medsci-skills 5.19.0 → 5.20.0

package/README.md

@@ -288,6 +288,8 @@ The E2E pipeline (`orchestrate --e2e`) produces everything up to `qc/`. The `sub
 
 ## What's New
 
+**v5.20** — reviewer-arithmetic gates. Five deterministic `self-review` detectors promoted from a peer-review cycle, each recomputing what a manuscript already prints: `check_table_percentages` (an `n (%)` cell vs its column denominator), `check_nested_group_comparison` (a P value comparing an analysed subset against the parent cohort that **contains** it — nested, invalid), `check_reported_p_from_counts` (rebuilds each 2×2 row and recomputes Fisher / Pearson χ² ± Yates in **pure stdlib**, calibrating the family on rows that reproduce), `check_dta_denominators` (sensitivity/specificity denominators vs the reference-standard category counts, behind a matching grand total), and `check_paired_difference_estimator` (an odd-n integer-scale median cannot be non-integer; a zero-width CI; an unnamed estimator). Plus `/peer-review` request-type discipline (disclosure vs computation) and impossibility-claim verification. Additive and backward-compatible; **55 skills / 46 guidelines / 57 integrity detectors / 22 domain-probe modules**. (See the [CHANGELOG](CHANGELOG.md) for v5.0–v5.19.)
+
 **v5.19** — reviewer-safety + reporting-checklist batch. A **PDF hidden-text / prompt-injection guard** for `/peer-review` — a PyMuPDF extractor plus a stdlib detector that catch a review-steering instruction hidden in a submitted PDF (white-on-white text, sub-visible fonts, off-page glyphs, invisible render mode, or a document-metadata field) before an LLM ingesting the text layer can be steered by it, and emit visible-only text to feed a model instead of the raw PDF; plus the **TARGET** (target-trial emulation; Cashin et al., *JAMA* 2025) and **REMARK** (prognostic tumour-marker; McShane et al.) reporting checklists. Additive and backward-compatible; **55 skills / 46 guidelines / 52 integrity detectors / 22 domain-probe modules**. (See the [CHANGELOG](CHANGELOG.md) for v5.0–v5.18.)
 
 **v5.18** — reliability & workflow-integrity batch. A new deterministic **response-claim gate** for `/revise` and `/peer-review` (verifies a response letter's "we added / we now cite X" against the *revised manuscript body* — a claimed-but-absent edit is caught before it reaches a reviewer; **detectors 50 → 51**), a **reframe / headline-change survivor scan** (`--retired-term` / `--old-value`, finds stale framing or superseded values left in the body, supplement, and figure legends after a reframe), a pre-drafting **backbone full-text readiness gate** for `/write-paper` (the backbone article is *used*, not merely detected), a **skill-registry consistency validator** (`capabilities.yml` ⇄ `skill.yml`, CI-enforced), AI-tool **citation-framing** guidance for `/academic-aio`, and **Demo 4** (PneumoniaMNIST CNN, the model-engineering lane end to end). Additive and backward-compatible; **55 skills / 46 guidelines / 51 integrity detectors / 22 domain-probe modules**. (See the [CHANGELOG](CHANGELOG.md) for v5.0–v5.17.)

package/metadata/distribution_files.json

@@ -3048,8 +3048,8 @@
     },
     {
       "path": "skills/peer-review/SKILL.md",
-      "size": 72603,
-      "sha256": "92b89c3edc60c1d621738904df2bd4932fcb70babc60e3e306d857c781cfd6e6"
+      "size": 75710,
+      "sha256": "484a289ceb990a3301fc3a8093c1e5fae2e8fb00e267d774e4d7df4e0591a748"
     },
     {
       "path": "skills/peer-review/references/aczel_2021_reviewer2_patterns.md",
@@ -3073,8 +3073,8 @@
     },
     {
       "path": "skills/peer-review/references/domain-probes/diagnostic_accuracy.md",
-      "size": 11967,
-      "sha256": "8c0a8bd30fe16507e2997ddba1bbcfb4ee7eade7df44f1981e1427cd3a28d233"
+      "size": 12675,
+      "sha256": "06f8a8e6fcfaf00f25d04e98b58ddf3344ddc1d772057c401a83ef71d012be90"
     },
     {
       "path": "skills/peer-review/references/domain-probes/equity_fairness.md",
@@ -3118,8 +3118,8 @@
     },
     {
       "path": "skills/peer-review/references/domain-probes/observational_confounding.md",
-      "size": 34795,
-      "sha256": "ff8cb910c7fb83ec08a04ba88866bdf1e0d2f43d972e2861b825595a85b035ee"
+      "size": 35839,
+      "sha256": "32343fb2ed0d20cb76593c8d07842d0628ce363a7c24a6739f10fa29b37e6398"
     },
     {
       "path": "skills/peer-review/references/domain-probes/polygenic_risk_score.md",
@@ -3778,8 +3778,8 @@
     },
     {
       "path": "skills/self-review/references/domain-probes/diagnostic_accuracy.md",
-      "size": 11967,
-      "sha256": "8c0a8bd30fe16507e2997ddba1bbcfb4ee7eade7df44f1981e1427cd3a28d233"
+      "size": 12675,
+      "sha256": "06f8a8e6fcfaf00f25d04e98b58ddf3344ddc1d772057c401a83ef71d012be90"
     },
     {
       "path": "skills/self-review/references/domain-probes/equity_fairness.md",
@@ -3823,8 +3823,8 @@
     },
     {
       "path": "skills/self-review/references/domain-probes/observational_confounding.md",
-      "size": 34795,
-      "sha256": "ff8cb910c7fb83ec08a04ba88866bdf1e0d2f43d972e2861b825595a85b035ee"
+      "size": 35839,
+      "sha256": "32343fb2ed0d20cb76593c8d07842d0628ce363a7c24a6739f10fa29b37e6398"
     },
     {
       "path": "skills/self-review/references/domain-probes/polygenic_risk_score.md",
@@ -3961,6 +3961,41 @@
       "size": 6349,
       "sha256": "431c5f14ad2c4f59e85b2a5e86150ac708d476e8f2f1da859d81953f522b2117"
     },
+    {
+      "path": "skills/self-review/scripts/check_dta_denominators.py",
+      "size": 7798,
+      "sha256": "98c116aa7c42f045565e6e8a24df2b15751aef4a5e7e012ec7e5b5ec17a73850"
+    },
+    {
+      "path": "skills/self-review/scripts/check_dta_denominators_challenge/expected/bad.txt",
+      "size": 506,
+      "sha256": "c784ee7acd5521c7789b6893f4b86deb5c9401e1b1195dc1b77be088af2dda96"
+    },
+    {
+      "path": "skills/self-review/scripts/check_dta_denominators_challenge/expected/ok.txt",
+      "size": 119,
+      "sha256": "875e7cfd99ff0c4e2daa1270f9a5dce30597cc2e58e50c085b7c30d843d57fda"
+    },
+    {
+      "path": "skills/self-review/scripts/check_dta_denominators_challenge/fixture/dta_bad.md",
+      "size": 281,
+      "sha256": "49ed72428ff7cbcc16470fad4f7e63ef20c529fff6a0866d26003cc6cb3f75c7"
+    },
+    {
+      "path": "skills/self-review/scripts/check_dta_denominators_challenge/fixture/dta_ok.md",
+      "size": 280,
+      "sha256": "3ed363e1a00af0db022c522254788d76c3618becff8479b91e98fab52fa77760"
+    },
+    {
+      "path": "skills/self-review/scripts/check_dta_denominators_challenge/problem.md",
+      "size": 1584,
+      "sha256": "fcb419b743fa08256f3098c9a96ee7277819e8b769732d7f4763053535af61a5"
+    },
+    {
+      "path": "skills/self-review/scripts/check_dta_denominators_challenge/verify.sh",
+      "size": 1410,
+      "sha256": "3d6582b582ab5fef657dbb1e8b75a5d478669ccd88fff52e7ebe644ce5ba51f4"
+    },
     {
       "path": "skills/self-review/scripts/check_editorial_impression.py",
       "size": 22032,
@@ -3976,11 +4011,81 @@
       "size": 5298,
       "sha256": "342bc1126a420111da9c0fe38c6cd68cc45542b0e485202689d7219de9d80712"
     },
+    {
+      "path": "skills/self-review/scripts/check_nested_group_comparison.py",
+      "size": 6351,
+      "sha256": "44a7d7ce9913dcf33386de3fae510bc2f1e388c59d338ae482028920b327b37a"
+    },
+    {
+      "path": "skills/self-review/scripts/check_nested_group_comparison_challenge/expected/clean.txt",
+      "size": 104,
+      "sha256": "de4af34352139fba95d091bd65da9c34ffba15d205c1f5ebe53bf1eb23dbe01d"
+    },
+    {
+      "path": "skills/self-review/scripts/check_nested_group_comparison_challenge/expected/nested.txt",
+      "size": 301,
+      "sha256": "497f6f5ce6604a8cce134a165d72768d5bb724458912f926a5cbfbbb22243671"
+    },
+    {
+      "path": "skills/self-review/scripts/check_nested_group_comparison_challenge/fixture/clean.md",
+      "size": 328,
+      "sha256": "882880e171b845160d2f228a45d74fbab55ad9c65aa5ef0f19fc53ff5f4dbb92"
+    },
+    {
+      "path": "skills/self-review/scripts/check_nested_group_comparison_challenge/fixture/nested.md",
+      "size": 243,
+      "sha256": "40d41141e8c60a8338f8e33ac2a597ec2ee4f42fa92ce3cc94cee7be20f52b2d"
+    },
+    {
+      "path": "skills/self-review/scripts/check_nested_group_comparison_challenge/problem.md",
+      "size": 1494,
+      "sha256": "f02cff50c0c86f28ea144cb06b190420d9a02f2217d15c271157943b3f6a4984"
+    },
+    {
+      "path": "skills/self-review/scripts/check_nested_group_comparison_challenge/verify.sh",
+      "size": 1478,
+      "sha256": "c84ba913c3ba0f4d9ecc9bf84a77836513314c967491c2bb0d8666b1baf0281e"
+    },
     {
       "path": "skills/self-review/scripts/check_null_calibration.py",
       "size": 8585,
       "sha256": "e5bd71c515554b3acbaefd3807e584364a536dc6e1101bec5b524bc26b341309"
     },
+    {
+      "path": "skills/self-review/scripts/check_paired_difference_estimator.py",
+      "size": 6612,
+      "sha256": "e4a4c4c59b7dc50a8eeace2a9dfa7ba988120b4ede5f307d9836c37b40a2b76f"
+    },
+    {
+      "path": "skills/self-review/scripts/check_paired_difference_estimator_challenge/expected/bad.txt",
+      "size": 550,
+      "sha256": "3f9204146334e1f18c67997cb208d1e3702e9f2c7ed865f31f342c03ca962935"
+    },
+    {
+      "path": "skills/self-review/scripts/check_paired_difference_estimator_challenge/expected/ok.txt",
+      "size": 111,
+      "sha256": "4706dc20a9e9391a032acf5b91ce71b6ec619174ffc7e676eee4fc2cf91d4e40"
+    },
+    {
+      "path": "skills/self-review/scripts/check_paired_difference_estimator_challenge/fixture/paired_bad.md",
+      "size": 257,
+      "sha256": "a268c335cc776f26f6b366044d40d75538fba87e2730c58cc8cbeb66780f44eb"
+    },
+    {
+      "path": "skills/self-review/scripts/check_paired_difference_estimator_challenge/fixture/paired_ok.md",
+      "size": 287,
+      "sha256": "0be43e775f0e6d45162c9d0dd23e591b916992f2abd041ff71483920cb4d8a90"
+    },
+    {
+      "path": "skills/self-review/scripts/check_paired_difference_estimator_challenge/problem.md",
+      "size": 1505,
+      "sha256": "ede7ecbef797e18958abd037d3534f11b71e1027f87a3a99c310a24484da69ed"
+    },
+    {
+      "path": "skills/self-review/scripts/check_paired_difference_estimator_challenge/verify.sh",
+      "size": 1531,
+      "sha256": "b3172bbdc15868b8d4f16534e1c0049ec8b694ac864f22a989ee00aa84d218e4"
+    },
     {
       "path": "skills/self-review/scripts/check_panel_diversity.py",
       "size": 15324,
@@ -3996,6 +4101,41 @@
       "size": 17564,
       "sha256": "007100ed2ec23d01a9352feae10c5af23e7b3e88d91bb7fda7c3c899772aaa85"
     },
+    {
+      "path": "skills/self-review/scripts/check_reported_p_from_counts.py",
+      "size": 8813,
+      "sha256": "86771e8dfbad9d9680fa7cfd15cc98b63a247a318cd86c4162d852359c82a252"
+    },
+    {
+      "path": "skills/self-review/scripts/check_reported_p_from_counts_challenge/expected/bad.txt",
+      "size": 308,
+      "sha256": "b9d0da79b8eedda2c8a17ef6852400e6b7adb5e556d561cf544c8a5e5a8fb875"
+    },
+    {
+      "path": "skills/self-review/scripts/check_reported_p_from_counts_challenge/expected/ok.txt",
+      "size": 112,
+      "sha256": "cddb1acfec3c0a27e18c24755373a0bcb5521eae3d6ad6e97b06a03eabfafdbc"
+    },
+    {
+      "path": "skills/self-review/scripts/check_reported_p_from_counts_challenge/fixture/p_bad.md",
+      "size": 206,
+      "sha256": "b22b3cae304a8986a9b9ec49377601dc96f791fbbc0e04cd8adaf7c9952707dc"
+    },
+    {
+      "path": "skills/self-review/scripts/check_reported_p_from_counts_challenge/fixture/p_ok.md",
+      "size": 205,
+      "sha256": "15546bf0a9be157265dd3ec80be46d32b7fbe5386ad6e31ea6c16ea669abf3ac"
+    },
+    {
+      "path": "skills/self-review/scripts/check_reported_p_from_counts_challenge/problem.md",
+      "size": 1572,
+      "sha256": "e2df6d50acbce0da65f1af6f767a63e7007a855f40e5ef7b549d8348bd830133"
+    },
+    {
+      "path": "skills/self-review/scripts/check_reported_p_from_counts_challenge/verify.sh",
+      "size": 1509,
+      "sha256": "1a1fddcdef6d23a3c6e2f83c924180c2703355953223d5d82cd89e01615e436d"
+    },
     {
       "path": "skills/self-review/scripts/check_reviewer_team_consistency.py",
       "size": 15481,
@@ -4016,10 +4156,45 @@
       "size": 13579,
       "sha256": "c61bd5e485d2bce4a6efd7bf2a7980aa0d31ebecb876a7c6895fd12cb6bf7f1c"
     },
+    {
+      "path": "skills/self-review/scripts/check_table_percentages.py",
+      "size": 8674,
+      "sha256": "102cc795660fd41c40219b006dc657edd8d77f9195cf5b8fd2e5e470d3c28f3a"
+    },
+    {
+      "path": "skills/self-review/scripts/check_table_percentages_challenge/expected/bad.txt",
+      "size": 301,
+      "sha256": "6fed046d75f3a02818e0a3ac4aa0ad907b17eab7e05e9629467e1d54bea51856"
+    },
+    {
+      "path": "skills/self-review/scripts/check_table_percentages_challenge/expected/ok.txt",
+      "size": 112,
+      "sha256": "2fb34506bf65909fdb040acf085c4c18810b52497b6b1a53a2e08509a4465fd8"
+    },
+    {
+      "path": "skills/self-review/scripts/check_table_percentages_challenge/fixture/table_bad.md",
+      "size": 121,
+      "sha256": "841bcb7d75c8e51512c59de171d82c9959b72976eef29f631ba9440c873357d6"
+    },
+    {
+      "path": "skills/self-review/scripts/check_table_percentages_challenge/fixture/table_ok.md",
+      "size": 235,
+      "sha256": "ea5999c08b15a928606ca154c504545ed5804e45bade45ffca4d12b5849cc0bc"
+    },
+    {
+      "path": "skills/self-review/scripts/check_table_percentages_challenge/problem.md",
+      "size": 2072,
+      "sha256": "0af0e84599a8ac52a33ce33bc3244803ccc34d6f8f7f8de24234ac3b64a9a61a"
+    },
+    {
+      "path": "skills/self-review/scripts/check_table_percentages_challenge/verify.sh",
+      "size": 2041,
+      "sha256": "b4c9eb16f291841c1d6dc74a5554a49b5dc60d3562d0e328e01a78797f8d32ad"
+    },
     {
       "path": "skills/self-review/skill.yml",
-      "size": 2381,
-      "sha256": "9499a87d985d981eac9418d6cd95ed79c1d9724aba8854b7d6713d89679d9f1d"
+      "size": 3278,
+      "sha256": "84ea9430003ffc8f1b2e009600ac3b4614abb84a697ec1891a39086aa65fe404"
     },
     {
       "path": "skills/setup-medsci/SKILL.md",

package/metadata/distribution_manifest.json

@@ -1,6 +1,6 @@
 {
   "schema_version": 1,
-  "version": "5.19.0",
+  "version": "5.20.0",
   "owned_skills": [
     "academic-aio",
     "add-journal",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "medsci-skills",
-  "version": "5.19.0",
+  "version": "5.20.0",
   "description": "MedSci Skills — a medical/scientific research skill suite for AI coding agents (Claude Code, Codex, Cursor, Copilot). The npm package is a terminal-friendly installer shortcut; the canonical distribution remains the GitHub repository and the Claude Code plugin marketplace.",
   "license": "SEE LICENSE IN LICENSE",
   "homepage": "https://github.com/Aperivue/medsci-skills#readme",

package/skills/peer-review/SKILL.md

@@ -368,6 +368,13 @@ calibration, optimistic validation reporting, selective outcome reporting). Each
 it as a partner (Aczel-compliant), and calibrate severity (design-level → Major #1). Model
 the anchoring and phrasing; do not copy — they are synthetic teaching examples.
 
+**Request-type discipline (classify every Major's ask before it ships).** Sort each request into two kinds:
+
+- **Disclosure** — the study already holds the answer and has not printed it (the analysis unit; the subset's characteristics; a CI already computed; whether the model was trained on this cohort; the reading order). It costs the authors nothing to produce and *surfaces* errors; the highest-value comments are almost always this kind, including one that forces an over-claiming title to be softened.
+- **Computation** — the authors must produce a number that does not yet exist (test this difference; bootstrap a CI; give an effect size). It creates a **new, unreviewed error surface** produced under revision deadline by authors who will not re-check it and accepted next round by a reviewer who reads its existence as compliance.
+
+A computation request must carry an explicit justification that the existing tables cannot answer the question; otherwise reword it as disclosure or drop it. Prefer **naming the estimator** you want (e.g. *Hodges–Lehmann pseudomedian*) over a loose phrase (*"paired median differences"*), which authors adopt verbatim (an odd-n integer-scale "median difference" is impossible — `check_paired_difference_estimator.py`). A comment may be **both** — split it: never *request* a subset-vs-parent-cohort P value, because the groups are nested and the test is invalid (`check_nested_group_comparison.py`, and the observational/DTA domain probes); ask for the subset's characteristics (disclosure) and judge representativeness by magnitude. This is not "ask for less" — a short review with two computation requests is worse than a long one with ten disclosure requests.
+
 Generate `{manuscript_id}_review_draft.md`:
 
 ```markdown
@@ -468,7 +475,7 @@ After drafting, verify mechanically:
     Comments contain no Reject-grade language (including value-judgment deferral to the board) left
     inconsistent with a softer recommendation.
 13. **Observational-confounding QC** (if Phase 2E applied): For any covariate imbalanced by exposure in Table 1 but absent from the adjustment set (O1), confirm the comment requests a concrete extended-adjustment sensitivity model, not a vague "adjust for more confounders." Confirm a selection/collider structure (O3) or an undisclosed complete-case collapse from a structural-zero dose covariate (O5) is raised at design-level severity, and that any E-value request (O6) targets the declared primary estimate rather than a supporting one.
-14. **Verify-your-own-criticism** (all reviews): For each Major framed as a technical inaccuracy or a citation–claim mismatch, confirm the reviewer's own assertion was checked against a current authoritative source (full paper, CrossRef, arXiv). Downgrade unverified technical claims to a hedged "Please verify…"; keep confirmed ones firm. Watch for status drift (a "preprint" since published; a method since adapted) before asserting the manuscript is wrong.
+14. **Verify-your-own-criticism** (all reviews): For each Major framed as a technical inaccuracy or a citation–claim mismatch, confirm the reviewer's own assertion was checked against a current authoritative source (full paper, CrossRef, arXiv). Downgrade unverified technical claims to a hedged "Please verify…"; keep confirmed ones firm. Watch for status drift (a "preprint" since published; a method since adapted) before asserting the manuscript is wrong. **This extends to any assertion of arithmetic or statistical *impossibility* derived from the manuscript's own summary statistics** — the highest-embarrassment class, asserted with certainty and trivially falsified by the authors. Any claim containing *requires / cannot / impossible / must / contradicts* must be restated as an explicit premise→conclusion pair and stress-tested for a counterexample before submission. Two recurring traps: **a quantile or IQR statement does not constrain the tail**, and **an agreement coefficient (κ, ICC) does not constrain the marginal distribution.** Separately, for a **reviewer-requested** new statistic that appears Resolved at revision, re-derive it from the manuscript's own cells before accepting — its existence is not evidence of its correctness, and a requested analysis is the highest-prior-probability location for error in the revision (`~/.claude/rules/peer-review-response-verification.md`).
 15. **Image-synthesis QC** (if Phase 2K applied): Confirm the determinism/information-ceiling point (IS1) is raised whenever the manuscript reads a same-reader source→source+synthetic gain as added diagnostic information without a source→label baseline, that undescribed slice/mask provenance (IS2) is surfaced as a leakage/circularity concern rather than a reporting nicety, that quantitative agreement is checked at the lesion/target level not only globally (IS3), and that a biological-information claim built on image similarity alone is tempered (IS4). Per Phase 2F, confirm IS2/IS4 were treated as unfixable-in-current-form when present.
 16. **Reference-integrity QC** (all original-research reviews): Confirm the load-bearing Introduction/Discussion citations (those used as evidence the method or premise works) were spot-checked — a cited paper doing a different task, a duplicate reference, or a wrong year/author is a Minor (Major if the premise rests on it), and any unconfirmed suspicion is phrased "please verify" rather than asserted.
 
@@ -494,6 +501,8 @@ Fix all issues found, then present to user.
 - [ ] Reference-baseline ratio (`wc / 545w`) reported; ratio > 2.0 trimmed
 - [ ] Hard cap 1400 words not exceeded
 - [ ] AI pattern density within thresholds (em-dash ≤2/1000w; structural rule-of-three ≤2/Major; significance inflation 0/Major; hedged Minor ≥50%)
+- [ ] Every Major's ask classified disclosure vs computation; each computation request justified (existing tables cannot answer it) and its estimator named; no subset-vs-parent-cohort P value requested
+- [ ] Impossibility claims (requires/cannot/impossible/must/contradicts) restated as premise→conclusion + counterexample-tested; reviewer-requested new statistics re-derived from the manuscript's own cells (correctness ≠ presence)
 - [ ] Fatal flaw hierarchy stated in Confidential Comments (if applicable)
 - [ ] Reject recommendations (if used): §1C condition checklist (design-level flaw + speculative practical value 3-trigger + novelty gap) explicitly verified — at least 2 of 3 conditions met
 - [ ] AI/method/review Major Revision (or Reconsider) recommendations: Phase 2F contribution/value gate checked; weak novelty + weak utility not silently softened; for review articles, weak-novelty/no-distinct-contribution treated as unfixable-in-current-form (escalate toward Reject); unfixable defects govern tier over fixable list; confidential note carries no Reject-grade language left inconsistent with a softer recommendation

package/skills/peer-review/references/domain-probes/diagnostic_accuracy.md

@@ -17,6 +17,7 @@ A checklist for **diagnostic test accuracy (DTA) primary studies** — an index
 **D2 — Spectrum + sampling design**:
 - How were participants sampled — a **single-gate** consecutive/random series of those with the target condition suspected (preserves spectrum and prevalence), or a **two-gate** case-control design (known cases vs healthy controls)? Two-gate sampling **overestimates** accuracy and distorts spectrum; it is a MAJOR interpretive limit when the headline is clinical accuracy, not just proof-of-concept.
 - Is the study **prospective** with a pre-specified index threshold, or retrospective with a threshold chosen on the same data (D-linked to optimism)? Spectrum (disease severity, comorbid mimics) should match the intended-use population.
+- **Representativeness-of-the-analysed-subset table**: when the analysed subset (e.g. the pathology-correlated cases) is compared with the parent cohort that *contains* it, a reported P value is **invalid** — nested groups, no valid null (`check_nested_group_comparison.py`). Judge representativeness by magnitude; the valid contrast is subset vs remainder. Separately, the sensitivity a/(a+c) and specificity d/(b+d) **denominators must equal the reference-standard positive / negative counts** in the characteristics table — grand-total agreement does **not** license a per-category split mismatch, which is exactly what lets it survive review under an Abstract headline (`check_dta_denominators.py`).
 
 **D3 — Blinding of index and reference interpretation**:
 - Was the **index test** interpreted blind to the reference standard and to clinical information that would not be available at the point of use, and vice versa? Unblinded interpretation (review bias) inflates agreement.

package/skills/peer-review/references/domain-probes/observational_confounding.md

@@ -67,6 +67,7 @@ A 17-probe checklist for observational studies (cohort, case-control, cross-sect
 - When a manuscript reports an effect-size **gradient or trend across nested / overlapping cohorts or subsets** — a "comparator-refinement gradient" where the exposed arm is held fixed and only the reference group is progressively refined (Frame A ⊃ B ⊃ C; e.g. 144 → 119 → 84 events as competing-cause cases are removed from the reference) — the across-frame shift is **attributable by construction**: the reference changed, not the exposure. The estimates are not independent and there is no contrast being tested.
 - Inferential language for that gradient — "the association **attenuated significantly** across frames", "the [confounder] **largely accounted for** the effect", "a **stepwise reduction**" — is unsupported without a formal **difference / interaction test**. Two acceptable resolutions: (a) keep the gradient **descriptive** (report each frame's estimate and state that the shift follows by construction from refining the reference, not from a tested difference), or (b) fit **one model** with a frame-membership × exposure interaction (or a formally nested-model comparison) and report that interaction term.
 - This is the cohort comparator-refinement analogue of the sensitivity-envelope rule in `~/.claude/rules/scope-coherence-gate.md` (a numeric analysis introduced as if estimable/tested without the supporting test). An inferential "attenuated/accounted-for/reduced across overlapping subsets" claim with no difference/interaction test → MAJOR; downgrade to descriptive or add the interaction model.
+- **A P value comparing an analysed subset with the parent cohort that contains it is invalid, not merely uninformative** — the groups are nested (the subset's patients are *inside* the "full cohort"), so the two-group test has no valid null. This is the RC5 baseline-testing failure's observational sibling, and here it is worse than uninformative. The shape appears when authors answer a selection-bias comment with a "representativeness of the analysed subset" table (full cohort n vs subset n, with P values). Judge representativeness by the **magnitude** of the difference; if a test is genuinely wanted, the valid contrast is **subset vs remainder** (n_full − n_subset), which is disjoint. Flag any subset-vs-parent P as MAJOR (the reported P is uninterpretable). The self-review sibling automates this as a deterministic gate (`scripts/check_nested_group_comparison.py`); on the reviewer side, do not *request* such a table either — ask for the subset's characteristics (disclosure) and read representativeness off the magnitudes.
 
 **O11 — Complex-survey design & weighting (NHANES / KNHANES / CHNS and other multistage probability samples)**:
 - When the data are a **complex probability survey** (NHANES, KNHANES, CHNS, BRFSS, national health-examination surveys), unweighted or naively-weighted analysis is wrong. Check that the analysis is **design-based**: the **sampling weight** is applied AND the **stratification** (e.g. NHANES `SDMVSTRA`) and **primary sampling unit / cluster** (`SDMVPSU`) are declared so that variances use Taylor linearization or replicate weights — a weighted point estimate with **model-based (unclustered) standard errors** still understates uncertainty.

package/skills/self-review/references/domain-probes/diagnostic_accuracy.md

@@ -17,6 +17,7 @@ A checklist for **diagnostic test accuracy (DTA) primary studies** — an index
 **D2 — Spectrum + sampling design**:
 - How were participants sampled — a **single-gate** consecutive/random series of those with the target condition suspected (preserves spectrum and prevalence), or a **two-gate** case-control design (known cases vs healthy controls)? Two-gate sampling **overestimates** accuracy and distorts spectrum; it is a MAJOR interpretive limit when the headline is clinical accuracy, not just proof-of-concept.
 - Is the study **prospective** with a pre-specified index threshold, or retrospective with a threshold chosen on the same data (D-linked to optimism)? Spectrum (disease severity, comorbid mimics) should match the intended-use population.
+- **Representativeness-of-the-analysed-subset table**: when the analysed subset (e.g. the pathology-correlated cases) is compared with the parent cohort that *contains* it, a reported P value is **invalid** — nested groups, no valid null (`check_nested_group_comparison.py`). Judge representativeness by magnitude; the valid contrast is subset vs remainder. Separately, the sensitivity a/(a+c) and specificity d/(b+d) **denominators must equal the reference-standard positive / negative counts** in the characteristics table — grand-total agreement does **not** license a per-category split mismatch, which is exactly what lets it survive review under an Abstract headline (`check_dta_denominators.py`).
 
 **D3 — Blinding of index and reference interpretation**:
 - Was the **index test** interpreted blind to the reference standard and to clinical information that would not be available at the point of use, and vice versa? Unblinded interpretation (review bias) inflates agreement.

package/skills/self-review/references/domain-probes/observational_confounding.md

@@ -67,6 +67,7 @@ A 17-probe checklist for observational studies (cohort, case-control, cross-sect
 - When a manuscript reports an effect-size **gradient or trend across nested / overlapping cohorts or subsets** — a "comparator-refinement gradient" where the exposed arm is held fixed and only the reference group is progressively refined (Frame A ⊃ B ⊃ C; e.g. 144 → 119 → 84 events as competing-cause cases are removed from the reference) — the across-frame shift is **attributable by construction**: the reference changed, not the exposure. The estimates are not independent and there is no contrast being tested.
 - Inferential language for that gradient — "the association **attenuated significantly** across frames", "the [confounder] **largely accounted for** the effect", "a **stepwise reduction**" — is unsupported without a formal **difference / interaction test**. Two acceptable resolutions: (a) keep the gradient **descriptive** (report each frame's estimate and state that the shift follows by construction from refining the reference, not from a tested difference), or (b) fit **one model** with a frame-membership × exposure interaction (or a formally nested-model comparison) and report that interaction term.
 - This is the cohort comparator-refinement analogue of the sensitivity-envelope rule in `~/.claude/rules/scope-coherence-gate.md` (a numeric analysis introduced as if estimable/tested without the supporting test). An inferential "attenuated/accounted-for/reduced across overlapping subsets" claim with no difference/interaction test → MAJOR; downgrade to descriptive or add the interaction model.
+- **A P value comparing an analysed subset with the parent cohort that contains it is invalid, not merely uninformative** — the groups are nested (the subset's patients are *inside* the "full cohort"), so the two-group test has no valid null. This is the RC5 baseline-testing failure's observational sibling, and here it is worse than uninformative. The shape appears when authors answer a selection-bias comment with a "representativeness of the analysed subset" table (full cohort n vs subset n, with P values). Judge representativeness by the **magnitude** of the difference; if a test is genuinely wanted, the valid contrast is **subset vs remainder** (n_full − n_subset), which is disjoint. Flag any subset-vs-parent P as MAJOR (the reported P is uninterpretable). The self-review sibling automates this as a deterministic gate (`scripts/check_nested_group_comparison.py`); on the reviewer side, do not *request* such a table either — ask for the subset's characteristics (disclosure) and read representativeness off the magnitudes.
 
 **O11 — Complex-survey design & weighting (NHANES / KNHANES / CHNS and other multistage probability samples)**:
 - When the data are a **complex probability survey** (NHANES, KNHANES, CHNS, BRFSS, national health-examination surveys), unweighted or naively-weighted analysis is wrong. Check that the analysis is **design-based**: the **sampling weight** is applied AND the **stratification** (e.g. NHANES `SDMVSTRA`) and **primary sampling unit / cluster** (`SDMVPSU`) are declared so that variances use Taylor linearization or replicate weights — a weighted point estimate with **model-based (unclustered) standard errors** still understates uncertainty.

package/skills/self-review/scripts/check_dta_denominators.py

@@ -0,0 +1,182 @@
+#!/usr/bin/env python3
+"""check_dta_denominators.py — cross-check diagnostic-accuracy sensitivity /
+specificity denominators against the reference-standard category counts in the
+characteristics table.
+
+Sensitivity is a/(a+c) over the disease-positive patients; specificity is d/(b+d)
+over the disease-negative patients. Those denominators must equal the reference-
+standard positive / negative counts printed in the characteristics table. When
+they disagree while the grand total still matches (e.g. table 14/19, Results
+15/18, both summing to 33), a totals check passes and the split-level error — which
+sits under an Abstract headline — survives multiple review rounds. This detector
+recovers both category counts and both denominators and asserts equality per
+category, refusing to accept grand-total agreement as passing.
+
+Also `STAGE_ROWSUM`: a staging-confusion breakdown must satisfy
+correctly-staged + over-staged + under-staged == n.
+
+Runs on FIRST submissions, not only revisions — the source incident had both
+numbers on the page from submission one.
+
+Stdlib-only. Reads the manuscript, never writes it.
+
+Usage:
+  python3 check_dta_denominators.py --manuscript paper.md [--strict] [--quiet] [--json]
+Exit: 0 clean; with --strict, 1 on any mismatch; 2 on input/usage error.
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import sys
+from dataclasses import dataclass, field, asdict
+
+SEP_RE = re.compile(r"^\s*\|?[\s:|-]*-[\s:|-]*\|?\s*$")
+NEG_LABEL_RE = re.compile(r"\b(p?N\s*0|node[- ]negative|disease[- ](?:free|negative)|absent|benign|non-?malignant)\b", re.I)
+POS_LABEL_RE = re.compile(r"\b(p?N\s*[1-9]|node[- ]positive|disease[- ]positive|present|malignant|metasta\w+)\b", re.I)
+# "sensitivity ... 13/18" or "13 of 18"; bounded gap so it does not cross a
+# sentence boundary but tolerates a decimal percentage (e.g. "72.2% (13/18)").
+SENS_RE = re.compile(r"sensitivit\w*[^\n]{0,40}?\b(\d+)\s*(?:/|of)\s*(\d+)", re.I)
+SPEC_RE = re.compile(r"specificit\w*[^\n]{0,40}?\b(\d+)\s*(?:/|of)\s*(\d+)", re.I)
+CORRECT_RE = re.compile(r"correctl\w*[^.\n]*?\b(\d+)", re.I)
+OVER_RE = re.compile(r"over[- ]?stag\w*[^.\n]*?\b(\d+)", re.I)
+UNDER_RE = re.compile(r"under[- ]?stag\w*[^.\n]*?\b(\d+)", re.I)
+STAGE_N_RE = re.compile(r"\b(?:in|of|among)\s+(\d+)\s+(?:patients|exams|cases|lesions)\b", re.I)
+
+
+@dataclass
+class Finding:
+    kind: str
+    severity: str
+    detail: str
+
+
+@dataclass
+class Report:
+    source: str
+    findings: list[Finding] = field(default_factory=list)
+
+    @property
+    def n_flag(self) -> int:
+        return sum(1 for f in self.findings if f.severity == "MAJOR")
+
+    @property
+    def verdict(self) -> str:
+        return "DENOMINATOR MISMATCH" if self.n_flag else "OK"
+
+
+def _split_row(line: str) -> list[str]:
+    s = line.strip()
+    if s.startswith("|"):
+        s = s[1:]
+    if s.endswith("|"):
+        s = s[:-1]
+    return [c.strip() for c in s.split("|")]
+
+
+def _reference_counts(text: str):
+    """Return (neg_total, pos_total) summed over reference-standard rows, or None."""
+    neg = pos = 0
+    seen = False
+    lines = text.splitlines()
+    i, n = 0, len(lines)
+    while i < n:
+        if "|" in lines[i] and i + 1 < n and SEP_RE.match(lines[i + 1]) and "-" in lines[i + 1]:
+            j = i + 2
+            while j < n and "|" in lines[j] and lines[j].strip():
+                cells = _split_row(lines[j])
+                if cells:
+                    label = cells[0]
+                    cm = re.search(r"\b(\d[\d,]*)\b", " ".join(cells[1:]))
+                    if cm:
+                        cnt = int(cm.group(1).replace(",", ""))
+                        if NEG_LABEL_RE.search(label):
+                            neg += cnt; seen = True
+                        elif POS_LABEL_RE.search(label):
+                            pos += cnt; seen = True
+                j += 1
+            i = j
+        else:
+            i += 1
+    return (neg, pos) if seen else None
+
+
+def audit(text: str, source: str) -> Report:
+    rep = Report(source=source)
+    ref = _reference_counts(text)
+    sens = SENS_RE.search(text)
+    spec = SPEC_RE.search(text)
+
+    if ref and (sens or spec):
+        neg_ct, pos_ct = ref
+        sens_denom = int(sens.group(2)) if sens else None
+        spec_denom = int(spec.group(2)) if spec else None
+        if sens_denom is not None and pos_ct and sens_denom != pos_ct:
+            rep.findings.append(Finding(
+                "DTA_DENOMINATOR_MISMATCH", "MAJOR",
+                f"sensitivity denominator {sens_denom} (disease-positive) ≠ reference-standard "
+                f"positive count {pos_ct} from the characteristics table"))
+        if spec_denom is not None and neg_ct and spec_denom != neg_ct:
+            rep.findings.append(Finding(
+                "DTA_DENOMINATOR_MISMATCH", "MAJOR",
+                f"specificity denominator {spec_denom} (disease-negative) ≠ reference-standard "
+                f"negative count {neg_ct} from the characteristics table"))
+        # note when the grand totals still agree — that is what hides it
+        if rep.n_flag and sens_denom and spec_denom and neg_ct + pos_ct == sens_denom + spec_denom:
+            rep.findings.append(Finding(
+                "GRAND_TOTAL_AGREES", "INFO",
+                f"grand totals agree ({neg_ct}+{pos_ct} = {sens_denom}+{spec_denom} = "
+                f"{neg_ct + pos_ct}), so a totals-only check passes — the split-level "
+                f"mismatch is the defect"))
+
+    c, o, u = CORRECT_RE.search(text), OVER_RE.search(text), UNDER_RE.search(text)
+    sn = STAGE_N_RE.search(text)
+    if c and o and u and sn:
+        s = int(c.group(1)) + int(o.group(1)) + int(u.group(1))
+        tot = int(sn.group(1))
+        if s != tot:
+            rep.findings.append(Finding(
+                "STAGE_ROWSUM", "MAJOR",
+                f"correctly {c.group(1)} + over {o.group(1)} + under {u.group(1)} = {s} ≠ n={tot}"))
+    return rep
+
+
+def format_report(rep: Report, color: bool) -> str:
+    tag = {"OK": "\033[92m", "DENOMINATOR MISMATCH": "\033[91m"}.get(rep.verdict, "") if color else ""
+    end = "\033[0m" if color else ""
+    out = [f"{tag}== {rep.verdict} =={end}  {rep.source}", f"mismatches={rep.n_flag}"]
+    if not rep.findings:
+        out.append("sens/spec denominators reconcile with the reference-standard category counts.")
+        return "\n".join(out)
+    for f in sorted(rep.findings, key=lambda x: (x.severity != "MAJOR", x.kind, x.detail)):
+        out.append(f"[{f.severity:<5}] {f.kind}  {f.detail}")
+    return "\n".join(out)
+
+
+def main(argv: list[str] | None = None) -> int:
+    ap = argparse.ArgumentParser(description=__doc__,
+                                 formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("--manuscript", required=True, help="manuscript markdown/text")
+    ap.add_argument("--strict", action="store_true", help="exit 1 on any mismatch")
+    ap.add_argument("--quiet", action="store_true", help="suppress the report; exit code only")
+    ap.add_argument("--json", action="store_true", help="emit JSON instead of a text report")
+    args = ap.parse_args(argv)
+    try:
+        text = open(args.manuscript, encoding="utf-8").read()
+    except OSError as e:
+        print(f"error: cannot read manuscript: {e}", file=sys.stderr)
+        return 2
+    rep = audit(text, args.manuscript)
+    if not args.quiet:
+        if args.json:
+            print(json.dumps({"source": rep.source, "verdict": rep.verdict,
+                              "findings": [asdict(f) for f in rep.findings]},
+                             ensure_ascii=False, indent=2))
+        else:
+            print(format_report(rep, color=sys.stdout.isatty()))
+    return 1 if (args.strict and rep.n_flag) else 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

package/skills/self-review/scripts/check_dta_denominators_challenge/expected/bad.txt

@@ -0,0 +1,5 @@
+== DENOMINATOR MISMATCH ==  fixture/dta_bad.md
+mismatches=2
+[MAJOR] DTA_DENOMINATOR_MISMATCH  sensitivity denominator 18 (disease-positive) ≠ reference-standard positive count 19 from the characteristics table
+[MAJOR] DTA_DENOMINATOR_MISMATCH  specificity denominator 15 (disease-negative) ≠ reference-standard negative count 14 from the characteristics table
+[INFO ] GRAND_TOTAL_AGREES  grand totals agree (14+19 = 18+15 = 33), so a totals-only check passes — the split-level mismatch is the defect

package/skills/self-review/scripts/check_dta_denominators_challenge/expected/ok.txt

@@ -0,0 +1,3 @@
+== OK ==  fixture/dta_ok.md
+mismatches=0
+sens/spec denominators reconcile with the reference-standard category counts.

package/skills/self-review/scripts/check_dta_denominators_challenge/fixture/dta_bad.md

@@ -0,0 +1,12 @@
+## Table 1. Reference standard (pathological N stage)
+
+| Pathological N stage | Patients |
+| --- | --- |
+| pN0 | 14 |
+| pN1 | 13 |
+| pN2 | 6 |
+
+## Results
+
+Diagnostic performance was assessed in 33 patients. The specificity was 93.3% (14/15)
+and the sensitivity was 72.2% (13/18).

package/skills/self-review/scripts/check_dta_denominators_challenge/fixture/dta_ok.md

@@ -0,0 +1,12 @@
+## Table 1. Reference standard (pathological N stage)
+
+| Pathological N stage | Patients |
+| --- | --- |
+| pN0 | 14 |
+| pN1 | 13 |
+| pN2 | 6 |
+
+## Results
+
+Diagnostic performance was assessed in 33 patients. The specificity was 100% (14/14)
+and the sensitivity was 68.4% (13/19).

package/skills/self-review/scripts/check_dta_denominators_challenge/problem.md

@@ -0,0 +1,29 @@
+# Challenge card — DTA-denominator gate (self-review)
+
+## Problem
+Sensitivity is a/(a+c) over disease-positive patients and specificity d/(b+d) over
+disease-negative patients, so those denominators must equal the reference-standard
+positive / negative counts in the characteristics table. When they disagree while
+the grand total still matches (table 14/19, Results 15/18, both = 33), a totals-only
+check passes and the split-level error — under an Abstract headline — survives
+multiple review rounds. It is present from the first submission, not introduced by a
+revision.
+
+## What the gate does
+`check_dta_denominators.py` recovers the reference-standard negative / positive
+counts from the characteristics table (pN0 vs pN1+pN2; node-negative vs -positive;
+absent vs present) and the sensitivity / specificity denominators from the Results
+prose, and asserts each denominator equals its category count — explicitly noting
+when the grand totals agree, because that is what hides it. It also checks a staging
+breakdown sums (`STAGE_ROWSUM`: correct + over + under == n).
+
+## Fixture (synthetic only — no real manuscript, no PII)
+- `fixture/dta_bad.md` — table pN0=14 / pN1+pN2=19; Results specificity 14/15 and
+  sensitivity 13/18; grand totals both 33.
+- `fixture/dta_ok.md` — Results 14/14 and 13/19 match the table.
+
+## Expected
+- `expected/bad.txt` — two `DTA_DENOMINATOR_MISMATCH` + `GRAND_TOTAL_AGREES`; exit 1 under `--strict`.
+- `expected/ok.txt` — `OK`; exit 0.
+
+`verify.sh` diffs both outputs and asserts the exit-code contract. Network-free, stdlib-only.