dsh-math-modeling-agent 0.4.1 → 0.5.0

package/skills/math-modeling-agent/scripts/verification-recipes.mjs

@@ -0,0 +1,52 @@
+#!/usr/bin/env node
+const RECIPES = Object.freeze({
+  data: Object.freeze(['schema', 'ids', 'references', 'hash', 'independent-parser']),
+  numeric: Object.freeze(['recompute', 'domain', 'strict', 'tolerance', 'counterexample']),
+  feasibility: Object.freeze(['all-constraints', 'actual-output', 'strict-and-tolerance']),
+  optimization: Object.freeze(['objective-scope', 'coverage', 'initialization', 'local-global-label']),
+  simulation: Object.freeze(['baseline', 'bounds', 'sampling', 'independent-implementation']),
+  report: Object.freeze(['nine-sections', 'why-how-what-check', 'coverage', 'symbols', 'citations']),
+  reproducibility: Object.freeze(['snapshot', 'command', 'environment', 'hash', 'clean-rerun']),
+  'branch-coverage': Object.freeze(['candidate-enumeration', 'branch-switch']),
+})
+
+function recipeFor(recipeId) {
+  const requiredChecks = RECIPES[recipeId]
+  if (!requiredChecks) throw new TypeError('unknown verification recipe: ' + recipeId)
+  return { id: recipeId, requiredChecks: [...requiredChecks] }
+}
+
+function nonIndependent(value) {
+  const text = [value?.independenceKey, value?.derivedFrom, value?.methodFamily].filter(Boolean).join(' ').toLowerCase()
+  return ['same summary', 'solver summary', 'deterministic result', 'same result', 'reprint'].some(marker => text.includes(marker))
+}
+
+function countIndependent(verifications) {
+  if (!Array.isArray(verifications)) throw new TypeError('verifications must be an array')
+  const keys = new Set()
+  const rejected = []
+  for (const [index, verification] of verifications.entries()) {
+    const key = typeof verification?.independenceKey === 'string' ? verification.independenceKey.trim() : ''
+    if (!key) { rejected.push({ index, reason: 'missing independenceKey' }); continue }
+    if (nonIndependent(verification)) { rejected.push({ index, reason: 'summary/deterministic/reprint provenance is not independent' }); continue }
+    if (keys.has(key)) { rejected.push({ index, reason: 'duplicate independenceKey' }); continue }
+    keys.add(key)
+  }
+  return { count: keys.size, keys: [...keys], rejected }
+}
+
+function runRecipe(recipeId, input = {}) {
+  const recipe = recipeFor(recipeId)
+  if (recipeId === 'branch-coverage') {
+    const declared = Number.isInteger(input.declaredCandidates) ? input.declaredCandidates : 0
+    const executed = new Set(Array.isArray(input.executedCandidates) ? input.executedCandidates : [])
+    if (declared > executed.size) throw new Error('branch-switch-not-tested')
+  }
+  const checks = input.checks ?? {}
+  const missing = recipe.requiredChecks.filter(check => !checks[check] || !['PASS', 'FAIL', 'INCONCLUSIVE'].includes(checks[check].verdict))
+  const failed = recipe.requiredChecks.filter(check => checks[check]?.verdict === 'FAIL')
+  const verdict = failed.length > 0 ? 'FAIL' : (missing.length > 0 ? 'INCONCLUSIVE' : 'PASS')
+  return { recipeId, verdict, missing, failed, checks }
+}
+
+export { RECIPES, countIndependent, recipeFor, runRecipe }

package/skills/math-modeling-agent/scripts/verification-runner.mjs

@@ -0,0 +1,8 @@
+#!/usr/bin/env node
+export async function runWithBenchmark({ recipeId, expensive = false, benchmark, run }) {
+  if (typeof run !== 'function') throw new TypeError('run must be a function')
+  if (expensive && (!benchmark || benchmark.verdict !== 'PASS')) throw new Error('benchmark evidence is required before an expensive recipe')
+  const startedAt = Date.now()
+  const result = await run()
+  return { recipeId, benchmark: benchmark ?? null, result, elapsedMs: Date.now() - startedAt }
+}

package/skills/math-modeling-audit/SKILL.md

@@ -1,41 +1,41 @@
----
-name: math-modeling-audit
-description: This skill should be used when the user asks to "独立核验数学模型", "检查推导", "寻找反例", "审计建模报告", "按 MCM/ICM 终审框架打分", "verify this model", or needs an artifact-only adversarial audit of mathematical claims, data, code, results, citations, or a competition paper.
-whenToUse: Use for existing artifacts and papers; do not take over open-ended model construction or silently rewrite the audited work.
-user-invocable: true
----
-
-# Math Modeling Audit
-
-## Goal
-
-Independently decide which claims in an existing artifact pass, fail, or remain inconclusive, and why.
-
-## Interface
-
-Accept a paper, model, derivation, code/result artifact, or MathModelingAgent run. Optionally accept the original problem, competition year/code, target claims, and assurance level.
-
-Return per-claim verdicts, evidence levels, counterexamples, reproducibility findings, residual risk, and—only for MCM/ICM judging intent—the fixed final-panel report.
-
-## Workflow
-
-1. Freeze the artifact set and record review coverage.
-2. Reconstruct claims, assumptions, obligations, and evidence without trusting the author’s summary.
-3. Apply `references/verification-protocol.md` and `references/evidence-levels.md`.
-4. Audit data, parameters, leakage, citations, and reproducibility with `references/data-citation-audit.md`.
-5. Recompute applicable formulas and results through independent tools.
-6. Seek boundary cases, counterexamples, alternative explanations, and simpler models.
-7. Output PASS, FAIL, or INCONCLUSIVE per claim; do not silently repair the source.
-8. For MCM/ICM scoring intent only, load `references/mcm-icm-final-judge.md`, validate arithmetic with `scripts/mcm-score.mjs`, and preserve its fourteen-section output.
-
-## Invariants
-
-- Paper prose and solver success are not verification artifacts.
-- Missing evidence is reported, never supplied on the author’s behalf.
-- Lean verifies the formal statement, not its natural-language fidelity or data pipeline.
-- A score never feeds back into the solver’s SOLVED gate.
-- Disqualification risk stops ordinary award scoring.
-
-## Boundaries
-
-Do not edit the audited artifact. Do not reward complexity or presentation without evidence. Do not load the large MCM/ICM rubric for generic audits.
+---
+name: math-modeling-audit
+description: This skill should be used when the user asks to "独立核验数学模型", "检查推导", "寻找反例", "审计建模报告", "按 MCM/ICM 终审框架打分", "verify this model", or needs an artifact-only adversarial audit of mathematical claims, data, code, results, citations, or a competition paper.
+whenToUse: Use for existing artifacts and papers; do not take over open-ended model construction or silently rewrite the audited work.
+user-invocable: true
+---
+
+# Math Modeling Audit
+
+## Goal
+
+Independently decide which claims in an existing artifact pass, fail, or remain inconclusive, and why.
+
+## Interface
+
+Accept a paper, model, derivation, code/result artifact, or MathModelingAgent run. Optionally accept the original problem, competition year/code, target claims, and assurance level.
+
+Return per-claim verdicts, evidence levels, counterexamples, reproducibility findings, residual risk, and—only for MCM/ICM judging intent—the fixed final-panel report.
+
+## Workflow
+
+1. Freeze the artifact set and record review coverage.
+2. Reconstruct claims, assumptions, obligations, and evidence without trusting the author’s summary.
+3. Apply `references/verification-protocol.md` and `references/evidence-levels.md`.
+4. Audit data, parameters, leakage, citations, and reproducibility with `references/data-citation-audit.md`.
+5. Recompute applicable formulas and results through independent tools.
+6. Seek boundary cases, counterexamples, alternative explanations, and simpler models.
+7. Output PASS, FAIL, or INCONCLUSIVE per claim; do not silently repair the source.
+8. For project initial review, evaluate the current run and paper-evidence readiness. For complete-paper final review, freeze the manuscript and evidence graph, load references/mcm-icm-final-judge.md, validate arithmetic with scripts/mcm-score.mjs, and preserve its fourteen-section output. Findings can request a correction lineage but never mutate the audited parent.
+
+## Invariants
+
+- Paper prose and solver success are not verification artifacts.
+- Missing evidence is reported, never supplied on the author’s behalf.
+- Lean verifies the formal statement, not its natural-language fidelity or data pipeline.
+- A score never feeds back into the solver’s SOLVED gate.
+- Disqualification risk stops ordinary award scoring.
+
+## Boundaries
+
+Do not edit the audited artifact. Do not reward complexity or presentation without evidence. Do not load the large MCM/ICM rubric for generic audits.