medsci-skills 4.1.0

package/skills/self-review/tests/fixtures/cohort_clean.md

@@ -0,0 +1,21 @@
+# Results
+
+## Incidence
+
+During follow-up there were 120 events over 50,000 person-years, an incidence
+rate of 2.4 per 1,000 person-years.
+
+## Study population
+
+Of 4,252 participants, 583 had missing data, leaving 3,669 in the analytic cohort.
+
+## Risk by tier
+
+The cohort was partitioned into three mutually exclusive risk tiers.
+
+| Risk tier | n | events |
+| --- | --- | --- |
+| Low | 5,000 | 30 |
+| Medium | 4,000 | 50 |
+| High | 3,019 | 35 |
+| Total | 12,019 | 115 |

package/skills/self-review/tests/fixtures/cohort_partition.csv

@@ -0,0 +1,5 @@
+stratum,n,events
+Low,5000,30
+Medium,4000,50
+High,3498,45
+Total,12019,115

package/skills/self-review/tests/fixtures/coverage_analysis/31_delong_nested_added_value.csv

@@ -0,0 +1,3 @@
+model,auc,delta_auc,p
+base,0.71,,
+added,0.72,0.01,0.44

package/skills/self-review/tests/fixtures/coverage_analysis/table1_demographics.csv

@@ -0,0 +1,3 @@
+group,n
+A,10
+B,12

package/skills/self-review/tests/fixtures/coverage_clean.md

@@ -0,0 +1,13 @@
+# Methods
+
+## Statistical Analysis
+
+Missing covariate data were handled with multiple imputation (m = 20). A
+sensitivity analysis excluding events in the first year was pre-specified.
+
+# Results
+
+Baseline demographics are shown in Table 1. After multiple imputation the
+estimates were unchanged, and the sensitivity analysis excluding early events did
+not alter the conclusions. The DeLong test of the nested added-value model showed
+no incremental discrimination.

package/skills/self-review/tests/fixtures/coverage_manuscript.md

@@ -0,0 +1,11 @@
+# Methods
+
+## Statistical Analysis
+
+Missing covariate data were handled with multiple imputation (m = 20). A
+sensitivity analysis excluding events in the first year was pre-specified.
+
+# Results
+
+Baseline demographics are shown in Table 1. The sensitivity analysis excluding
+early events did not change the conclusions.

package/skills/self-review/tests/fixtures/panel_collapse.json

@@ -0,0 +1,27 @@
+[
+  {
+    "reviewer_id": "R1",
+    "expertise_area": "Methodology (search/screening/PRISMA)",
+    "major": [
+      {"heading": "Search strategy thin", "comment": "The search strategy omits two major databases and grey literature.", "location": "Methods"},
+      {"heading": "Heterogeneity not pooled correctly", "comment": "Random-effects pooling is applied despite I2 above 95%.", "location": "Results"}
+    ],
+    "minor": []
+  },
+  {
+    "reviewer_id": "R2",
+    "expertise_area": "Clinical",
+    "major": [
+      {"heading": "Clinical overclaim", "comment": "Clinical actionability is asserted beyond the evidence.", "location": "Discussion"}
+    ],
+    "minor": []
+  },
+  {
+    "reviewer_id": "R3",
+    "expertise_area": "Statistics (pooling/heterogeneity)",
+    "major": [
+      {"heading": "Heterogeneity echo", "comment": "I2 heterogeneity is high and pooling is questionable.", "location": "Results"}
+    ],
+    "minor": []
+  }
+]

package/skills/self-review/tests/fixtures/panel_good.json

@@ -0,0 +1,32 @@
+{
+  "research_type": "sr_ma",
+  "reviewers": [
+    {
+      "reviewer_id": "R1",
+      "expertise_area": "Methodology (search/screening/PRISMA)",
+      "major": [
+        {"heading": "Search strategy not reproducible", "comment": "The search strategy is described in one line with no database-specific syntax; eligibility criteria are prose rather than a numbered list.", "location": "Methods, Search"},
+        {"heading": "Duplicate record handling unclear", "comment": "Screening counts in the PRISMA flow do not reconcile with the records identified.", "location": "Figure 1"}
+      ],
+      "minor": []
+    },
+    {
+      "reviewer_id": "R2",
+      "expertise_area": "Clinical",
+      "major": [
+        {"heading": "Overreaching clinical recommendation", "comment": "Clinical actionability is asserted but generalizability to community settings is not supported.", "location": "Discussion"},
+        {"heading": "Applicability to the target population", "comment": "The included studies are tertiary-center only; clinical relevance to primary care is overstated.", "location": "Discussion"}
+      ],
+      "minor": []
+    },
+    {
+      "reviewer_id": "R3",
+      "expertise_area": "Statistics (pooling/heterogeneity)",
+      "major": [
+        {"heading": "Heterogeneity not addressed", "comment": "I2 exceeds 90% yet a fixed pooling is reported without justification.", "location": "Results"},
+        {"heading": "Pooling model mis-specified", "comment": "Random-effects model assumptions and the confidence interval method are not stated.", "location": "Methods, Synthesis"}
+      ],
+      "minor": []
+    }
+  ]
+}

package/skills/self-review/tests/fixtures/panel_monoculture.json

@@ -0,0 +1,32 @@
+{
+  "research_type": "sr_ma",
+  "reviewers": [
+    {
+      "reviewer_id": "R1",
+      "expertise_area": "Methodology (search/screening/PRISMA)",
+      "major": [
+        {"heading": "Calibration absent", "comment": "No calibration plot or Brier score is reported.", "location": "Results"},
+        {"heading": "Discrimination overstated", "comment": "AUC discrimination is reported without a confidence interval.", "location": "Results"},
+        {"heading": "Heterogeneity ignored", "comment": "I2 heterogeneity is high but not explored.", "location": "Results"}
+      ],
+      "minor": []
+    },
+    {
+      "reviewer_id": "R2",
+      "expertise_area": "Clinical",
+      "major": [
+        {"heading": "Confidence intervals missing", "comment": "Pooled estimates lack a confidence interval.", "location": "Table 2"},
+        {"heading": "Multiplicity uncorrected", "comment": "Multiplicity across many outcomes is not corrected.", "location": "Methods"}
+      ],
+      "minor": []
+    },
+    {
+      "reviewer_id": "R3",
+      "expertise_area": "Statistics (pooling/heterogeneity)",
+      "major": [
+        {"heading": "Missing data handling", "comment": "Missing data and imputation are not described.", "location": "Methods"}
+      ],
+      "minor": []
+    }
+  ]
+}

package/skills/self-review/tests/fixtures/refadeq_letter.md

@@ -0,0 +1,13 @@
+# Synthetic letter to the editor
+
+We read with interest the recent analysis of the index biomarker [@smith2019]
+and offer three observations. First, the reported effect direction is
+consistent with two prior cohorts [@jones2020; @lee2021], strengthening the
+external plausibility of the finding. Second, the authors' interpretation of the
+subgroup pattern echoes an earlier commentary [@kim2018] but appears to
+understate the confidence-interval width reported in the original registry
+[@park2022]. Third, the clinical implication should be read against the
+guideline statement [@choi2017], which cautions against acting on a single
+cross-sectional measurement [@brown2015]. We commend the authors for releasing
+their analytic code [@white2016], which permits the kind of scrutiny that
+strengthens the literature.

package/skills/self-review/tests/fixtures/refadeq_original_fixed.md

@@ -0,0 +1,42 @@
+# Synthetic original-research draft (Methods named methods now cited)
+
+## Introduction
+
+Chronic disease burden continues to rise worldwide [@smith2019]. Risk
+stratification tools have proliferated over the last decade [@jones2020], yet
+external validation remains uneven [@lee2021]. Prior cohorts established the
+prognostic value of the index biomarker [@kim2018], and registry analyses
+extended this to older adults [@park2022]. The evidence base nonetheless lacks
+competing-risk-aware modeling [@choi2017].
+
+## Methods
+
+### Study population
+
+We assembled a retrospective single-center cohort of adults meeting the
+eligibility criteria below. The analytic dataset was frozen before any modeling.
+
+### Statistical analysis
+
+We fitted a Fine-Gray subdistribution hazard model for the competing risk of
+non-event mortality [@finegray1999]. Missing covariate data were addressed with
+multiple imputation (MICE) under the fully conditional specification
+[@white2011]. The robustness of the primary association to unmeasured
+confounding was quantified with the E-value [@vanderweele2017]. Estimated
+glomerular filtration rate was computed with the CKD-EPI 2021 equation
+[@inker2021]. Discrimination was summarized with the concordance statistic
+[@harrell1996].
+
+## Results
+
+The cohort comprised consecutive patients over the study window. The primary
+model reproduced the previously reported direction of effect [@smith2019].
+
+## Discussion
+
+Our findings align with earlier prognostic work [@brown2015] and extend it to a
+competing-risk framework [@white2016]. The magnitude is consistent with two
+external cohorts [@green2018; @black2019]. Mechanistically, the association is
+plausible [@gray2020], although residual confounding cannot be excluded
+[@adams2021]. Generalizability is limited by the single-center design
+[@baker2014].

package/skills/self-review/tests/fixtures/refadeq_original_uncited.md

@@ -0,0 +1,40 @@
+# Synthetic original-research draft (Methods names methods with zero citations)
+
+## Introduction
+
+Chronic disease burden continues to rise worldwide [@smith2019]. Risk
+stratification tools have proliferated over the last decade [@jones2020], yet
+external validation remains uneven [@lee2021]. Prior cohorts established the
+prognostic value of the index biomarker [@kim2018], and registry analyses
+extended this to older adults [@park2022]. The evidence base nonetheless lacks
+competing-risk-aware modeling [@choi2017].
+
+## Methods
+
+### Study population
+
+We assembled a retrospective single-center cohort of adults meeting the
+eligibility criteria below. The analytic dataset was frozen before any modeling.
+
+### Statistical analysis
+
+We fitted a Fine-Gray subdistribution hazard model for the competing risk of
+non-event mortality. Missing covariate data were addressed with multiple
+imputation (MICE), generating twenty imputed datasets. The robustness of the
+primary association to unmeasured confounding was quantified with the E-value.
+Estimated glomerular filtration rate was computed with the CKD-EPI 2021
+equation. Discrimination was summarized with the concordance statistic.
+
+## Results
+
+The cohort comprised consecutive patients over the study window. The primary
+model reproduced the previously reported direction of effect [@smith2019].
+
+## Discussion
+
+Our findings align with earlier prognostic work [@brown2015] and extend it to a
+competing-risk framework [@white2016]. The magnitude is consistent with two
+external cohorts [@green2018; @black2019]. Mechanistically, the association is
+plausible [@gray2020], although residual confounding cannot be excluded
+[@adams2021]. Generalizability is limited by the single-center design
+[@baker2014].

package/skills/self-review/tests/fixtures/scope_bad.md

@@ -0,0 +1,9 @@
+# Methods
+
+This was a cross-sectional study of health-screening participants assessed at a
+single visit.
+
+# Conclusion
+
+Individuals with a negative scan may safely extend their screening interval, and
+routine surveillance can be deferred.

package/skills/self-review/tests/fixtures/scope_clean.md

@@ -0,0 +1,8 @@
+# Methods
+
+This was a longitudinal cohort followed prospectively for ten years.
+
+# Conclusion
+
+Higher coronary calcium scores were associated with incident cardiovascular
+events over the follow-up period.

package/skills/self-review/tests/fixtures/scope_surrogate.md

@@ -0,0 +1,8 @@
+# Methods
+
+The primary endpoint was a longitudinal cohort outcome over ten years.
+
+# Conclusion
+
+Because coronary calcium was categorized as present versus absent, a binary
+surrogate, statin therapy can be deferred in calcium-negative individuals.

package/skills/self-review/tests/fixtures/style_bad.md

@@ -0,0 +1,13 @@
+# **METHODS**
+
+Studies were eligible if they reported a diagnostic accuracy estimate and enrolled
+adult participants with the target condition.
+
+Outcome definitions are given in Methods §2.
+
+The adjusted odds ratio was 1.25 and the hazard ratio was 2.305.
+
+# **DISCUSSION**
+
+During the preparation of this manuscript, the authors used a generative AI tool to
+assist with language editing. Generative AI was not used to create any figures.

package/skills/self-review/tests/fixtures/style_clean.md

@@ -0,0 +1,11 @@
+# **METHODS**
+
+Studies were eligible if they met the following criteria: (1) reported a diagnostic
+accuracy estimate; (2) enrolled adult participants; (3) used an acceptable reference
+standard.
+
+The adjusted odds ratio was 1.25 and the hazard ratio was 2.30.
+
+# **DISCUSSION**
+
+These findings extend prior work on the target condition and support further study.

package/skills/self-review/tests/fixtures/table1_by_exposure.csv

@@ -0,0 +1,11 @@
+covariate,exposed,unexposed,p_value
+Age (years),54.2,53.8,0.21
+"Sex, male (%)",61.0,59.4,0.34
+BMI (kg/m2),25.1,24.8,0.08
+Hypertension (%),30.0,28.0,0.12
+Diabetes (%),12.0,11.0,0.40
+Uric acid (mg/dL),6.1,5.4,0.0007
+"Smoking, pack-years",18.0,9.0,0.001
+HDL (mg/dL),48.0,55.0,<0.001
+Total cholesterol (mg/dL),195,188,0.010
+HbA1c (%),5.9,5.6,0.0009

package/skills/self-review/tests/test_artifact_coverage.sh

@@ -0,0 +1,44 @@
+#!/usr/bin/env bash
+# Regression test for the artifact-coverage gate (self-review Phase 2.5f).
+# Synthetic, PII-free fixtures reproduce: (a) a Methods-promised multiple-
+# imputation analysis that never reaches Results (FORWARD), (b) an analysis-bearing
+# output CSV (a DeLong nested added-value table) present on disk but unmentioned in
+# the manuscript (REVERSE). The clean manuscript reports both and mentions the disk
+# outputs, so it reconciles.
+# Stdlib-only (python3).
+set -u
+
+HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+SCRIPT="$HERE/../scripts/check_artifact_coverage.py"
+BAD="$HERE/fixtures/coverage_manuscript.md"
+CLEAN="$HERE/fixtures/coverage_clean.md"
+ADIR="$HERE/fixtures/coverage_analysis"
+OUT="$(mktemp -t cov_XXXX).json"
+trap 'rm -f "$OUT"' EXIT
+
+fail=0
+check() { local label="$1"; shift
+    if "$@" >/dev/null 2>&1; then printf '  PASS  %s\n' "$label"
+    else printf '  FAIL  %s\n' "$label"; fail=$((fail+1)); fi
+}
+has_verdict() { python3 -c "
+import json,sys
+d=json.load(open('$OUT'))
+assert any(c['verdict']=='$1' for c in d['claims']), '$1 not found'
+"; }
+
+[[ -f "$SCRIPT" ]] || { echo "ENV-ERR: script missing" >&2; exit 2; }
+
+# (1) bad manuscript + analysis dir: promised-absent + disk-unreported, Major -> exit 1
+python3 "$SCRIPT" --manuscript "$BAD" --analysis-dir "$ADIR" --out "$OUT" --strict --quiet >/dev/null 2>&1
+check "exit 1 under --strict (Major present)" test "$?" -eq 1
+check "JSON artifact written" test -s "$OUT"
+check "PROMISED_ABSENT detected (MI promised, absent from Results)" has_verdict PROMISED_ABSENT
+check "DISK_UNREPORTED detected (delong nested CSV unmentioned)"    has_verdict DISK_UNREPORTED
+
+# (2) clean manuscript: reports MI + sensitivity, mentions disk outputs -> exit 0
+python3 "$SCRIPT" --manuscript "$CLEAN" --analysis-dir "$ADIR" --strict --quiet >/dev/null 2>&1
+check "exit 0 on clean manuscript (all reconciled)" test "$?" -eq 0
+
+echo "fail=$fail"; [[ "$fail" -eq 0 ]] && echo "ALL PASS" || echo "FAILURES: $fail"
+exit "$fail"

package/skills/self-review/tests/test_claim_artifact.sh

@@ -0,0 +1,50 @@
+#!/usr/bin/env bash
+# Regression test for the claim-vs-artifact cross-check (self-review Phase 2.5f).
+# Synthetic fixture reproduces: (a) a primary re-designated at manuscript stage,
+# (b) an E-value (2.79) that does not recompute from its stated primary HR 1.34,
+# (c) a correctly-arithmetic E-value attached to a non-primary (cancer) estimate.
+# Stdlib-only (python3).
+set -u
+
+HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+SCRIPT="$HERE/../scripts/check_claim_artifact.py"
+MAN="$HERE/fixtures/claim_manuscript.md"
+PRE="$HERE/fixtures/claim_prereg.md"
+OUT="$(mktemp -t ca_XXXX).json"
+trap 'rm -f "$OUT"' EXIT
+
+fail=0
+check() { local label="$1"; shift
+    if "$@" >/dev/null 2>&1; then printf '  PASS  %s\n' "$label"
+    else printf '  FAIL  %s\n' "$label"; fail=$((fail+1)); fi
+}
+has_verdict() { python3 -c "
+import json,sys
+d=json.load(open('$OUT'))
+assert any(c['verdict']=='$1' for c in d['claims']), '$1 not found'
+"; }
+
+[[ -f "$SCRIPT" ]] || { echo "ENV-ERR: script missing" >&2; exit 2; }
+
+python3 "$SCRIPT" --manuscript "$MAN" --prereg "$PRE" --out "$OUT" --strict >/dev/null 2>&1
+check "exit 1 under --strict (Major present)" test "$?" -eq 1
+check "JSON artifact written" test -s "$OUT"
+check "PRIMARY_REASSIGNED detected"  has_verdict PRIMARY_REASSIGNED
+check "EVALUE_ARITHMETIC detected (2.79 vs HR 1.34)" has_verdict EVALUE_ARITHMETIC
+check "EVALUE_NON_PRIMARY detected (cancer sHR)"     has_verdict EVALUE_NON_PRIMARY
+
+# Clean case: primary matches prereg, correct primary E-value, no reassignment.
+CLEAN="$(mktemp -t ca_clean_XXXX).md"
+trap 'rm -f "$OUT" "$CLEAN"' EXIT
+cat > "$CLEAN" <<'EOF'
+## Methods
+The primary analysis was the association between emphysema and all-cause mortality
+in the complete-case multivariable Cox model.
+## Results
+The E-value for the primary association (HR 1.34) was 2.02.
+EOF
+python3 "$SCRIPT" --manuscript "$CLEAN" --prereg "$PRE" --strict >/dev/null 2>&1
+check "exit 0 on clean manuscript (matching primary, correct E-value)" test "$?" -eq 0
+
+echo "fail=$fail"; [[ "$fail" -eq 0 ]] && echo "ALL PASS" || echo "FAILURES: $fail"
+exit "$fail"

package/skills/self-review/tests/test_classical_style.sh

@@ -0,0 +1,44 @@
+#!/usr/bin/env bash
+# Regression test for the classical-style body lint (self-review §J).
+# Synthetic, PII-free fixtures reproduce: a § symbol self-reference + an in-body
+# AI-disclosure paragraph (both Major), eligibility prose, and mixed OR/HR decimals
+# (Minor). The clean fixture uses a numbered eligibility list, consistent decimals,
+# no § symbol, and no in-body disclosure.
+# Stdlib-only (python3).
+set -u
+
+HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+SCRIPT="$HERE/../scripts/check_classical_style.py"
+BAD="$HERE/fixtures/style_bad.md"
+CLEAN="$HERE/fixtures/style_clean.md"
+OUT="$(mktemp -t style_XXXX).json"
+trap 'rm -f "$OUT"' EXIT
+
+fail=0
+check() { local label="$1"; shift
+    if "$@" >/dev/null 2>&1; then printf '  PASS  %s\n' "$label"
+    else printf '  FAIL  %s\n' "$label"; fail=$((fail+1)); fi
+}
+has_verdict() { python3 -c "
+import json,sys
+d=json.load(open('$OUT'))
+assert any(c['verdict']=='$1' for c in d['claims']), '$1 not found'
+"; }
+
+[[ -f "$SCRIPT" ]] || { echo "ENV-ERR: script missing" >&2; exit 2; }
+
+# (1) bad manuscript: Major (§ + in-body disclosure) -> exit 1
+python3 "$SCRIPT" --manuscript "$BAD" --out "$OUT" --strict --quiet >/dev/null 2>&1
+check "exit 1 under --strict (Major present)" test "$?" -eq 1
+check "JSON artifact written" test -s "$OUT"
+check "SECTION_SYMBOL detected"        has_verdict SECTION_SYMBOL
+check "INBODY_AI_DISCLOSURE detected"  has_verdict INBODY_AI_DISCLOSURE
+check "ELIGIBILITY_PROSE detected"     has_verdict ELIGIBILITY_PROSE
+check "DECIMAL_INCONSISTENCY detected" has_verdict DECIMAL_INCONSISTENCY
+
+# (2) clean manuscript: numbered eligibility, consistent decimals, no §/disclosure -> exit 0
+python3 "$SCRIPT" --manuscript "$CLEAN" --strict --quiet >/dev/null 2>&1
+check "exit 0 on clean manuscript" test "$?" -eq 0
+
+echo "fail=$fail"; [[ "$fail" -eq 0 ]] && echo "ALL PASS" || echo "FAILURES: $fail"
+exit "$fail"

package/skills/self-review/tests/test_cohort_arithmetic.sh

@@ -0,0 +1,49 @@
+#!/usr/bin/env bash
+# Regression test for the cohort-arithmetic gate (self-review Phase 2.5 / 2.5b).
+# Synthetic, PII-free fixtures reproduce: (a) an incidence rate that does not
+# recompute from its events/person-years, (b) a complete-case footnote that does
+# not balance (total - missing != complete), (c) an ordinal tier partition whose
+# stratum denominators and events sum above the stated total.
+# Stdlib-only (python3).
+set -u
+
+HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+SCRIPT="$HERE/../scripts/check_cohort_arithmetic.py"
+BAD="$HERE/fixtures/cohort_bad.md"
+CLEAN="$HERE/fixtures/cohort_clean.md"
+PART="$HERE/fixtures/cohort_partition.csv"
+OUT="$(mktemp -t coh_XXXX).json"
+trap 'rm -f "$OUT"' EXIT
+
+fail=0
+check() { local label="$1"; shift
+    if "$@" >/dev/null 2>&1; then printf '  PASS  %s\n' "$label"
+    else printf '  FAIL  %s\n' "$label"; fail=$((fail+1)); fi
+}
+has_verdict() { python3 -c "
+import json,sys
+d=json.load(open('$OUT'))
+assert any(c['verdict']=='$1' for c in d['claims']), '$1 not found'
+"; }
+
+[[ -f "$SCRIPT" ]] || { echo "ENV-ERR: script missing" >&2; exit 2; }
+
+# (1) bad manuscript: all three discrepancies, Major present -> exit 1 under --strict
+python3 "$SCRIPT" --manuscript "$BAD" --out "$OUT" --strict --quiet >/dev/null 2>&1
+check "exit 1 under --strict (Major present)" test "$?" -eq 1
+check "JSON artifact written" test -s "$OUT"
+check "RATE_BACKCALC detected (5.0 vs 120/50000*1000=2.4)" has_verdict RATE_BACKCALC
+check "CASCADE_SUM detected (4252-583=3669 != 3667)"       has_verdict CASCADE_SUM
+check "PARTITION_OVERLAP detected (sum 12498 != 12019)"     has_verdict PARTITION_OVERLAP
+
+# (2) clean manuscript: everything balances -> exit 0
+python3 "$SCRIPT" --manuscript "$CLEAN" --strict --quiet >/dev/null 2>&1
+check "exit 0 on clean manuscript (all arithmetic balances)" test "$?" -eq 0
+
+# (3) --data CSV partition: stratum n sum above total -> PARTITION_OVERLAP, exit 1
+python3 "$SCRIPT" --manuscript "$CLEAN" --data "$PART" --out "$OUT" --strict --quiet >/dev/null 2>&1
+check "exit 1 with overlapping --data partition CSV" test "$?" -eq 1
+check "PARTITION_OVERLAP from --data CSV" has_verdict PARTITION_OVERLAP
+
+echo "fail=$fail"; [[ "$fail" -eq 0 ]] && echo "ALL PASS" || echo "FAILURES: $fail"
+exit "$fail"

package/skills/self-review/tests/test_confounding_completeness.sh

@@ -0,0 +1,66 @@
+#!/usr/bin/env bash
+# Regression test for the confounding-completeness gate (self-review Phase 2.5e).
+# Synthetic, PII-free fixture mirroring the canonical failure pattern: five
+# measured covariates imbalanced by exposure (uric acid, pack-years, HDL, total
+# cholesterol, HbA1c) that are absent from an age/sex/BMI/HTN/DM adjustment set.
+# Stdlib-only (python3); no pandas/statsmodels.
+set -u
+
+HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+SCRIPT="$HERE/../scripts/check_confounding_completeness.py"
+FIXTURE="$HERE/fixtures/table1_by_exposure.csv"
+OUT="$(mktemp -t cc_XXXX).json"
+trap 'rm -f "$OUT"' EXIT
+
+fail=0
+check() {  # check "label" expr...
+    local label="$1"; shift
+    if "$@" >/dev/null 2>&1; then printf '  PASS  %s\n' "$label"
+    else printf '  FAIL  %s\n' "$label"; fail=$((fail+1)); fi
+}
+
+[[ -f "$SCRIPT" ]]  || { echo "ENV-ERR: script missing"  >&2; exit 2; }
+[[ -f "$FIXTURE" ]] || { echo "ENV-ERR: fixture missing" >&2; exit 2; }
+
+# 1. Runs and writes JSON.
+python3 "$SCRIPT" --table1 "$FIXTURE" \
+    --adjusted-list "age, sex, BMI, hypertension, diabetes" \
+    --out "$OUT" --strict >/dev/null 2>&1
+rc=$?
+check "exit 1 under --strict (unadjusted-imbalanced present)" test "$rc" -eq 1
+check "JSON artifact written" test -s "$OUT"
+
+# 2. Exactly five unadjusted-imbalanced covariates.
+n=$(python3 -c "import json,sys; print(json.load(open('$OUT'))['n_unadjusted_imbalanced'])" 2>/dev/null)
+check "n_unadjusted_imbalanced == 5" test "$n" = "5"
+
+# 3. Each expected offender flagged; adjusted/balanced ones not.
+for cov in "Uric acid" "Smoking, pack-years" "HDL" "Total cholesterol" "HbA1c"; do
+    check "flagged: $cov" python3 -c "
+import json
+d=json.load(open('$OUT'))
+assert any('$cov'.lower() in f['covariate'].lower() and f['verdict']=='UNADJUSTED_IMBALANCED' for f in d['findings'])
+"
+done
+for cov in "Age" "BMI"; do
+    check "not flagged: $cov" python3 -c "
+import json
+d=json.load(open('$OUT'))
+assert not any('$cov'.lower() in f['covariate'].lower() and f['verdict']=='UNADJUSTED_IMBALANCED' for f in d['findings'])
+"
+done
+
+# 4. '<0.001' p-value cell parsed as imbalanced (HDL row uses '<0.001').
+check "verdict MAJOR_CANDIDATE" python3 -c "
+import json; assert json.load(open('$OUT'))['verdict']=='MAJOR_CANDIDATE'
+"
+
+# 5. Clean case: adjusting for everything yields exit 0.
+python3 "$SCRIPT" --table1 "$FIXTURE" \
+    --adjusted-list "age, sex, BMI, hypertension, diabetes, uric acid, smoking, HDL, total cholesterol, HbA1c" \
+    --strict >/dev/null 2>&1
+check "exit 0 when all imbalanced covariates adjusted" test "$?" -eq 0
+
+echo "ran=$(( ${fail} + 0 )) fail=$fail"
+[[ "$fail" -eq 0 ]] && echo "ALL PASS" || echo "FAILURES: $fail"
+exit "$fail"

package/skills/self-review/tests/test_panel_diversity.sh

@@ -0,0 +1,55 @@
+#!/usr/bin/env bash
+# Regression test for the panel lens-diversity gate (self-review Phase 2.6).
+# Synthetic, PII-free reviewer-output fixtures reproduce: (a) a healthy panel
+# spanning three distinct axes (clean), (b) a monoculture where every major
+# falls in one family AND two expected axes are uncovered (UNCOVERED_AXIS +
+# FAMILY_MONOCULTURE), (c) a fully-redundant reviewer lens (LENS_COLLAPSE).
+# Stdlib-only (python3).
+set -u
+
+HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+SCRIPT="$HERE/../scripts/check_panel_diversity.py"
+GOOD="$HERE/fixtures/panel_good.json"
+MONO="$HERE/fixtures/panel_monoculture.json"
+COLLAPSE="$HERE/fixtures/panel_collapse.json"
+OUT="$(mktemp -t paneldiv_XXXX).json"
+trap 'rm -f "$OUT"' EXIT
+
+fail=0
+check() { local label="$1"; shift
+    if "$@" >/dev/null 2>&1; then printf '  PASS  %s\n' "$label"
+    else printf '  FAIL  %s\n' "$label"; fail=$((fail+1)); fi
+}
+has_verdict() { python3 -c "
+import json,sys
+d=json.load(open('$OUT'))
+assert any(c['verdict']=='$1' for c in d['claims']), '$1 not found'
+"; }
+no_verdict() { python3 -c "
+import json,sys
+d=json.load(open('$OUT'))
+assert all(c['verdict']!='$1' for c in d['claims']), '$1 unexpectedly found'
+"; }
+
+[[ -f "$SCRIPT" ]] || { echo "ENV-ERR: script missing" >&2; exit 2; }
+
+# (1) healthy diverse panel (sr_ma) -> exit 0, no Major claims
+python3 "$SCRIPT" --panel "$GOOD" --out "$OUT" --strict --quiet >/dev/null 2>&1
+check "exit 0 (diverse panel)" test "$?" -eq 0
+check "no UNCOVERED_AXIS on clean panel" no_verdict UNCOVERED_AXIS
+check "no FAMILY_MONOCULTURE on clean panel" no_verdict FAMILY_MONOCULTURE
+
+# (2) monoculture + uncovered axes -> exit 1 with both Major verdicts
+python3 "$SCRIPT" --panel "$MONO" --out "$OUT" --strict --quiet >/dev/null 2>&1
+check "exit 1 (monoculture)" test "$?" -eq 1
+check "UNCOVERED_AXIS detected" has_verdict UNCOVERED_AXIS
+check "FAMILY_MONOCULTURE detected" has_verdict FAMILY_MONOCULTURE
+
+# (3) fully-redundant reviewer (no research type) -> LENS_COLLAPSE flag, exit 0
+python3 "$SCRIPT" --panel "$COLLAPSE" --out "$OUT" --strict --quiet >/dev/null 2>&1
+check "exit 0 (collapse is flag-only)" test "$?" -eq 0
+check "LENS_COLLAPSE detected" has_verdict LENS_COLLAPSE
+check "no UNCOVERED_AXIS without research type" no_verdict UNCOVERED_AXIS
+
+echo "fail=$fail"; [[ "$fail" -eq 0 ]] && echo "ALL PASS" || echo "FAILURES: $fail"
+exit "$fail"