# Mock Peer-Review Report — Round 2 (revised resubmission) **Manuscripts assessed:** - **Paper D** (Nature Perspective) — *From disease name to design specification: a self-correcting AI-agent campaign for EoE* - **Paper E** (research article / bioRxiv) — *An integrated computational pipeline for EoE* **Review model:** The same three reviewers who returned Major revision in round 1 re-reviewed the revised resubmissions, each briefed on their prior concerns and asked for a point-by-point RESOLVED / PARTIALLY RESOLVED / NOT ADDRESSED verdict, plus a handling-editor synthesis. ## Outcome — progression across rounds | | Reviewer 1 (clinical) | Reviewer 2 (comp-bio/stats) | Reviewer 3 (AI-methods/ethics) | |---|---|---|---| | **Round 1** | Major revision | Major revision | Major revision | | **Round 2** | **Minor revision** | **Minor revision** | **Minor revision** | **Editor's decision (both papers): Minor revision.** Across all three re-reviews: **0 concerns Not Addressed, 17 Resolved, 14 Partially Resolved.** The partials are second-order — correct fixes that could be carried one step further (into every restatement of a claim, into computation, or into the companion paper) — not reopened science. *Generated 2026-07-08 · reviews produced by independent model instances under distinct expert personas. Advisory only — verify any specific factual claim before acting on it.* --- ## Editor's Decision Letter # Editor's Decision Letter **Manuscript pair:** Paper D ("From disease name to design specification: a self-correcting AI-agent campaign for EoE," Nature Perspective) and Paper E ("An integrated computational pipeline for EoE...," research article) Dear Authors, Thank you for this thorough and, in most respects, genuinely responsive resubmission. All three reviewers independently conclude that you engaged with the round-1 letter in good faith rather than defensively — the ledger denominator, the concordance statistics, the leave-one-cohort-out diagnostics, the SIGLEC6 and sST2 tissue-selectivity caveats, and the abstract-level caveat migration are all singled out, across all three reviews, as real fixes rather than rhetorical ones. That said, all three reviewers also converge on a second-order version of the same problem that opened round 1: several fixes are diagnosed correctly in prose but not carried through to computation, to every restatement of the claim, or to the companion paper. This round is about closing that gap, not reopening the science. --- ## Decision: Paper D — **Minor Revision** ## Decision: Paper E — **Minor Revision** All three reviewers independently recommend minor revision for both manuscripts. This is convergent, not split. Neither paper requires new analysis of a kind not already listed below; both are close to the bar for their venue. --- ## Essential Revisions — Paper D 1. **Supply the denominator for the 12-finding set, not just the 8-episode ledger (R2, R3).** The ledger's selection rule and prospective/reconstructed status are now well documented, but the total candidate pool and selection rule behind the 12 "principal findings" is not stated. Without it, the 8/3/1 novelty split (Fig. 3a) cannot be distinguished from a curated, flattering ratio. Add one explicit sentence stating the candidate pool size (or generative process) and the selection rule, and state plainly — as R2 requests — that the split describes the curated set, not a measured base rate. 2. **Move the "self-correction" scope statement to first use, not mid-paper (R2, R3).** The scoping sentence itself is exactly right ("a human-supervised system with a functioning correction loop, not an agent that reliably catches its own errors") but currently arrives several thousand words in. Per the same "caveat travels with the claim" principle already applied successfully to Paper E's Abstract, this sentence — or an equivalent one — must appear in the Standfirst or the opening section, since the title and Standfirst are Paper D's "abstract." 3. **Align Paper D's own prose with Paper E's therapeutic-landscape and DQ2.2/celiac corrections (R1, R2, R3 — all three).** Paper E now correctly states dupilumab's approval status, sharpens the mast-cell-depletion gap, discloses the DRB1-only pipeline scope, and states the celiac exclusion. Paper D still carries the older, unqualified "narrow targeted-therapy landscape" framing and states the DQ2.2/wheat mechanism as if pipeline-derived. A reader of Paper D alone is currently misled on both points. One sentence each, mirroring Paper E's language, resolves this. 4. **Complete the AI-authorship, competing-interest, and author-contribution sections (R1, R2, R3 — all three, and previously flagged in round 1).** This is now a two-round-unresolved placeholder in a paper whose thesis is that trust is earned through disclosure. This is a hard gate, achievable purely editorially, and must close before acceptance. 5. **Add the interpretive paragraph on the agent/human correction boundary (R3).** The mechanical 5/3 split is now reported, but the paper does not yet state the generalizable claim that follows from it — that agent self-correction, on this evidence, is bounded to claims reducible to an orthogonal quantitative re-check, while claims requiring literature breadth or normative judgment required a human. This is the paper's most valuable unstated finding; one paragraph converts a fact into a contribution. 6. **Soften "applied appropriate statistics where methods were specified" to match the concession that follows it (R2).** As written this sentence overclaims relative to the explicit statement, two paragraphs later, that the agent never audited its own statistical scaffolding. 7. **Flag the 1/28 null model's limitation (R3).** State explicitly that the calculation assumes exchangeable ranking across the eight candidate foods and does not account for prior population-level allergen-prevalence correlation with epitope burden — a one-sentence limitation, not a recomputation. --- ## Essential Revisions — Paper E 1. **Compute the patient-level SIGLEC6 statistic (R1, R2, R3 — all three, in near-identical language).** The cell-level P=3×10⁻³¹ is now correctly labeled as pseudoreplicated, and the paper states that "the per-patient contrast is the appropriate basis for a diagnostic cutoff" — but that statistic is never actually computed. With only 11 vs. 12 subjects this is a same-day calculation on data already in hand (a Mann-Whitney or mixed-effects model analogous to the existing Fig. 3d treatment) and is the single most-repeated unclosed item across all three reviews. This gates the diagnostic-cutoff claim at the level that matters for patient selection. 2. **Either run the requested GWAS hypergeometric/permutation test or fully retract the enrichment-flavored framing, including the "expected risk-gene-versus-effector split" language (R1, R2, R3).** The current text withdraws the word "enrichment" but keeps a distributional claim about where genetic risk concentrates within the signature, resting on n=7 with no test — and Fig. 4c's odds ratios remain visually adjacent to signature membership in a way the text disclaims but the figure still implies. Pick one: compute the matched-background test, or strip both the prose and the figure of enrichment-shaped framing. 3. **Propagate the mhcnuggets/netMHCIIpan-4.1 concordance finding into a sensitivity check on the index-case result itself (R2, echoed by R1's and R3's concern about the food-ranking claim's robustness).** The Jaccard of 0.20–0.62 on strong-binder calls means the two tools disagree on roughly half of nominal strong binders — yet Fig. 7's food-ranking, which drove the index-case test, is reported as if from a single consensus pipeline. Re-run the ranking using netMHCIIpan-4.1 alone on shared alleles and report whether milk/soy still place top-two. This is the one remaining check that bears directly on the paper's most rhetorically load-bearing result. 4. **Re-attach caveats at every restatement, not just first mention (R1, R2, R3).** This was the round-1 ask and it is now met at first mention (Abstract) but recurs unmet at second mention: the Discussion restates the refractory core as lending "credibility" without its single-cohort caveat; "reproducible endotypes" appears uncaveated even though it rests on the same six cohorts used for both derivation and validation; Fig. 6 presents sST2 in the same visual register as the two tissue-based assays despite the text's own tempering. Apply the same caveat-placement standard already achieved for the Abstract to every subsequent occurrence, including at least one figure-level (not only textual) treatment for sST2. 5. **State the refractory-core paired-cohort n explicitly next to the claim (R1, R2), and state the coverage threshold and characterization of the 168 genes excluded from the leave-one-cohort-out test (R3).** Both are narrow, computable transparency additions using data already in hand. 6. **Resolve the Methods/Results asymmetry on DRB1-only scope (R1, R2, R3).** The Results section now correctly discloses that DQ2.2 presentation was asserted from literature, not computed, and that celiac was excluded by a single tTG-IgA result (soften this to acknowledge IgA-deficiency false-negative risk, per R1). Add the equivalent one-line disclosure to Methods and Table 2 so a reader consulting either alone isn't misled. --- ## Disagreements Requiring Author Adjudication None of the three reviewers substantively disagree in direction this round — as in round 1, this is a convergent, not split, set of reviews. The only difference in emphasis: R1 continues to weight clinical-plausibility consequences (celiac false-negative modes, tissue-selectivity risk) more heavily, while R2 and R3 continue to weight statistical execution and rhetorical-calibration gaps (caveat recurrence, enrichment framing, null-model exchangeability) more heavily. Treat these as complementary; both sets of fixes are required. R3 additionally proposes an interpretive paragraph (Paper D, item 5 above) that is not explicitly requested by R1 or R2 but is compatible with and strengthens their concerns — authors should include it, as it directly serves the paper's stated purpose rather than adding new claims. --- ## Path to Acceptance — Paper D The substantive scientific scoping is done and done honestly — what remains is placement, completion, and one added interpretive layer, all editorial: surface the self-correction scope statement at first use rather than mid-paper, supply the missing denominator for the 12-finding set, align Paper D's own prose with Paper E's already-corrected dupilumab and DQ2.2/celiac language rather than leaving the fix siloed in the companion paper, close out the two-round-outstanding disclosure placeholders, and add the short paragraph generalizing the agent/human correction boundary. None of this requires new analysis. ## Path to Acceptance — Paper E The statistical scaffolding requested in round 1 has been substantively built — concordance stats, heterogeneity/LOCO diagnostics, derivation-cohort relabeling are all real, checkable additions. What remains is finishing three specific computations already within reach of data in hand (the patient-level SIGLEC6 statistic, the GWAS background-matched test or a full retraction of enrichment-shaped language, and a netMHCIIpan-alone sensitivity check on the index-case food ranking), plus applying the caveat-placement standard the authors already met in the Abstract consistently to every later restatement of the same claims. We anticipate one final, focused pass will be sufficient for acceptance. --- ## Full Reviews (with point-by-point response to prior concerns) ## Reviewer 1 ### Reviewer 1: Senior EoE clinician-scientist and mucosal immunologist (translational/clinical lens) --- ## PAPER D — "From disease name to design specification..." **Brief summary** This Perspective reports a human-supervised AI-agent campaign that carried EoE target discovery through structure-grounded design specifications, using an 8-episode correction ledger and a 12-finding novelty classification as evidence that the process was calibrated and "self-correcting." The revision now supplies a denominator for the ledger (214 artifacts → 12 principal findings → 8 correction episodes), scopes "self-correction" explicitly to one machine-initiated reversal plus three human catches, and reframes the index-case test as illustrative rather than decisive. **Strengths** - The taxonomy of catches — conceptual (MHC-I vs II), evidential (missed TCR paper), ethical (HLA-burden enrollment) — remains the paper's most durable contribution; the HLA-equity catch is a genuinely instructive example of why human oversight of trial design cannot be automated away. - The revised text is now explicit and unflinching about what "self-correcting" does *not* mean: "none of the eight episodes reversed a core methodological decision... the agent did not audit its own statistical scaffolding, and we do not claim it did." This is the right level of candor. - The denominator accounting (214 artifacts → 12 findings → 8-episode ledger, with an explicit statement that the ledger was reconstructed post hoc and is a lower bound) directly answers what was previously an unfalsifiable claim. **Major concerns** 1. **Dupilumab framing is now accurate but the "gap" claim still needs a sharper clinical citation.** The revision correctly states dupilumab is "FDA-approved for EoE in adults and children (≥1 year)" (Paper E, mirrored in Paper D's therapeutic-landscape references) and reframes the gap as "not a lack of any therapy but the absence of agents directed at specific disease drivers." This is clinically correct and a real improvement. However, Paper D's own text still opens with "a large treatment-refractory population and a narrow targeted-therapy landscape" without naming dupilumab or its approval — the accurate framing lives in Paper E's Introduction, not in Paper D. For a Perspective making claims about the therapeutic landscape, this should be stated once, explicitly, in Paper D's own prose, not left implicit or delegated to the companion article. 2. **The index-case reframing is verbally consistent but still structurally load-bearing.** The text now calls it "a single illustrative case (n=1)," reports the 1/28 chance calculation, and states plainly it should not be read as "the campaign's decisive evidence." Good. But the case is still the *only* clinical validation event in the entire campaign, is given its own named sub-heading treatment in Fig. 3b ("Blinded clinical calibration"), and is the culminating narrative beat of the "What the campaign found" section. A reader will still walk away treating it as the emotional climax of the paper, hedges notwithstanding. This is a presentation-weight problem, not a wording problem, and rhetorical structure, not just adjectives, communicates confidence. 3. **Wheat/DQ2.2 mechanistic claim needs the same hedge in Paper D as in Paper E.** Paper D states as fact that "wheat gliadin was heavily presented by the patient's celiac-associated DQ2.2 heterodimer" — but Paper E's revision now clarifies this DQ2.2 mechanism was *not* an output of the DRB1-only pipeline but a literature-drawn mechanistic interpretation, and that celiac disease was clinically excluded (tTG-IgA negative). Paper D omits both clarifications entirely. A reader of Paper D alone would reasonably (and wrongly) infer the DQ2.2 presentation claim was computed by the same pipeline used for the DRB1 predictions, and would not know celiac disease had been ruled out. This is a cross-paper consistency gap that matters clinically — an uncontrolled celiac confound would have undermined the entire wheat-discordance narrative. **Minor concerns** 1. SIGLEC6 tissue-selectivity caveat (basophils, other mast pools) — properly added in Paper E — is not mentioned anywhere in Paper D, even though Paper D features SIGLEC6 as a headline "novel claim upheld under test." Cross-reference or restate briefly. 2. Ref. 15 (Jacobse et al.) is now cited in-text ("concordant with prior multi-cohort syntheses of the EoE tissue transcriptome^15^") — resolved. 3. Competing-interest and author-contribution sections remain bracketed placeholders ("[Authors to complete]"). Understandable for a resubmission still in internal review, but this is now the second round with this unresolved — should be flagged as a hard gate before proof stage. **Feasible revisions** 1. Add one sentence early in Paper D's therapeutic-landscape paragraph naming dupilumab's approval status and the specific mast-cell-depletion gap, matching Paper E's Introduction language, rather than leaving this only in the companion paper. 2. Add the DQ2.2-is-a-literature-derived-interpretation-not-a-pipeline-output clarification, and the celiac-exclusion fact, to Paper D's wheat-discordance paragraph — one sentence each suffices. 3. Add a one-line SIGLEC6 tissue-selectivity caveat where SIGLEC6 is introduced in Paper D's "What the campaign found" section. 4. Consider demoting the index-case from its own Fig. 3b sub-panel heading to an inline callout within the broader calibration figure, or add a figure-legend line stating explicitly "not powered as a validation study" directly under the panel (not just in body text), so the caveat travels with the image. **Recommendation:** **Minor revision.** The round-1 essential revisions for Paper D (denominator, index-case de-escalation, self-correction scoping, dupilumab framing) are substantively resolved in the text; the remaining issues are consistency/completeness gaps between the two linked papers and residual rhetorical weighting, not new evidentiary problems. --- ## PAPER E — "An integrated computational pipeline for eosinophilic esophagitis..." **Brief summary** This research article details the full discovery-to-design pipeline: a 567-gene meta-signature from 9 cohorts, single-cell validation showing 14-fold mast-cell expansion, three structure-grounded therapeutic leads (SIGLEC6, IL1RL1/ST2, CCL26) each with a companion diagnostic, and an antigen-directed MHC-II arm culminating in a blinded single-patient food-trigger test. The revision adds heterogeneity statistics, concordance metrics for the two epitope predictors, explicit derivation-cohort caveats attached to headline numbers, a SIGLEC6 tissue-selectivity caveat, sST2 specificity caveats, and a DQ2.2/celiac clarification. **Strengths** - The Abstract now front-loads caveats at first mention rather than burying them: "a candidate treatment-refractory molecular core (observed in the single available paired pre/post-treatment cohort, and therefore provisional)" and "≥92% of severe-endotype patients (a derivation-cohort estimate, not an independently validated one)" both appear in the same breath as the headline claim. This is exactly the fix requested and it is done correctly, not just in the Discussion. - The concordance analysis between mhcnuggets and netMHCIIpan-4.1 (Spearman 0.74–0.84 across five shared alleles, Jaccard strong-binder overlap 0.20–0.62) is a genuine, specific, appropriately-interpreted addition — it distinguishes rank-agreement from threshold-agreement and uses the latter to justify treating presentation as "a prior, not a proof," which is clinically the correct framing for MHC-II binding predictions in general. - The SIGLEC6 tissue-selectivity caveat is well-executed clinically: it correctly identifies basophils and other tissue mast-cell pools as the off-target risk, correctly flags this against the depleting-antibody/ADC modality specifically, and proposes a concrete in-silico mitigation (public expression atlas query) and a design fallback (non-depleting/restricted-payload ADC). This is precisely the kind of clinical-plausibility check that was missing before. - The sST2 caveat is honest and appropriately differentiating: "soluble ST2 is a systemic analyte with known lack of esophageal specificity and appreciable population-level background variation, so its companion-diagnostic confidence is lower than the tissue-based readouts." This directly satisfies my round-1 concern rather than paying lip service to it. **Major concerns** 1. **The DQ2.2/DRB1 gap is disclosed but not resolved, and the disclosure creates a new internal tension.** The revision now states plainly that "our epitope predictions were generated on HLA-DRB1 alleles; the DQ2.2-gliadin presentation invoked to explain the wheat discordance is therefore a mechanistic interpretation drawn from the established celiac-disease literature, not an output of our DRB1 pipeline." This is honest and appropriately labeled a limitation. But it means the wheat-discordance narrative — repeatedly held up across both papers as "the most instructive result" — rests on a mechanism the pipeline did not actually model. The paper should be more careful not to let the strength of the *narrative* (presentation ≠ pathology) outrun the fact that the specific mechanistic claim underpinning the wheat case (DQ2.2 presentation) is asserted, not computed. I would ask that the Results text state directly, adjacent to the wheat discussion (not just in the two-clarifications paragraph), that the DQ2.2 presentation level itself was never quantified by this pipeline — the reader should not come away thinking the tool "showed" DQ2.2 over-presentation the way it showed DRB1 burden for milk and soy. 2. **Celiac exclusion is a single-patient point-in-time serology result, not a general safeguard, and the paper should say so.** "Tissue transglutaminase IgA negative" rules out active celiac disease in this one patient at this one test but does not establish absence of gluten-sensitivity more broadly (seronegative celiac, non-celiac gluten sensitivity) nor does it validate the DQ2.2-tolerance interpretation as generalizable. This is a minor but real overreach in an otherwise appropriately hedged paragraph — as written, "so the tolerated-wheat observation is not confounded by undiagnosed celiac disease" is stated with more certainty than a single tTG-IgA result supports (IgA deficiency can cause false-negative tTG-IgA; this is not mentioned). 3. **SIGLEC6 IHC statistic (P=3×10⁻³¹) — n reported, but the diagnostic-cutoff implication is still asserted, not derived.** The revision correctly discloses the unit-of-analysis issue (166,420 cells across 11 EoE/12 healthy subjects, cell-level not patient-level) and correctly states "the per-patient contrast is the appropriate basis for a diagnostic cutoff" — but no per-patient contrast is actually shown. With n=11 EoE / 12 healthy subjects, a patient-level effect size and (nonparametric, given the small n) test statistic should be computable from the same single-cell object and reported here; this is a fully in-silico, immediately achievable fix, and without it the companion-diagnostic addressability numbers built on this assay remain ungrounded at the level that matters for actual patient selection. 4. **Refractory-core Fig. 4d unit-of-analysis** — addressed for SIGLEC6 IHC but not explicitly for the PPI-refractory persistence result. The paired-cohort n (patients, not samples) supporting "9 of 10 targets... persisted in non-responders" should be stated explicitly next to the claim, as was done for the IHC statistic. **Minor concerns** 1. GWAS overlap: the revision correctly withdraws the enrichment claim ("we do not assert statistical enrichment... because the analysis lacked a fixed, pre-specified GWAS candidate-gene universe screened against a matched background") — this is a defensible and honest retreat rather than a fix, and I accept it as adequate, though a matched-background hypergeometric test (using, e.g., all druggable/secreted genes as the universe) remains straightforward to compute and would strengthen rather than weaken the paper. 2. I²/leave-one-out diagnostics are reported for CCL26, ALOX15, DSG1, POSTN, IL1RL1 — SIGLEC6 (36.6%) is included, satisfying my ask for the core headline genes. 3. The sST2 "repurposing an established cardiac-biomarker platform" phrase is retained in the text, but is now immediately followed by the specificity caveat — acceptable, though I'd still suggest moving "cardiac-biomarker" out of the sentence that introduces the assay, since the platform pedigree is not really a point in favor of an esophageal diagnostic and reads as promotional framing despite the adjacent hedge. **Feasible revisions** 1. Compute and report a patient-level (n=11 vs 12) effect size/statistic for the SIGLEC6⁺ mast-fraction comparison, alongside (not instead of) the cell-level P-value, and use the patient-level number to anchor the diagnostic-cutoff claim. 2. State the paired-cohort patient n explicitly in the Fig. 4d refractory-core sentence. 3. Add one sentence adjacent to the wheat-discordance discussion clarifying that DQ2.2 presentation level was not computed by this pipeline (currently this only appears in the Methods-adjacent "two clarifications" paragraph, somewhat separated from the narrative claim itself). 4. Soften "not confounded by undiagnosed celiac disease" to acknowledge tTG-IgA's known false-negative modes (e.g., IgA deficiency) — one clause suffices. 5. Compute a hypergeometric/permutation enrichment test against a matched background (e.g., all secreted/surface druggable genes, or all genes tested for DE) for the GWAS overlap, even if reported as a secondary/exploratory statistic alongside the current honest disclaimer. **Recommendation:** **Minor revision.** All five round-1 statistical/clinical asks that fall in my domain (concordance stats, GWAS enrichment framing, heterogeneity diagnostics, SIGLEC6 selectivity, sST2 specificity, DQ2.2/celiac) have been substantively engaged with rather than deflected; remaining gaps (patient-level SIGLEC6 statistic, DQ2.2 narrative-vs-computation distinction, refractory-core n) are targeted, computable, and do not require new data. --- ## Response to prior concerns — Paper D 1. **Ground the correction ledger with a denominator.** RESOLVED. The revision states "the campaign produced 214 lineage-tracked deliverables. From these we defined a set of twelve principal findings... it is these twelve that were classified against the literature," and separately: "the eight-episode correction ledger... is a *separate* selection... the selection rule was 'a course-correction occurred'... this ledger was reconstructed from the campaign's artifacts and transcripts *after* the fact, not logged prospectively... it is therefore a lower bound." All three requested elements (total count, selection rule, prospective-vs-reconstructed) are present and specific. 2. **De-escalate the index-case as evidence.** PARTIALLY RESOLVED. The wording is fully de-escalated ("a single illustrative case (n = 1), not a validation study," chance calculation of 1/28 = 0.036 given, "we do not present it as the campaign's decisive evidence"). However, structurally the case still closes the "What the campaign found" section and receives a named sub-heading in Fig. 3b ("Blinded clinical calibration"), giving it outsized narrative real estate relative to its evidentiary weight — see Major Concern #2 above. 3. **Scope "self-correction" to what is actually demonstrated.** RESOLVED. The text now states directly: "Of the eight, five were machine-initiated and three were human catches... none of the eight episodes reversed a *core methodological* decision... 'Self-correcting' here therefore denotes a human-supervised system with a functioning correction loop, not an agent that reliably catches its own errors." This is an explicit, well-scoped statement matching the editor's requested distinction, and the title is retained with this scoping sentence rather than being changed — a permissible resolution per the editor's stated latitude. 4. **Update the therapeutic-landscape framing.** PARTIALLY RESOLVED. Paper E's Introduction now correctly states dupilumab's approval status and sharpens the actual gap to mast-cell depletion. Paper D itself, however, still frames the landscape only as "a large treatment-refractory population and a narrow targeted-therapy landscape" without naming dupilumab's approval in its own prose (see Major Concern #1) — the fix lives in the companion paper, not in Paper D directly. 5. **Housekeeping (ref. 15 citation; competing-interest/author-contribution placeholders).** PARTIALLY RESOLVED. Ref. 15 is now cited in text — resolved. Competing-interest and author-contribution sections remain bracketed placeholders — not addressed, though this is a pre-publication administrative gate rather than a scientific one. --- ## Response to prior concerns — Paper E 1. **Attach caveats at first mention, especially in Abstract.** RESOLVED. The Abstract now reads "a candidate treatment-refractory molecular core (observed in the single available paired pre/post-treatment cohort, and therefore provisional)" and "addresses ≥92% of severe-endotype patients (a derivation-cohort estimate, not an independently validated one)" — both caveats travel with the headline number in the same sentence, in the Abstract itself, not deferred to Discussion. 2. **Report concordance statistics for mhcnuggets vs. netMHCIIpan-4.1.** RESOLVED. "Per-allele Spearman rank correlation ranged 0.74–0.84 (mean 0.79) and strong-binder agreement (Jaccard) ranged 0.20–0.62 (mean 0.42)," computed on 2,424 peptides per allele across five shared DRB1 alleles, with an honest interpretation ("agree well on rank-ordering... only moderately on the hard strong-binder threshold"). 3. **Address circularity/independence of validation cohorts.** RESOLVED. The text explicitly states the addressability figures "are derivation-cohort estimates — computed on the same cohorts used to derive the signature and endotypes, with no independent held-out validation set — and should be read as upper bounds pending prospective evaluation." 4. **Add a formal background/enrichment test for the GWAS-gene overlap.** PARTIALLY RESOLVED. The paper now correctly withdraws the enrichment claim rather than asserting it improperly ("we do not assert statistical enrichment"), which is an honest and defensible response to the concern, but the actual requested test (hypergeometric/permutation against a matched background) is not performed — the gap is disclosed rather than closed. Since this is a computable in-silico addition, I do not consider disclosure alone sufficient to fully resolve. 5. **Add heterogeneity/robustness diagnostics for the meta-signature.** RESOLVED. Leave-one-cohort-out results (399/399 robust-core genes surviving) and I² statistics for CCL26 (97.6%), ALOX15 (97.6%), DSG1 (96.4%), POSTN (71.8%), IL1RL1 (72.9%), SIGLEC6 (36.6%) are all reported, covering the requested headline genes plus additional ones. 6. **Address SIGLEC6 tissue-selectivity/off-target risk.** RESOLVED. Explicit paragraph added: SIGLEC6 "is not exclusive to esophageal mast cells — it is also expressed on basophils and mast-cell pools in other tissues — so a depleting agent carries a systemic off-target mast/basophil-depletion risk," with a concrete mitigation path proposed. 7. **Resolve DQ2.2/celiac and DRB1-vs-DQ prediction gap.** PARTIALLY RESOLVED. Celiac exclusion is now stated (tTG-IgA negative) and the DRB1-only scope of the actual pipeline is disclosed clearly as a limitation. However, as noted in Major Concern #1/#2, the disclosure is somewhat siloed from the wheat narrative itself, and the celiac-exclusion claim is stated with more certainty than a single serologic test supports (no acknowledgment of tTG-IgA false-negative modes, e.g. IgA deficiency). 8. **Report unit-of-analysis and n for extreme p-values.** PARTIALLY RESOLVED. For SIGLEC6 IHC, the unit-of-analysis problem is now explicitly disclosed and the subject-level n (11 EoE, 12 healthy) is given — but no patient-level statistic is actually computed to replace or accompany the cell-level P=3×10⁻³¹ as the basis for the diagnostic cutoff, so the underlying request (a properly-powered statistic for the claim actually being made) is only partly met. For the refractory-core (Fig. 4d), no explicit n is stated at all — not addressed. 9. **Temper the sST2 diagnostic framing.** RESOLVED. The text now explicitly flags sST2 as lower-confidence than the tissue-based assays: "soluble ST2 is a systemic analyte with known lack of esophageal specificity and appreciable population-level background variation, so its companion-diagnostic confidence is lower than the tissue-based readouts and it would require esophagus-anchored cutoff calibration." --- **Cross-cutting comment:** This is a substantially more disciplined pair than the version I first reviewed — the authors engaged with essentially every specific, numbered request rather than making generic gestures toward "limitations," and several fixes (the mhcnuggets/netMHCIIpan concordance numbers, the SIGLEC6 tissue-selectivity caveat, the Abstract-level caveat migration) are exactly what was asked for, not partial substitutes. The residual issues are narrow and almost all computable from data already in hand: a patient-level SIGLEC6 statistic, an actual GWAS background-matched test, an explicit n for the refractory-core claim, and better narrative-level (not just Methods-level) separation between what the DRB1 pipeline computed and what the DQ2.2/wheat mechanism merely asserts from outside literature. None of this requires new science, and I'd expect one more focused pass to close it out. ## Reviewer 2 ### Reviewer 2: Computational biologist / biostatistician — multi-cohort transcriptomics, MHC-binding prediction, reproducibility --- ## PAPER D (Nature Perspective) **Brief summary** This Perspective uses an EoE discovery-to-design campaign as a case study to argue for a human-AI operating model in agentic science, centering on a correction ledger, a novelty classification of twelve findings, and a blinded single-patient calibration. The revision adds a denominator for the ledger, explicit "self-correction" scoping, and a chance-probability calculation for the index case. **Strengths** - The five-practice distillation (orthogonal confirmation for novelty claims, auditable ledger, human sign-off on equity-relevant design, calibration against withheld ground truth, provenance tracking) is a genuinely reusable contribution independent of the EoE result set. - The HLA-burden recruitment catch remains the strongest single exhibit in the piece and is now stated with appropriate weight rather than as a throwaway anecdote. - The provenance graph (Fig. 4, 36 artifacts/83 edges) is a concrete, checkable claim rather than an assertion. **Major concerns** 1. The 12-finding "principal findings" set is still selected by the authors' own judgment of what counts as "substantive enough to assert as results," with no external or pre-specified rule; this is now disclosed rather than fixed. That is an acceptable resolution for a Perspective but means Fig. 3a's 8/3/1 split is a description of curation, not a measurement of calibration rate — the text should say this once, plainly, rather than leaving the reader to infer it from the denominator paragraph. 2. The chance calculation for the index case (1/28, permutation-confirmed per Paper E) is now present, which is good, but Paper D itself only states "chance probability 1/28 = 0.036" without stating the permutation confirmation performed in Paper E — a cross-reference gap between the two linked papers on the exact same number. 3. None of the eight ledger episodes still touch the statistical core (batch harmonization, meta-analysis, cell annotation, thresholds) — this is now explicitly conceded ("the agent did not audit its own statistical scaffolding") rather than obscured, which is the right move, but it somewhat undercuts the paper's own framing of the agent as having "applied appropriate statistics where methods were specified" in the Lessons section — that sentence still reads as a stronger statistical-competence claim than the ledger supports and should be softened to match the concession two paragraphs later. **Minor concerns** 1. Ref. 15 (Jacobse et al.) is now cited in-text ("concordant with prior multi-cohort syntheses of the EoE tissue transcriptome^15^") — confirmed resolved. 2. Competing-interest and author-contribution sections remain bracketed placeholders; for a paper substantially about disclosure and correction-transparency, this residual gap is a live inconsistency the editor already flagged and it is unresolved in this draft. 3. "Notably, once the panel was completed the agent's own predictor independently recovered the same DR7 β-casein window" is a nice result but is stated identically in both papers without either flagging whether this is a fully independent recovery or partially informed by having just read the reference paper that named the epitope — worth one clarifying clause on analysis order (was the panel gap fixed before or after reading ref 17's specific epitope window?). **Feasible revisions** 1. Add one sentence stating explicitly that the 8/3/1 novelty split is a description of the curated set, not a base rate over all campaign outputs. 2. Soften "applied appropriate statistics where methods were specified" to acknowledge, in the same breath, that statistical-scaffolding decisions were not subject to the self-correction loop described. 3. Complete the competing-interest/author-contribution placeholders before resubmission — this is pure editorial housekeeping and fully in scope. 4. Add a clause clarifying the temporal/informational relationship between reading ref. 17 and the predictor's independent recovery of the DR7 β-casein window, to preclude a hindsight-bias reading. **Response to prior concerns** 1. *Ground the ledger with a denominator (R1/R2/R3).* **RESOLVED.** New paragraph: "the campaign produced 214 lineage-tracked deliverables. From these we defined a set of twelve principal findings... The eight-episode correction ledger... is a *separate* selection... We are explicit that this ledger was reconstructed... after the fact, not logged prospectively." This directly answers (a), (b), (c). 2. *De-escalate the index-case as evidence (R1/R2/R3).* **RESOLVED.** Text now reads "This episode is a single illustrative case (n = 1), not a validation study... roughly 1 time in 28 (P = 0.036). We weight it as a concrete, hypothesis-generating demonstration... rather than as statistical proof, and we do not present it as the campaign's decisive evidence." "Strongest single piece of external validation" language is gone. 3. *Scope "self-correction" (R1/R3, R2's statistical-core point).* **RESOLVED.** The paper now states the 5/3 machine/human split, names the one clean unprompted reversal (S100A8/9→S100A4) explicitly, and states "none of the eight episodes reversed a core methodological decision... 'Self-correcting' here therefore denotes a human-supervised system with a functioning correction loop, not an agent that reliably catches its own errors." This is a direct, textual scoping — resolved, and R3's retitling option was left as author's latitude per the editor's note, which is acceptable. 4. *Update therapeutic-landscape framing (R1).* **RESOLVED** in Paper E ("now FDA-approved for EoE in adults and children (≥1 year)... The gap we target is therefore not a lack of any therapy but the absence of agents directed at specific disease drivers"); Paper D's own EoE-therapeutics sentence ("a large treatment-refractory population and a narrow targeted-therapy landscape") is compatible with this and not contradicted, though Paper D does not itself restate the dupilumab approval — acceptable since Paper E carries it and D cites the same refs. 5. *Housekeeping — ref. 15 citation and placeholders (R1/R2, R3).* **PARTIALLY RESOLVED.** Ref. 15 citation confirmed in text (resolved). Competing-interest/author-contribution sections remain explicitly bracketed placeholders — not resolved, and this was specifically flagged as a "live irony" worth fixing before acceptance. **Recommendation: Minor revision.** The substantive statistical/evidentiary concerns are now well-scoped and honestly caveated; what remains is housekeeping (placeholders) and small consistency edits, both squarely in scope and low-risk to defer to a final proof stage — but they should not be waved through unmentioned given the editor's explicit flag. --- ## PAPER E (research article) **Brief summary** This companion research article reports the EoE discovery pipeline itself: a 567-gene meta-signature from 9 harmonized cohorts, single-cell validation, endotype/refractory-core analysis, three structure-grounded therapeutic design specs with companion diagnostics, and an antigen-directed MHC-II arm calibrated against one blinded patient. The revision adds heterogeneity statistics, concordance metrics for the two epitope predictors, explicit derivation-cohort labeling, a walked-back GWAS claim, and a DQ2.2/DRB1 clarification. **Strengths** - The mhcnuggets vs. netMHCIIpan-4.1 concordance analysis (Spearman 0.74–0.84, Jaccard 0.20–0.62 on 2,424 peptides × 5 shared alleles) is exactly the requested statistic and is reported with an honest interpretive gloss distinguishing rank agreement from threshold agreement. - The leave-one-cohort-out robustness check (399/399 genes with sufficient coverage) plus per-gene I² reporting is a substantial and well-executed upgrade to signature credibility. - The SIGLEC6 tissue-selectivity caveat and the sST2 specificity caveat are now both present, specific, and actionable (naming basophils/other mast pools, naming the ELISA's lack of esophageal specificity). **Major concerns** 1. **Concordance statistic is reported but its consequence for the downstream burden analysis is not carried through.** Jaccard 0.20–0.62 (mean 0.42) on the strong-binder call means the two tools disagree on roughly half of nominal "strong binders" — yet Fig. 7's epitope-burden rankings (which drove food prioritization in the index-case test) are reported as if from a single consensus pipeline (mhcnuggets, "cross-validated against" netMHCIIpan). The paper does not state whether the index-case food-ranking result is robust to using netMHCIIpan-4.1 alone instead of mhcnuggets alone — given a moderate Jaccard, this is a directly checkable sensitivity analysis that was requested implicitly (E2) and is still missing in its downstream application, only supplied for the abstract binding-prediction comparison in isolation. 2. **GWAS overlap claim is now appropriately hedged in prose but the hedge does not fully resolve E4.** The revised text: "We present this as per-gene genetic annotation rather than a signature-level enrichment claim: because the analysis lacked a fixed, pre-specified GWAS candidate-gene universe screened against a matched background, we do not assert statistical enrichment." This is honest, but it resolves the concern by *withdrawing* the enrichment claim rather than *computing* the requested hypergeometric/permutation test. The Editor's letter asked for the test to be added, not for the claim to be retracted — retraction is an acceptable alternative resolution in principle, but the paper should be explicit that this is what happened (a claim removed, not a claim substantiated), and Fig. 4c's odds ratios (CAPN14 1.77, DSG1 2.22) are still displayed adjacent to signature membership in a way that visually implies aggregate enrichment even though the text disclaims it. 3. **Validation-cohort independence: resolved in prose, but the practical consequence for the paper's own confidence language is inconsistent.** The Abstract now says "a derivation-cohort estimate, not an independently validated one" for the ≥92% diagnostic figure, and Fig. 6's caption/text repeats "should be read as upper bounds pending prospective evaluation" — good. But the refractory-core claim, also derivation-cohort/single-cohort, is called "a treatment-refractory molecular core (observed in the single available paired pre/post-treatment cohort, and therefore provisional)" in the Abstract — also good — yet the Discussion still states "which both validates the approach and lends credibility to its more novel outputs, chief among them the mast-restricted target SIGLEC6 **and the treatment-refractory molecular core**" without re-attaching the single-cohort caveat at that specific sentence. This is precisely the "caveat at first mention only, not every headline restatement" pattern the editor asked to eliminate — it recurs here in the Discussion. 4. **Sample-size/unit-of-analysis for the SIGLEC6 IHC statistic is resolved quantitatively but raises a new question.** The text now states P=3×10⁻³¹ derives from 166,420 cells across 11 EoE and 12 healthy subjects and correctly flags this as a cell-level, not patient-level, statistic — this directly answers E8. However, no patient-level statistic (e.g., per-subject mean SIGLEC6⁺ mast fraction compared by Mann-Whitney across 11 vs 12 subjects, analogous to the Fig. 3d treatment) is actually computed and reported alongside it. The paper correctly diagnoses the pseudoreplication problem but does not fix it with the patient-level test that would let a reader judge whether the diagnostic threshold is powered at n=11/12. **Minor concerns** 1. The wheat/DQ2.2 resolution (E7) is thorough and includes the celiac-exclusion clarification (tTG-IgA negative) and the DRB1-only scope limitation — well done, but the Methods section's antigen-directed-arm paragraph still only mentions DRB1 predictions and does not add a sentence noting DQ/DP extension is future work, leaving a Methods/Results asymmetry (Results flags the limitation; Methods doesn't reflect it). 2. I² values for CCL26 and ALOX15 (97.6%) are very high; the paper appropriately attributes this to platform/cohort diversity and pivots to direction-consistency, but does not report the DerSimonian-Laird τ² alongside I², nor whether a random-effects vs. fixed-effects sensitivity changed which genes cleared the leave-one-out bar. 3. Fig. 7's "31,512 predictions" scale is stated but the shared-allele concordance analysis subsamples to 5 of 13 DRB1 alleles (2,424 peptides each) — worth stating what fraction of the full prediction set (31,512) this concordance check actually covers, since 5/13 alleles is well under half. **Feasible revisions** 1. Re-run the index-case food ranking using netMHCIIpan-4.1 alone (on the shared alleles) as a sensitivity check and report whether milk/soy still rank top-two — this directly tests whether the headline blinded-calibration result is an artifact of predictor choice. 2. Compute and report the patient-level (11 vs. 12 subject) SIGLEC6⁺ mast-fraction statistic alongside the existing cell-level P-value, to give the diagnostic-cutoff claim an appropriately powered basis. 3. Re-attach the single-paired-cohort caveat to the refractory-core mention in the Discussion paragraph, not just the Abstract and Fig. 4d legend. 4. Add one Methods sentence noting the antigen-directed arm's current DRB1-only scope and flagging DQ/DP extension as a defined next computational step, matching the Results-section disclosure. 5. Report DerSimonian-Laird τ² alongside I² for the headline genes, and state what fraction of the 31,512 total MHC-II predictions the 5-allele concordance check actually covers. **Response to prior concerns** 1. *Caveats at first mention (R1/R2/R3).* **PARTIALLY RESOLVED.** Abstract and most first-mentions (≥92% diagnostic addressability, refractory core) now carry caveats immediately — confirmed improvement. But the Discussion's restatement of the refractory core as lending "credibility" without re-attaching the single-cohort caveat shows the pattern recurring at second mention (see Major Concern 3). 2. *Report concordance statistics for mhcnuggets vs. netMHCIIpan-4.1 (R1/R2/R3).* **PARTIALLY RESOLVED.** The concordance numbers themselves (Spearman 0.74–0.84, Jaccard 0.20–0.62) are now reported exactly as requested — a real and substantial addition. But the moderate Jaccard is not propagated into a sensitivity check on the downstream index-case ranking that the abstract treats as a headline result (Major Concern 1), so the gate this item was meant to open (credibility of Fig. 7's specific food-ranking claim) is only half-resolved. 3. *Address circularity/independence of validation cohorts (R1/R2/R3).* **RESOLVED.** Explicit relabeling as derivation-cohort/upper-bound estimates now appears in the Abstract, Results, and Fig. 6 text: "These addressability figures are derivation-cohort estimates... and should be read as upper bounds pending prospective evaluation." This is the clearest, most complete resolution among the requested items. 4. *Formal GWAS enrichment test (R2/R3).* **PARTIALLY RESOLVED.** The claim is now explicitly withdrawn/hedged rather than tested ("we do not assert statistical enrichment"), which is a legitimate way to close the gap, but no hypergeometric/permutation test was actually run, and Fig. 4c's visual framing (odds ratios beside signature membership) still invites an enrichment reading the text disclaims. Authors should either run the simple background-matched test (fully feasible in silico, and the more complete resolution) or strip the odds-ratio display down to unambiguous single-gene annotation framing. 5. *Heterogeneity/robustness diagnostics for the meta-signature (R2, R1's ancestry point, R3's multiple-comparisons point).* **RESOLVED** for the core ask: leave-one-cohort-out (399/399 robust) and per-gene I² for CCL26, POSTN, DSG1, ALOX15, IL1RL1, SIGLEC6 are now reported. Forest plots per se are not shown as figures, but the equivalent numeric summary is present in text — acceptable substitute. τ² is not reported (minor gap noted above). 6. *SIGLEC6 tissue-selectivity caveat (R1).* **RESOLVED.** Explicit paragraph in Results naming basophils and other-tissue mast pools, quantifying the risk qualitatively, and recommending a non-depleting alternative if selectivity is insufficient. 7. *DQ2.2/celiac and DRB1-vs-DQ gap (R1/R2).* **RESOLVED** in Results (celiac exclusion stated, DQ2.2-vs-DRB1 scope gap explicitly flagged as a limitation); **NOT propagated to Methods** (minor concern 1 above), so call it resolved in substance with a small consistency gap remaining. 8. *Unit-of-analysis/n for extreme p-values (R2).* **PARTIALLY RESOLVED.** The cell-level vs. patient-level distinction is now explicitly named and the n (11 EoE/12 healthy subjects) is disclosed — this answers the diagnostic half of the request. But the actual patient-level statistic that would resolve the pseudoreplication concern is not computed (Major Concern 4). The Fig. 4d refractory-core result's unit of analysis (genes × patients within the one paired cohort) is not separately addressed at all. 9. *Temper sST2 diagnostic framing (R1).* **RESOLVED.** New sentence explicitly downgrades ST2 confidence relative to the tissue-based assays and calls for esophagus-anchored cutoff calibration. **Recommendation: Minor revision.** The great majority of the statistical-scaffolding requests have been substantively addressed with real numbers (concordance stats, I², leave-one-out, derivation-cohort relabeling); what remains is narrow and fully computable — a predictor-choice sensitivity check on the one downstream result that matters most rhetorically (the blinded index case), a patient-level version of the SIGLEC6 IHC statistic, and consistent caveat placement at every restatement of the refractory-core claim, not just its first mention. --- **Cross-cutting comment.** This is an unusually responsive revision: the authors engaged concretely with almost every numbered item from the round-1 letter rather than reasserting the original framing with added hedges, and the concordance/heterogeneity/derivation-cohort work in Paper E in particular represents real added rigor, not cosmetic wording changes. The residual issues in both papers are now second-order — a recurring-caveat placement problem in Paper E's Discussion, an unresolved placeholder in Paper D's disclosures, and one analysis (GWAS enrichment) resolved by claim-withdrawal rather than by running the requested test, which is defensible but should be labeled as such rather than presented as equivalent compliance. None of these blocks acceptance; all are fixable in a light final pass. ## Reviewer 3 ### Reviewer 3: AI-in-science methodologist and research-ethics reviewer (autonomous/agentic-science evaluation, reproducibility, equity in trial design, AI-disclosure norms) --- ## PAPER D — "From disease name to design specification: a self-correcting AI-agent campaign for EoE" **Brief summary.** This Perspective narrates a human-supervised AI agent's campaign on public EoE data, using an 8-episode "correction ledger" and a 12-finding novelty classification to argue that agentic science can be conducted in a calibrated, auditable, trustworthy manner. The revision adds a denominator for the ledger, de-escalates the index-case language, and explicitly scopes "self-correction" to one clean machine episode plus a human-supervised oversight loop. **Strengths** - The revised framing — "self-correcting" now denotes "a human-supervised system with a functioning correction loop, not an agent that reliably catches its own errors" — is exactly the kind of scoping the field needs and is stated plainly rather than buried. - The explicit statement that none of the eight episodes touched core statistical/methodological scaffolding is an unusually honest admission for a paper whose thesis is trust. - The HLA-agnostic recruitment redesign remains the strongest single piece of evidence in the paper and is now presented with the correct epistemic register (human catch, ethical failure invisible to quantitative checks). **Major concerns** 1. **The denominator is still incomplete.** The revision states the ledger was "reconstructed... after the fact, not logged prospectively" and is "a lower bound," which is a genuine improvement. But the Editor's request (R1/R2/R3 jointly) also asked for the total pool of claims/decisions from which the 12 "principal findings" were selected — i.e., how many candidate findings existed before the 12 were chosen, and what the selection rule was beyond "substantive enough to assert as results." The text gives the selection rule for the *ledger* (episodes where "a course-correction occurred") but not for the *12-finding set*. Without knowing whether 12 was drawn from 15 or 150 candidate claims, "sparingly asserted novelty" (3/12 novel) cannot be assessed as a calibration statistic — it could equally reflect a curation choice made to produce a flattering ratio. This is PARTIALLY, not fully, resolved. 2. **"Self-correcting" is now scoped in prose but not in the title.** Per the editor's latitude, retaining the title was permitted only with an explicit early scoping sentence — which the paper now supplies, but only in the middle of the "Calibration and self-correction" section, several thousand words in. A reader who stops at the Standfirst or the opening section still encounters unscoped "self-correcting" language ("a self-correcting AI-agent campaign," Standfirst; "self-correction" implied as the throughline of "Anatomy of the campaign"). The scoping needs to appear at or near first use, not four sections later — this is the same "attach caveats at first mention" principle the editor applied to Paper E's abstract, and Paper D's title/standfirst is its own "abstract." PARTIALLY RESOLVED. 3. **The 1/28 chance calculation is a permutation-of-convenience, not a null model.** The paper reports "chance probability roughly 1/28" for two true triggers landing in the top two of eight candidates, computed as if ranking were a uniform-random draw over C(8,2)=28 pairs. But this null does not reflect the actual decision structure: milk and soy are, a priori, the two allergens most likely to dominate any MHC-II presentation-burden ranking simply because casein/soy epitope repertoires are large relative to the other six foods (the paper itself notes "the two soy allergens carried the highest epitope burden" in Paper E, prior to knowing the patient's triggers). If high epitope burden independently correlates with allergenic prevalence in the general food-allergic population, the "chance" model is not exchangeable across foods and 1/28 overstates the surprise. This is a subtler point than the previous round's request, but it matters: the number is presented as if it settles the informativeness of the n=1 test, when the appropriate null would need to account for prior population-level allergen prevalence. I would not block acceptance on this point but it should be flagged as a remaining soft spot in a paper about calibration. 4. **Epistemic asymmetry between agent-caught and human-caught errors remains under-theorized.** The paper is now admirably clear on the mechanical fact (5 machine, 3 human), but does not engage with *why* the split falls this way, which is the more important scientific claim for a methods-facing Perspective. The S100A8/9→S100A4 reversal was catchable by the agent because it reduces to a directional consistency check across cohorts — a computable operation. The three human catches (MHC class I/II framing, missed prior art, HLA-equity) are catchable only by a system with either broader literature retrieval, disease-taxonomic priors, or a normative equity commitment — none of which the current agent architecture possesses natively. The paper's "Lessons" section gestures at this ("agent's failure modes are specific and predictable") but never draws the generalizable claim: agent self-correction, on this evidence, is currently bounded to claims reducible to an orthogonal quantitative test, and everything requiring domain judgement or values is out of reach. Stating this crisply would strengthen the paper's contribution to the agentic-science literature considerably. 5. **AI authorship/competing-interest sections remain placeholders.** "[Authors to complete]" and "[Author-specific entries to be completed]" persist in this revision. For a paper whose entire thesis is that trust is earned by disclosure, submitting with the disclosure sections unfinished is a live irony the editor already flagged (R3, round 1) and it has not been fixed. This is a hard gate for a paper on this topic, not a formality. **Minor concerns** 1. The therapeutic-landscape correction (dupilumab's FDA-approval status) was made in Paper E's Introduction but Paper D's own "EoE is a fitting proving ground" paragraph still describes only "a narrow targeted-therapy landscape" without naming dupilumab or its approval status — the two papers should be consistent on this point since D references companion work. 2. Ref. 15 (Jacobse et al.) is now cited in text ("concordant with prior multi-cohort syntheses of the EoE tissue transcriptome^15^") — confirmed resolved. 3. Figure 3's caption bundles the literature-calibration classification and the correction ledger and the index-case test into one dense figure; given how much interpretive weight this figure now carries, consider splitting for clarity. 4. "Notably, once the panel was completed the agent's own predictor independently recovered the same DR7 β-casein window" is repeated near-verbatim in both papers D and E — fine for companion papers, but Paper D should clarify this is not an independent replication (same model, same underlying prediction pipeline, run after the gap was identified) rather than a second confirmatory data point. **Feasible revisions** 1. State explicitly the total candidate pool (or the process that generated candidates) from which the 12 principal findings were drawn, and the selection rule, mirroring the transparency already achieved for the 8-episode ledger. 2. Move an unambiguous one-sentence scope statement for "self-correcting" into the Standfirst or the first paragraph of "The trust problem," not just into the mid-paper "Calibration" section. 3. Add a sentence acknowledging that the 1/28 null assumes exchangeable food-ranking probability and does not account for prior population-level allergen-prevalence correlation with epitope burden; note this as a further reason the result is hypothesis-generating only. 4. Add one paragraph explicitly generalizing the agent-vs-human split into a stated boundary condition ("agent self-correction, in this campaign, was limited to claims verifiable by an orthogonal quantitative re-analysis; claims requiring literature breadth, disease-taxonomic judgement, or normative/equity commitments were not machine-correctable"). 5. Complete the AI-authorship and competing-interest sections before resubmission — this is achievable editorially and should not wait for a later round. **Recommendation: Minor revision.** The substantive scientific scoping the editor demanded has been done in good faith and is largely successful; what remains is placement (surface the scoping earlier), completion (denominator for the 12-finding set, disclosure sections), and one additional interpretive paragraph on the agent/human boundary that would meaningfully strengthen the piece's contribution to the agentic-science methodology literature. --- ## PAPER E — "An integrated computational pipeline for eosinophilic esophagitis" **Brief summary.** This companion research article reports the disease-discovery pipeline underlying Paper D: a 567-gene meta-signature from nine harmonized cohorts, single-cell validation, three prioritized therapeutic leads with companion diagnostics, and an antigen-directed MHC-II arm calibrated against one index patient. The revision adds heterogeneity statistics, concordance metrics for the two epitope predictors, derivation-cohort caveats attached at first mention, a GWAS enrichment disclaimer, and resolution of the DQ2.2/DRB1 gap. **Strengths** - The concordance analysis for mhcnuggets vs. netMHCIIpan-4.1 (Spearman 0.74–0.84, Jaccard 0.20–0.62) is exactly the requested test, reported with appropriate interpretation ("agree well on rank-ordering... only moderately on the hard strong-binder threshold"). - Leave-one-cohort-out robustness (399/399 surviving genes) plus explicit I² heterogeneity reporting for headline genes is a genuine statistical upgrade and handled with the right amount of humility (I²=97.6% for CCL26 flagged, not hidden). - The wheat/DQ2.2/celiac resolution is now clean and specific (tTG-IgA negative, DRB1-only scope stated as a limitation) — a real fix, not a hand-wave. **Major concerns** 1. **The GWAS "enrichment" fix substitutes disclaimer for statistics, which is a defensible but incomplete response.** The editor asked for "a formal background/enrichment test... a hypergeometric or permutation test against a matched background gene set." The revision instead declines to make an enrichment claim at all ("we do not assert statistical enrichment... because the analysis lacked a fixed, pre-specified GWAS candidate-gene universe"). This is intellectually honest and an acceptable resolution *if* the paper truly makes no enrichment-flavored claim anywhere else — but it does: the very next sentence states "the genetic support maps to epithelial-barrier susceptibility genes... rather than the downstream effectors... the expected risk-gene-versus-effector split," which is itself a distributional/enrichment-flavored claim (a claim about where within the signature genetic risk concentrates) resting on an n=7 gene list with no test. Either compute the simple hypergeometric test the editor requested (matched background = e.g., all genes on the arrays, or a matched-size random draw from the signature) or drop the "expected split" framing along with the enrichment framing — right now the paper has partially retreated from the claim's statistical form while keeping its rhetorical form. PARTIALLY RESOLVED. 2. **Validation-cohort independence is stated but the consequence is not fully carried through the Abstract.** The Abstract now says diagnostics address "≥92% of severe-endotype patients (a derivation-cohort estimate, not an independently validated one)" — good, caveat at first mention, as required. But the refractory-core claim in the Abstract ("a candidate treatment-refractory molecular core (observed in the single available paired... cohort, and therefore provisional)") is caveated, while the endotype-reproducibility claim ("resolved the disease into reproducible endotypes") is not — "reproducible" here means "recurred across six cohorts we used for both derivation and validation of endotype structure," which is the same circularity issue as the diagnostics, just left unflagged. The word "reproducible" is doing exactly the overclaiming work the editor's Essential Revision E3 was meant to eliminate everywhere it appears, not just in the two instances that got caveats. 3. **SIGLEC6 P=3×10⁻³¹ pseudoreplication concern: resolved in substance, but the "appropriate basis" language is aspirational, not executed.** The revision correctly identifies the P-value as cell-level and explains why (166,420 cells, 11 vs 12 subjects) — a real and valuable fix. But it then says "the per-patient contrast is the appropriate basis for a diagnostic cutoff" without actually reporting the per-patient-level statistic (e.g., a patient-level Mann-Whitney or mixed-effects model treating subject as random effect, analogous to what was already done for the mast-cell-fraction test in Fig 3d). Given only 11 vs 12 subjects this is trivially computable and its absence is conspicuous — the paper diagnoses the pseudoreplication problem correctly but does not fix it with the number that would let a reader judge the real diagnostic effect size. PARTIALLY RESOLVED. 4. **The leave-one-cohort-out test covers only 399/567 genes** ("sufficient multi-cohort coverage"), and the paper does not state how the 399 were selected or what happens to the excluded 168 genes' robustness. If a substantial fraction of the signature is untestable by this method, that should be quantified (e.g., what coverage threshold was used, and do the excluded genes cluster in any particular pathway or direction) rather than silently set aside. 5. **sST2 tempering is present but the Methods/Fig. 6 addressability figure still treats it identically to the other two diagnostics.** The Discussion-adjacent text correctly notes "its companion-diagnostic confidence is lower than the tissue-based readouts," but Fig. 6a's headline "100% addressability" for IL1RL1/CCL26 is presented in the same bar-chart register as SIGLEC6, with the caveat only in prose below. A visual flag (e.g., an asterisk or shaded bar) in Fig. 6 itself would match the editor's "caveat at first mention, same object as the claim" standard, which the text meets but the figure does not. **Minor concerns** 1. The DRB1-only-vs-DQ2.2-mechanistic-explanation gap is now well-handled in prose, but Table 2 / Methods should carry a one-line note that DQ predictions are out of pipeline scope, so a reader consulting the table alone isn't misled. 2. "31,512 predictions" (ten allergens × thirteen alleles) — worth a one-line sanity check in Methods that this arithmetic is peptide-count × allele-count, not allergen × allele, since the number materially exceeds simple multiplication of the two headline counts and could confuse readers auditing the pipeline. 3. Given I²=97.6% for CCL26 (the single most important target in the paper), consider a supplementary forest plot, not just the summary statistic, since this is the single figure most likely to be requested by a skeptical reader. **Feasible revisions** 1. Either compute a formal hypergeometric/permutation enrichment test for the 7-gene GWAS overlap against a matched background (this is a same-day computation on data already in hand) or strike the "expected risk-gene-versus-effector split" framing so the paper isn't making an enrichment-shaped claim while disclaiming enrichment statistics. 2. Add the missing caveat clause to "reproducible endotypes" wherever it appears in Abstract/Results, parallel to the treatment already given to the refractory-core and diagnostic-addressability claims. 3. Compute and report the patient-level (n=11 vs 12) statistic for the SIGLEC6⁺ mast-fraction comparison, not just the cell-level P-value with a caveat about what it isn't. 4. State the coverage threshold used to select the 399/567 genes for leave-one-out testing, and characterize the excluded 168 (direction, module membership) rather than leaving them uncharacterized. 5. Add a visual (not just textual) confidence marker to Fig. 6 distinguishing the sST2 diagnostic's lower maturity from the two tissue-based assays. **Recommendation: Minor revision.** The paper has made real, checkable statistical improvements on nearly every editor-flagged item (concordance stats, heterogeneity/LOCO, DQ2.2 resolution, derivation-cohort relabeling), and the remaining gaps are narrow and computable — a hypergeometric test, a patient-level P-value, and consistent caveat placement — not a reopening of the discovery pipeline. --- **Cross-cutting comment.** Both manuscripts responded to the round-1 letter with unusual candor rather than defensive minimization — the "derivation-cohort estimate, not an independently validated one" language and the explicit statement that no ledger episode touched core statistical scaffolding are the kind of self-limiting admissions that make agentic-science claims auditable rather than merely asserted. The residual pattern across both papers is the same one flagged last round: caveats are now present almost everywhere requested, but a few headline words — "reproducible," "expected... split," "self-correcting" in the title — still carry more rhetorical confidence than the adjacent, more careful prose actually supports, and a couple of requested statistics (patient-level SIGLEC6 test, GWAS background test) were diagnosed correctly but not computed. Neither paper needs new science to close these; both need one more editing/computation pass applying the same standard of rigor the authors already demonstrated everywhere else.