never names, though it doesn’t need to for this run’s purpose: primate/human cochleae are roughly an order of magnitude larger by basilar-membrane length and perilymph volume than mouse, so a fixed injectate volume/dose delivered at the round window produces a far lower effective local concentration and a much longer diffusion path to the apex in primate than in mouse. That volumetric-dilution argument is a real, additional physical reason (beyond “species” as a label) the same dose regime would underperform in primate independent of anatomy differences in RWM permeability — worth adding to the hub as mechanism, not required for this run’s verdict.

Defects (file:line, severity, what is wrong, what would fix it)

BLOCKER — PANEL-PACKET/02-moderna-assessment.md:68 (“There is no cross-correction, no bystander effect… Every rescued cell must be individually reached… nothing else.”) asserts biological impossibility of non-cell-autonomous rescue as settled fact, while hypotheses/h04-mrna-lnp/phases/STRC mRNA-LNP Strategy B Audiogram Rescue.md:124 and hypotheses/h06-mrna-therapy/phases/STRC mRNA Therapy Hypothesis.md:85 already carry this as an open, unmodelled, potentially 2–3x-relevant effect in the same vault, uncited and unrebutted. Fix: either retrieve primary literature on stereocilin diffusion/uptake range (none currently in sources/lit/) and cite it, or soften the claim to “no evidence for, and the best available in-model-organism data (Iranfar 2026: bundle preservation ≈ transduction fraction, not amplified) argues against a large effect” — which is a defensible, evidence-backed claim the packet could have made instead.

MAJOR — PANEL-PACKET/02-moderna-assessment.md:136 (“no news at all for DFNB16”) and :62-64 (“there is no version of this that becomes one with more work”) overclaim closure of a question the vault’s own record (see above) leaves open. Fix: change to “no confirmed news” / “no evidence of a version of this with more work, pending a diffusion-range measurement that has not been attempted.”

MINOR — hypotheses/h03-mini-strc/index.md frontmatter deliv: 3 does not visibly distinguish “strategy-space plausible” from “plan-as-proposed near-zero in the only primate data point.” No numeric error, just a clarity gap given four cells out of thousands is closer to a null result than a “reduced” one. Fix: one added clause in next_step or the delivery line making explicit that 3/5 scores OHC-AAV delivery as a strategy, not the specific Anc80L65 single-vector/RWM combination.

MINOR — reviews/STRC Cure-Route Audit mRNA vs De-Novo Binder 2026-08-20.md:92’s claim that round-window delivery is “basal-to-apical non-uniform… higher at the base than the apex” is stated with more confidence than the actual retrieved mouse data supports (Landegger and Iranfar both show roughly flat apical-to-basal transduction in the numbers this packet’s own grounding cites). Not introduced by this run, but this run’s own grounding could have flagged the tension and didn’t. Fix: soften to “plausible by analogy to general round-window pharmacokinetics; not directly measured for AAV-STRC in the retrieved mouse studies.”

Where the packet is too pessimistic, if anywhere

The non-cell-autonomous-secretion closure (above) is the one place false certainty runs toward pessimism rather than hope — worth stating plainly since the brief asked for both directions. Everywhere else I checked, the packet’s pessimism is earned: the h03 downgrade is forced by a primary NHP paper with a named author overlap to the mouse paper being walked back, not by inference; the h09 downgrade cites a primary source (Chereau 2005) that directly contradicts the vault’s own prior structural justification; the “S tier is now empty” headline is the correct read of genuinely negative primate and lit-audit findings, not manufactured gloom. I did not find a case where a real positive finding was suppressed or downplayed.

What is genuinely good (specific, no padding)

• The h03 species+age double-extrapolation catch (GROUNDING/chief-andres-mateos-2022-nhp-ohc-primary.md, GROUNDING/chief-h03-deliv-rederivation.md addendum) is excellent primary-source detective work: it noticed the vault’s deliv: 5 rested on a mouse-P1 figure stacked with an inner-hair-cell clinical precedent for an outer-hair-cell hypothesis, retrieved the one primate paper that actually tests the same vector+route, and reported the exact count (four cells) rather than rounding it into a percentage that would have hidden how close to zero it is.
• Refusal to fabricate a replacement dB band (01-outcome-chain-disposition.md) is the correct call given the Bredberg 1968 and Schuknecht & Gacek 1993 sources genuinely do not contain the claimed transfer-function pairs — verified against the actual MinerU-parsed PDF text in GROUNDING/schuknecht1993-mineru/, which does show frequency-place-aligned, heterogeneous pathology exactly as the disposition describes, not a clean scalar curve.
• Catching the phantom Masaki-2009/210-kPa constant still executing in code four months after the parameter table was corrected (05-program-rerank.md:90, models/piezo_voltage_budget.py:262,271) — this is the right kind of check: a table fix that never propagated to the runtime is exactly the failure mode that fools future agents who grep the tables and never open the script.
• The KEYNOTE-942 Phase 2b vs Phase 3 hazard-ratio correction (02-moderna-assessment.md:109) — catching that the widely-quoted HR 0.51/0.41 figures are from the n=157 Phase 2b five-year follow-up, not the n=1,137 Phase 3 (which has not published an effect size), is a real statistical-rigor catch that most readers of the Moderna press release would miss.
• Honest labeling discipline throughout — PRIMARY vs SECONDARY vs INFERENCE tags are applied consistently and correctly (e.g., the organ-of-Corti-lacks-immune-cells claim is properly tagged SECONDARY, not elevated to fact it wasn’t retrieved as).