incremnt 0.8.0 → 0.8.2

package/src/summary-evals.js

@@ -6,6 +6,7 @@ import {
   canonicalExerciseName,
   checkpointContext,
   cycleSummaryContext,
+  dateOnlyString,
   executeCoachReadTool,
   normalizeExerciseName,
   workoutSummaryContext,
@@ -24,12 +25,30 @@ import {
 import { computeScoreBand } from './score-context.js';
 import { stripXMLTagBlocks } from './prompt-security.js';
 import { extractAskProgramDraft, hasProgramDraftBlock } from './program-draft.js';
+import { findAskAnswerExerciseMentions, verifyAskAnswer } from './ask-answer-verifier.js';
 
 const __filename = fileURLToPath(import.meta.url);
 const __dirname = path.dirname(__filename);
 
 export const summaryEvalFixturesRoot = path.resolve(__dirname, '../test/fixtures/summary-evals');
 
+const ASK_PROVENANCE_VERIFIER_KEYS = new Set([
+  'unsupported_weight_claim',
+  'unrouted_weight_claim',
+  'unsupported_weighted_set_claim',
+  'unrouted_weighted_set_claim',
+  'e1rm_without_records',
+  'direction_inversion',
+  'tool_replay_failed'
+]);
+
+const ASK_CLAIM_VERIFIER_KEYS = new Set([
+  'target_hit_contradiction',
+  'target_hit_without_session_evidence',
+  'clean_consistency_contradiction',
+  'unsupported_fatigue_recovery'
+]);
+
 export function defaultCaseSetName() {
   return process.env.SUMMARY_EVAL_CASE_SET || 'synthetic';
 }
@@ -108,7 +127,8 @@ export function buildSummaryEvalContext(snapshot, testCase) {
       // prelude to the routed context. Including it here means a live eval feeds
       // the model the same dump-prone material, so evaluateAskScoreVoice actually
       // guards the prompt, not just the checker.
-      const prelude = formatIncrementScorePrelude(scoreHistoryFromSnapshot(snapshot), { question });
+      const responseProfile = routed?.metadata?.responseProfile ?? routed?.metadata?.intent?.responseProfile ?? testCase.context?.responseProfile;
+      const prelude = formatIncrementScorePrelude(scoreHistoryFromSnapshot(snapshot), { question, responseProfile });
       const routedContext = routed?.context ?? null;
       const trainingData = testCase.context?.trainingData
         ?? (prelude && routedContext ? `${prelude}\n\n${routedContext}` : (routedContext ?? prelude));
@@ -222,6 +242,49 @@ function normalizeText(value) {
   return String(value ?? '').trim();
 }
 
+const askVerifierCacheKey = Symbol('summaryEvalAskVerifierCache');
+
+export function summaryEvalAskVerifierToday(testCase) {
+  return dateOnlyString(testCase.context?.today ?? testCase.today ?? new Date());
+}
+
+function getAskVerifierResult(output, context, snapshot, testCase) {
+  const answer = normalizeText(output);
+  const routingMetadata = context?.routedMetadata ?? testCase.context?.routedMetadata ?? {};
+  const today = summaryEvalAskVerifierToday(testCase);
+  const exclude = testCase.exclude ?? [];
+  const cacheKey = `${answer}\n${JSON.stringify(routingMetadata)}\n${today}\n${JSON.stringify(exclude)}`;
+  if (context && typeof context === 'object') {
+    if (!context[askVerifierCacheKey]) {
+      Object.defineProperty(context, askVerifierCacheKey, {
+        value: new Map(),
+        enumerable: false
+      });
+    }
+    const cached = context[askVerifierCacheKey].get(cacheKey);
+    if (cached) return cached;
+  }
+  const verification = verifyAskAnswer({
+    answer,
+    snapshot,
+    routingMetadata,
+    today,
+    exclude,
+    strictMentionProvenance: false
+  });
+  const result = {
+    verification,
+    provenanceFailures: (verification.blockingFailures ?? [])
+      .filter((failure) => ASK_PROVENANCE_VERIFIER_KEYS.has(failure.key)),
+    claimFailures: (verification.blockingFailures ?? [])
+      .filter((failure) => ASK_CLAIM_VERIFIER_KEYS.has(failure.key))
+  };
+  if (context && typeof context === 'object') {
+    context[askVerifierCacheKey].set(cacheKey, result);
+  }
+  return result;
+}
+
 function parseJsonOutput(output) {
   const normalized = normalizeText(output);
   if (!normalized) return null;
@@ -697,7 +760,7 @@ function evaluateNoInsight(output, testCase) {
   };
 }
 
-function evaluateShape(output, testCase) {
+function evaluateShape(output, testCase, context = null) {
   const normalized = testCase.surface === 'scoreCommentary'
     ? scoreCommentaryText(output)
     : normalizeText(output);
@@ -768,12 +831,20 @@ function evaluateShape(output, testCase) {
         reasons.push(`Checkpoint summaries must be 2-3 paragraphs, got ${paragraphs}.`);
       }
       break;
-    case 'ask':
-      if (sentences < 1 || sentences > 12) {
+    case 'ask': {
+      // Expansive answers are intentionally richer; the old 12-sentence cap was
+      // the pre-expansive policy. Allow more for expansive (still bounded so a
+      // genuine wall of text is flagged), keep the tight cap for defensive.
+      const profile = context?.routedMetadata?.responseProfile
+        ?? context?.routedMetadata?.intent?.responseProfile
+        ?? askResponseProfileFromTestCase(testCase);
+      const maxAskSentences = profile === 'expansive' ? 20 : 12;
+      if (sentences < 1 || sentences > maxAskSentences) {
         passed = false;
-        reasons.push(`Ask-coach answers must be 1-12 sentences, got ${sentences}.`);
+        reasons.push(`Ask-coach answers must be 1-${maxAskSentences} sentences, got ${sentences}.`);
       }
       break;
+    }
     case 'scoreCommentary':
       if (sentences < 1 || sentences > 8) {
         passed = false;
@@ -888,45 +959,6 @@ function hasFatigueLanguage(output) {
   return /\b(fatigue|fatigued|underrecovered|recovery debt|fatigue ceiling|limited by recovery|limited by fatigue|accumulated fatigue)\b/i.test(output);
 }
 
-function hasAskFatigueRecoveryLanguage(output) {
-  return hasFatigueLanguage(output)
-    || /\b(?:poor|low|bad|incomplete)\s+recovery\b/i.test(output)
-    || /\bunder[-\s]?recovery\b/i.test(output)
-    || /\brecovery\s+(?:limited|held back|caused|explains|drove|deficit|issue|problem)\b/i.test(output);
-}
-
-function hasAskFatigueRecoveryUncertaintyLanguage(output) {
-  const missingRecoveryData = /\b(?:no|not enough|without|missing|lack(?:ing)?|insufficient)\s+(?:\w+\s+){0,4}?(?:recovery|readiness|vitals?|sleep|hrv|heart rate|data|info|signals?|metrics?)\b/i.test(output);
-  const refusesInference = /\b(?:cannot|can't|do not|don't|does not|doesn't|would not|wouldn't|not enough|isn't enough|is not enough|no basis to|hard to)\s+(?:\w+\s+){0,12}?(?:infer|tie|connect|attribute|blame|claim|say|show|prove|know|call)\s+(?:\w+\s+){0,12}?(?:fatigue|recovery|readiness|why)\b/i.test(output);
-  const recoveryDoesNotExplain = /\b(?:fatigue|recovery|readiness)\b\s+(?:\w+\s+){0,10}?(?:cannot|can't|does not|doesn't|would not|wouldn't|isn't|is not)\s+(?:\w+\s+){0,10}?(?:explain|prove|show|tell|account for)\b/i.test(output);
-  return missingRecoveryData || refusesInference || recoveryDoesNotExplain;
-}
-
-function hasAskPositiveFatigueRecoveryAttribution(output) {
-  const concept = String.raw`(?:fatigue|fatigued|under[-\s]?recovered|under[-\s]?recovery|poor recovery|low recovery|incomplete recovery|recovery debt|fatigue ceiling|accumulated fatigue)`;
-  const causeVerb = String.raw`(?:because|due to|caused by|from|reflects?|suggests?|indicates?|points? to|explains?|limited|held back|drove|contributed to|tied to|tie\s+\w+\s+to)`;
-  const patterns = [
-    new RegExp(String.raw`\b${causeVerb}\b.{0,80}\b${concept}\b`, 'gi'),
-    new RegExp(String.raw`\b${concept}\b.{0,80}\b(?:caused|limited|held back|explains?|drove|led to|contributed to|accounts? for)\b`, 'gi')
-  ];
-  for (const pattern of patterns) {
-    for (const match of output.matchAll(pattern)) {
-      const start = Math.max(0, (match.index ?? 0) - 40);
-      const window = output.slice(start, (match.index ?? 0) + match[0].length);
-      if (!/\b(?:not|no|cannot|can't|doesn't|does not|would not|wouldn't|isn't|is not)\b/i.test(window)) {
-        return true;
-      }
-    }
-  }
-  return false;
-}
-
-function hasUnsupportedAskFatigueRecoveryClaim(output) {
-  if (!hasAskFatigueRecoveryLanguage(output)) return false;
-  if (hasAskPositiveFatigueRecoveryAttribution(output)) return true;
-  return !hasAskFatigueRecoveryUncertaintyLanguage(output);
-}
-
 function matchesHistoricalFamilyName(claimName, actualName) {
   const claimVariants = new Set(historicalExerciseVariants(claimName));
   const actualVariants = new Set(historicalExerciseVariants(actualName));
@@ -1086,37 +1118,6 @@ function evaluateWorkoutClaims(output, context, testCase) {
   };
 }
 
-function extractAskTargetHitClaims(text) {
-  const claims = [];
-  const patterns = [
-    /\b(?:you\s+)?hit(?:ting)?\s+all\s+(?:your\s+)?target(?:ed)?\s+reps?\b/gi,
-    /\b(?:you\s+)?hit\s+all\s+(?:the\s+)?targets?\b/gi,
-    /\b(?:you\s+)?hit\s+(?:the|your)\s+target\b(?!\s+(?:of|for|on))/gi
-  ];
-  for (const pattern of patterns) {
-    for (const match of text.matchAll(pattern)) {
-      claims.push({ text: match[0] });
-    }
-  }
-  return claims;
-}
-
-function extractAskCleanConsistencyClaims(text) {
-  const claims = [];
-  const patterns = [
-    /\bclean,\s+consistent\b/gi,
-    /\bclean\s+and\s+consistent\b/gi,
-    /\bconsistent\s+set\s+of\s+work\b/gi,
-    /\bacross\s+the\s+board\b/gi
-  ];
-  for (const pattern of patterns) {
-    for (const match of text.matchAll(pattern)) {
-      claims.push({ text: match[0] });
-    }
-  }
-  return claims;
-}
-
 function extractAskPlannedListClaims(text) {
   const claims = [];
   const pattern = /\((\s*\d+(?:\s*,\s*\d+){2,})\s+planned\s*\)/gi;
@@ -1137,77 +1138,6 @@ function sessionPlannedReps(session) {
   return values;
 }
 
-function findMentionedExercises(text, snapshot) {
-  const exercisesByName = new Map();
-  for (const session of snapshot?.sessions ?? []) {
-    for (const exercise of session.exercises ?? []) {
-      if (!exercise?.name) continue;
-      const normalizedName = normalizeExerciseName(exercise.name);
-      if (!normalizedName || exercisesByName.has(normalizedName)) continue;
-      exercisesByName.set(normalizedName, exercise.name);
-    }
-    for (const exercise of session.prescriptionSnapshot?.exercises ?? []) {
-      if (!exercise?.exerciseName) continue;
-      const normalizedName = normalizeExerciseName(exercise.exerciseName);
-      if (!normalizedName || exercisesByName.has(normalizedName)) continue;
-      exercisesByName.set(normalizedName, exercise.exerciseName);
-    }
-  }
-
-  const mentions = [];
-  for (const [normalizedName, displayName] of exercisesByName) {
-    const pattern = new RegExp(`\\b${escapeRegExp(displayName)}\\b`, 'gi');
-    for (const match of text.matchAll(pattern)) {
-      mentions.push({
-        index: match.index ?? -1,
-        end: (match.index ?? -1) + match[0].length,
-        name: displayName,
-        normalizedName
-      });
-    }
-  }
-  return mentions
-    .filter((mention, index, allMentions) => !allMentions.some((candidate, candidateIndex) =>
-      candidateIndex !== index &&
-      candidate.index <= mention.index &&
-      candidate.end >= mention.end &&
-      candidate.normalizedName.length > mention.normalizedName.length
-    ))
-    .sort((lhs, rhs) => lhs.index - rhs.index);
-}
-
-function findRecentSessionMisses(snapshot, { lookbackDays = 7, exerciseNames = null } = {}) {
-  const sessions = snapshot?.sessions ?? [];
-  const cutoff = Date.now() - lookbackDays * 24 * 60 * 60 * 1000;
-  const scopedExerciseNames = exerciseNames && exerciseNames.length > 0 ? new Set(exerciseNames) : null;
-  const misses = [];
-  for (const session of sessions) {
-    const completedAt = session.completedAt || session.date;
-    const completedTime = Date.parse(completedAt);
-    if (!Number.isFinite(completedTime) || completedTime < cutoff) continue;
-    const targetByExercise = new Map();
-    for (const planned of session.prescriptionSnapshot?.exercises ?? []) {
-      const target = Number(planned.targetReps);
-      if (Number.isFinite(target) && target > 0) {
-        targetByExercise.set(normalizeExerciseName(planned.exerciseName), target);
-      }
-    }
-    for (const exercise of session.exercises ?? []) {
-      const normalizedExerciseName = normalizeExerciseName(exercise.name);
-      if (scopedExerciseNames && !scopedExerciseNames.has(normalizedExerciseName)) continue;
-      const target = targetByExercise.get(normalizedExerciseName);
-      if (!Number.isFinite(target)) continue;
-      for (const set of exercise.sets ?? []) {
-        const reps = Number(set.reps);
-        if (set.isComplete && Number.isFinite(reps) && reps < target) {
-          misses.push({ sessionId: session.id, exerciseName: exercise.name, reps, target });
-        }
-      }
-    }
-  }
-  return misses;
-}
-
 function findNearestMentionedExercise(mentions, index) {
   let candidate = null;
   for (const mention of mentions) {
@@ -1217,170 +1147,6 @@ function findNearestMentionedExercise(mentions, index) {
   return candidate;
 }
 
-function hasAskFatigueSupport(snapshot, lookbackDays = 7) {
-  const cutoff = Date.now() - lookbackDays * 24 * 60 * 60 * 1000;
-  const withinCutoff = (dateValue) => {
-    const ms = Date.parse(dateValue);
-    return Number.isFinite(ms) && ms >= cutoff;
-  };
-
-  const vitalsSummaries = snapshot?.vitalsSummaries ?? [];
-  if (vitalsSummaries.some((entry) => withinCutoff(entry.date))) return true;
-
-  const metrics = snapshot?.healthMetrics ?? {};
-  for (const key of ['restingHR', 'hrv', 'sleep']) {
-    const readings = Array.isArray(metrics[key]) ? metrics[key] : [];
-    if (readings.some((reading) => withinCutoff(reading.date))) return true;
-  }
-
-  for (const session of snapshot?.sessions ?? []) {
-    const completedAt = session.completedAt || session.date;
-    if (!withinCutoff(completedAt)) continue;
-    for (const exercise of session.exercises ?? []) {
-      const reps = (exercise.sets ?? [])
-        .map((set) => Number(set.reps))
-        .filter((value) => Number.isFinite(value) && value > 0);
-      if (reps.length < 2) continue;
-      const first = reps[0];
-      const last = reps[reps.length - 1];
-      if (first > 0 && (first - last) / first >= 0.3) return true;
-    }
-  }
-
-  return false;
-}
-
-function parseWeightNumber(raw) {
-  return Number(String(raw).replace(/,/g, ''));
-}
-
-function extractAskWeightClaims(text) {
-  const claims = [];
-  // Accept comma-grouped thousands ("40,500 kg") as a single number so volume
-  // figures are not shredded into bogus "500 kg" / "000 kg" claims. Volume/total
-  // figures are excluded by isVolumeWeightClaim at the call sites, not by a
-  // magnitude cap — heavy machine work (leg press, sled) legitimately exceeds
-  // 1000 kg, and a fabricated heavy load must still be graded.
-  const pattern = /\b(\d{1,3}(?:,\d{3})+(?:\.\d+)?|\d+(?:\.\d+)?)\s*(?:kg|kilograms?)\b/gi;
-  for (const match of text.matchAll(pattern)) {
-    claims.push({
-      text: match[0],
-      value: parseWeightNumber(match[1]),
-      index: match.index ?? -1
-    });
-  }
-  return claims;
-}
-
-function extractAskWeightedSetClaims(text) {
-  const claims = [];
-  // A weight×reps pair is only unambiguous with "x"/"×" (e.g. "70 kg x 5"), or
-  // an explicit "for N rep(s)". Bare "X kg for N" is NOT a rep claim — N is
-  // almost always a SET count ("70 kg for 4 working sets") or a duration, and
-  // treating it as reps flags real data as a fabricated pair. So match only the
-  // unambiguous forms; the plain-weight loop still grounds the weight itself.
-  const pattern = /\b(\d{1,3}(?:,\d{3})+(?:\.\d+)?|\d+(?:\.\d+)?)\s*(?:kg|kilograms?)\s*(?:(?:x|×)\s*(\d+)|for\s+(\d+)\s*reps?)\b/gi;
-  for (const match of text.matchAll(pattern)) {
-    claims.push({
-      text: match[0],
-      weight: parseWeightNumber(match[1]),
-      reps: Number(match[2] ?? match[3]),
-      index: match.index ?? -1,
-      end: (match.index ?? -1) + match[0].length
-    });
-  }
-  return claims;
-}
-
-function allowedWeightsForExercise(snapshot, normalizedExerciseName) {
-  const weights = [];
-  for (const session of snapshot?.sessions ?? []) {
-    for (const exercise of session.exercises ?? []) {
-      if (normalizeExerciseName(exercise.name) !== normalizedExerciseName) continue;
-      for (const set of exercise.sets ?? []) {
-        const weight = Number(set.weight);
-        if (Number.isFinite(weight)) weights.push(weight);
-      }
-    }
-    for (const exercise of session.prescriptionSnapshot?.exercises ?? []) {
-      if (normalizeExerciseName(exercise.exerciseName) !== normalizedExerciseName) continue;
-      const targetWeight = Number(exercise.targetWeight);
-      if (Number.isFinite(targetWeight)) weights.push(targetWeight);
-      for (const targetSet of exercise.targetSets ?? []) {
-        const weight = Number(targetSet.weight ?? targetSet.targetWeight);
-        if (Number.isFinite(weight)) weights.push(weight);
-      }
-    }
-  }
-  return weights;
-}
-
-function weightClaimSupported(claim, allowedWeights) {
-  return allowedWeights.some((weight) => Math.abs(weight - claim.value) < 0.01);
-}
-
-function isEstimatedOneRepMaxWeightClaim(text, claim) {
-  const start = Math.max(0, claim.index - 40);
-  const end = Math.min(text.length, claim.index + claim.text.length + 40);
-  const window = text.slice(start, end);
-  return /\b(?:estimated\s+)?(?:1rm|one[-\s]?rep\s+max)\b/i.test(window);
-}
-
-function isVolumeWeightClaim(text, claim) {
-  // A kg figure in a clause about volume/tonnage/total load is a workload total
-  // (e.g. "weekly strength volume fell from 44,000 kg to 40,500 kg"), not an
-  // exercise load. Scope to the claim's clause so a fabricated exercise load
-  // earlier in the same sentence is still graded.
-  return /\b(?:volume|tonnage|total\s+(?:load|work|volume|tonnage))\b/i.test(claimClause(text, claim));
-}
-
-function claimClause(text, claim) {
-  const boundaries = [
-    '\n',
-    '. ',
-    ';',
-    ', while',
-    ', whereas',
-    ', but',
-    ' while ',
-    ' whereas ',
-    ' but '
-  ];
-  let start = 0;
-  for (const boundary of boundaries) {
-    const index = text.lastIndexOf(boundary, claim.index);
-    if (index >= 0) start = Math.max(start, index + boundary.length);
-  }
-
-  let end = text.length;
-  for (const boundary of boundaries) {
-    const index = text.indexOf(boundary, claim.index + claim.text.length);
-    if (index >= 0) end = Math.min(end, index);
-  }
-  return text.slice(start, end);
-}
-
-// Returns the sentence containing the claim, so context guards can look at the
-// whole clause rather than a fixed-width window (body-weight phrasing can put
-// the "body weight" anchor well before the kg figure).
-function claimSentence(text, claim) {
-  const before = text.slice(0, claim.index);
-  const startBreak = Math.max(before.lastIndexOf('. '), before.lastIndexOf('\n'));
-  const start = startBreak >= 0 ? startBreak + 1 : 0;
-  const after = text.slice(claim.index);
-  const breaks = [after.indexOf('. '), after.indexOf('\n')].filter((i) => i >= 0);
-  const end = breaks.length ? claim.index + Math.min(...breaks) : text.length;
-  return text.slice(start, end);
-}
-
-// Body-weight figures ("body weight is up 0.6 kg", "80.0 kg latest") are not
-// exercise-load claims. findNearestMentionedExercise would otherwise attribute
-// them to the previously named lift and flag a correct answer as a
-// hallucination, so skip any kg figure stated in a body-weight clause.
-function isBodyWeightClaim(text, claim) {
-  return /\bbody\s*weight\b|\bbodyweight\b|\bweight\s+trend\b/i.test(claimSentence(text, claim));
-}
-
 function askWorkingTopSetRows(snapshot) {
   const rows = [];
   for (const session of stableSortByDateDesc(snapshot?.sessions ?? [], (session) => session.completedAt ?? session.date)) {
@@ -1529,7 +1295,7 @@ const SCORE_NUMBER = '\\d{1,3}(?:\\.\\d+)?';
 // number directly followed by one of these is left alone.
 const NON_SCORE_UNIT =
   '(?:kg|kilo|lbs?|pounds?|reps?|sets?|%|percent|pct|x\\b|for\\s+\\d|sessions?|days?|nights?|weeks?|months?|' +
-  'years?|yrs?|hrs?|hours?|mins?|minutes?|secs?|seconds?|rpe|rir|am|pm|out\\s+of|of\\b|/\\s*\\d)';
+  'years?|yrs?|h\\b|hrs?|hours?|mins?|minutes?|secs?|seconds?|bpm|ms|rpe|rir|am|pm|out\\s+of|of\\b|/\\s*\\d)';
 
 // Heuristic, not a parser. Flags a component name followed — within a short,
 // period/newline-free gap (one clause) — by a score-like number that is not a
@@ -1632,11 +1398,19 @@ function evaluateAskSelfReference(output, testCase) {
 // On a question that is not about the Increment Score, the coach must not
 // volunteer the bare overall score number (e.g. "your score is 92/100"). The
 // prelude withholds the number for non-score questions; this guards the answer.
-function evaluateAskVolunteeredScore(output, testCase) {
+function evaluateAskVolunteeredScore(output, testCase, context = {}) {
   if (testCase.surface !== 'ask') {
     return { key: 'ask_volunteered_score', passed: true, reason: 'Not an ask answer.' };
   }
   const question = testCase.context?.question ?? testCase.question ?? '';
+  const responseProfile = context?.routedMetadata?.responseProfile
+    ?? context?.routedMetadata?.intent?.responseProfile
+    ?? testCase.context?.routedMetadata?.responseProfile
+    ?? testCase.context?.routedMetadata?.intent?.responseProfile
+    ?? testCase.context?.responseProfile;
+  if (responseProfile === 'expansive') {
+    return { key: 'ask_volunteered_score', passed: true, reason: 'Expansive Ask answers may name the rounded score headline.' };
+  }
   if (isScoreQuestion(question)) {
     return { key: 'ask_volunteered_score', passed: true, reason: 'Question is about the score; naming it is allowed.' };
   }
@@ -1682,7 +1456,14 @@ function evaluateAskStaleness(output, snapshot, testCase) {
   };
 }
 
-function evaluateAskClaims(output, snapshot, testCase) {
+function askResponseProfileFromTestCase(testCase) {
+  return testCase?.context?.routedMetadata?.responseProfile
+    ?? testCase?.context?.routedMetadata?.intent?.responseProfile
+    ?? testCase?.context?.responseProfile
+    ?? null;
+}
+
+function evaluateAskClaims(output, snapshot, testCase, context = null) {
   if (testCase.surface !== 'ask') {
     return { key: 'ask_claims', passed: true, reason: 'Not an ask answer.' };
   }
@@ -1693,26 +1474,9 @@ function evaluateAskClaims(output, snapshot, testCase) {
   }
 
   const failures = [];
-  const mentionedExercises = findMentionedExercises(output, snapshot);
-  const scopedExerciseNames = uniqueStrings(mentionedExercises.map((mention) => mention.normalizedName));
-
-  const targetHitClaims = extractAskTargetHitClaims(normalized);
-  if (targetHitClaims.length > 0) {
-    const misses = findRecentSessionMisses(snapshot, { exerciseNames: scopedExerciseNames });
-    if (misses.length > 0) {
-      const sample = misses[0];
-      failures.push(`Ask answer claims targets hit ("${targetHitClaims[0].text}") but recent session ${sample.sessionId} has ${sample.exerciseName} at ${sample.reps} reps below target ${sample.target}.`);
-    }
-  }
-
-  const cleanConsistencyClaims = extractAskCleanConsistencyClaims(normalized);
-  if (cleanConsistencyClaims.length > 0) {
-    const misses = findRecentSessionMisses(snapshot, { exerciseNames: scopedExerciseNames });
-    if (misses.length > 0) {
-      const sample = misses[0];
-      failures.push(`Ask answer frames missed target reps as "${cleanConsistencyClaims[0].text}", but recent session ${sample.sessionId} has ${sample.exerciseName} at ${sample.reps} reps below target ${sample.target}.`);
-    }
-  }
+  const { claimFailures } = getAskVerifierResult(output, context, snapshot, testCase);
+  failures.push(...claimFailures.map((failure) => failure.reason));
+  const mentionedExercises = findAskAnswerExerciseMentions(output, snapshot);
 
   for (const claim of extractAskPlannedListClaims(normalized)) {
     const uniquePlanned = new Set(claim.reps);
@@ -1742,23 +1506,6 @@ function evaluateAskClaims(output, snapshot, testCase) {
     }
   }
 
-  for (const claim of extractAskWeightClaims(normalized)) {
-    if (isEstimatedOneRepMaxWeightClaim(normalized, claim)) continue;
-    if (isVolumeWeightClaim(normalized, claim)) continue;
-    if (isBodyWeightClaim(normalized, claim)) continue;
-    const referencedExercise = findNearestMentionedExercise(mentionedExercises, claim.index);
-    if (!referencedExercise) continue;
-    const allowedWeights = allowedWeightsForExercise(snapshot, referencedExercise.normalizedName);
-    if (allowedWeights.length === 0) continue;
-    if (!weightClaimSupported(claim, allowedWeights)) {
-      failures.push(`Ask answer asserts ${claim.text} for ${referencedExercise.name}, but that weight is not present in recorded or planned sets for that exercise.`);
-    }
-  }
-
-  if (hasUnsupportedAskFatigueRecoveryClaim(normalized) && !hasAskFatigueSupport(snapshot)) {
-    failures.push('Ask answer uses fatigue/recovery language but the snapshot has no recent vitals, sleep, or rep-dropoff signals to support it.');
-  }
-
   return {
     key: 'ask_claims',
     passed: failures.length === 0,
@@ -1826,55 +1573,6 @@ function askToolEvidenceRows(toolResults = []) {
   return rows;
 }
 
-function askToolEvidenceWeights(rows = []) {
-  const weights = [];
-  for (const row of rows) {
-    for (const set of row.sets ?? []) {
-      const weight = Number(set.weight);
-      if (Number.isFinite(weight)) weights.push(weight);
-    }
-    const topWeight = Number(row.topSet?.weight);
-    if (Number.isFinite(topWeight)) weights.push(topWeight);
-    const previousTopWeight = Number(row.comparedToPreviousSession?.previousTopSet?.weight);
-    if (Number.isFinite(previousTopWeight)) weights.push(previousTopWeight);
-  }
-  return weights;
-}
-
-function askToolEvidenceSetPairs(rows = []) {
-  const pairs = [];
-  for (const row of rows) {
-    for (const set of row.sets ?? []) {
-      const weight = Number(set.weight);
-      const reps = Number(set.reps);
-      if (Number.isFinite(weight) && Number.isFinite(reps)) pairs.push({ weight, reps });
-    }
-    const topWeight = Number(row.topSet?.weight);
-    const topReps = Number(row.topSet?.reps);
-    if (Number.isFinite(topWeight) && Number.isFinite(topReps)) pairs.push({ weight: topWeight, reps: topReps });
-    const previousTopWeight = Number(row.comparedToPreviousSession?.previousTopSet?.weight);
-    const previousTopReps = Number(row.comparedToPreviousSession?.previousTopSet?.reps);
-    if (Number.isFinite(previousTopWeight) && Number.isFinite(previousTopReps)) {
-      pairs.push({ weight: previousTopWeight, reps: previousTopReps });
-    }
-  }
-  return pairs;
-}
-
-function toolEvidenceSupportsWeightClaim(claim, rows) {
-  if (weightClaimSupported(claim, askToolEvidenceWeights(rows))) return true;
-  return false;
-}
-
-function toolEvidenceSupportsWeightedSetClaim(claim, rows) {
-  if (askToolEvidenceSetPairs(rows).some((pair) => (
-    Math.abs(pair.weight - claim.weight) < 0.01 && pair.reps === claim.reps
-  ))) {
-    return true;
-  }
-  return false;
-}
-
 function compareToolEvidenceRecency(lhs, rhs) {
   const lhsDaysAgo = Number(lhs?.daysAgo);
   const rhsDaysAgo = Number(rhs?.daysAgo);
@@ -1890,14 +1588,6 @@ function newestToolEvidenceRow(rows = [], predicate = () => true) {
     .sort(compareToolEvidenceRecency)[0] ?? null;
 }
 
-function latestComparableToolRow(rows = []) {
-  return newestToolEvidenceRow(rows, (row) => row.comparedToPreviousSession?.loadDirection) ?? null;
-}
-
-function isWithinWeightedSetClaim(claim, weightedSetClaims) {
-  return weightedSetClaims.some((setClaim) => claim.index >= setClaim.index && claim.index < setClaim.end);
-}
-
 function rowIsStaleForEval(row, testCase) {
   const daysAgo = Number(row?.daysAgo);
   const cutoff = Number(testCase.staleness?.maxRecentDays ?? row?.recencyCutoffDays);
@@ -1952,80 +1642,22 @@ function evaluateAskToolProvenance(output, context, testCase, snapshot) {
 
   const routedMetadata = context?.routedMetadata ?? {};
   const toolsUsed = new Set(routedMetadata.toolsUsed ?? []);
-  const { toolResults, replayFailures } = routedToolResultsForEval(snapshot, context);
+  const { toolResults } = routedToolResultsForEval(snapshot, context);
   const evidenceRows = askToolEvidenceRows(toolResults);
-  const mentionedExercises = findMentionedExercises(output, snapshot);
-  const unroutedMentionNames = new Set();
-  const failures = [...replayFailures];
+  const failures = [];
   for (const toolName of uniqueStrings(testCase.requiredTools)) {
     if (!toolsUsed.has(toolName)) {
       failures.push(`Expected routed Ask Coach context to use ${toolName}.`);
     }
   }
 
-  if (/\b(?:estimated\s+)?(?:e1rm|1rm|one[- ]rep max)\b/i.test(output) && !toolsUsed.has('get_records')) {
-    failures.push('Ask answer mentions e1RM/1RM, but routed context did not use get_records.');
-  }
-
-  const weightedSetClaims = extractAskWeightedSetClaims(output);
-  for (const claim of weightedSetClaims) {
-    if (isEstimatedOneRepMaxWeightClaim(output, claim)) continue;
-    const referencedExercise = findNearestMentionedExercise(mentionedExercises, claim.index);
-    if (!referencedExercise) continue;
-    const rows = evidenceRows.filter((row) => row.normalizedName === referencedExercise.normalizedName);
-    if (rows.length === 0) {
-      unroutedMentionNames.add(referencedExercise.normalizedName);
-      failures.push(`Ask answer asserts ${claim.text} for ${referencedExercise.name}, but ${referencedExercise.name} was not present in routed tool outputs.`);
-      continue;
-    }
-    if (!toolEvidenceSupportsWeightedSetClaim(claim, rows)) {
-      failures.push(`Ask answer asserts ${claim.text} for ${referencedExercise.name}, but routed tool outputs for ${referencedExercise.name} did not include that weight/reps pair.`);
-    }
-  }
-
-  for (const claim of extractAskWeightClaims(output)) {
-    if (isWithinWeightedSetClaim(claim, weightedSetClaims)) continue;
-    if (isEstimatedOneRepMaxWeightClaim(output, claim)) continue;
-    if (isVolumeWeightClaim(output, claim)) continue;
-    if (isBodyWeightClaim(output, claim)) continue;
-    const referencedExercise = findNearestMentionedExercise(mentionedExercises, claim.index);
-    if (!referencedExercise) continue;
-    const rows = evidenceRows.filter((row) => row.normalizedName === referencedExercise.normalizedName);
-    if (rows.length === 0) {
-      unroutedMentionNames.add(referencedExercise.normalizedName);
-      failures.push(`Ask answer asserts ${claim.text} for ${referencedExercise.name}, but ${referencedExercise.name} was not present in routed tool outputs.`);
-      continue;
-    }
-    if (!toolEvidenceSupportsWeightClaim(claim, rows)) {
-      failures.push(`Ask answer asserts ${claim.text} for ${referencedExercise.name}, but routed tool outputs for ${referencedExercise.name} did not include that weight.`);
-    }
-  }
+  const { provenanceFailures } = getAskVerifierResult(output, context, snapshot, testCase);
+  failures.push(...provenanceFailures.map((failure) => failure.reason));
 
   const exerciseNames = evidenceRows.map((row) => row.exerciseName);
-  for (const mention of mentionedExercises) {
+  for (const mention of findAskAnswerExerciseMentions(output, snapshot)) {
     const rows = evidenceRows.filter((row) => row.normalizedName === mention.normalizedName);
-    if (rows.length === 0) {
-      if (toolResults.length > 0 && !unroutedMentionNames.has(mention.normalizedName)) {
-        unroutedMentionNames.add(mention.normalizedName);
-        failures.push(`Ask answer mentions ${mention.name}, but ${mention.name} was not present in routed tool outputs.`);
-      }
-      continue;
-    }
-    const comparable = latestComparableToolRow(rows);
-    if (comparable) {
-      const direction = comparable.comparedToPreviousSession.loadDirection;
-      const previous = comparable.comparedToPreviousSession.previousTopSet;
-      const windows = directionEvaluationWindows(output, mention.name, exerciseNames);
-      if (direction === 'up' && windows.some(hasUnqualifiedDeclineLanguage)) {
-        failures.push(`Ask answer frames ${mention.name} as declining/drop-off, but routed ${comparable.toolName} evidence says top load increased from ${previous.weight} kg to ${comparable.topSet.weight} kg.`);
-      }
-      if (direction === 'down' && windows.some(hasUnqualifiedImprovementLanguage)) {
-        failures.push(`Ask answer frames ${mention.name} as improving, but routed ${comparable.toolName} evidence says top load decreased from ${previous.weight} kg to ${comparable.topSet.weight} kg.`);
-      }
-      if (direction === 'flat' && windows.some((window) => hasUnqualifiedDeclineLanguage(window) || hasUnqualifiedImprovementLanguage(window))) {
-        failures.push(`Ask answer invents a load direction for ${mention.name}, but routed ${comparable.toolName} evidence says top load was flat at ${comparable.topSet.weight} kg.`);
-      }
-    }
+    if (rows.length === 0) continue;
 
     const latestDatedRow = newestToolEvidenceRow(rows, (row) => row.daysAgo != null);
     if (latestDatedRow && rowIsStaleForEval(latestDatedRow, testCase)) {
@@ -2665,18 +2297,18 @@ export function evaluateSummaryOutputFromSnapshot(testCase, snapshot, output) {
 
   const checks = [
     evaluateNoInsight(visibleOutput, testCase),
-    evaluateShape(visibleOutput, testCase),
+    evaluateShape(visibleOutput, testCase, context),
     evaluateRequiredMentions(visibleOutput, testCase),
     evaluateAnyOfMentions(visibleOutput, testCase),
     evaluateForbiddenPhrases(visibleOutput, testCase),
     evaluateForbiddenMentions(visibleOutput, testCase),
     evaluateExerciseMentions(visibleOutput, snapshot, context, testCase.surface, testCase),
     evaluateWorkoutClaims(visibleOutput, context, testCase),
-    evaluateAskClaims(visibleOutput, snapshot, testCase),
+    evaluateAskClaims(visibleOutput, snapshot, testCase, context),
     evaluateAskDirectionalConsistency(visibleOutput, snapshot, testCase),
     evaluateAskScoreVoice(visibleOutput, testCase),
     evaluateAskSelfReference(visibleOutput, testCase),
-    evaluateAskVolunteeredScore(visibleOutput, testCase),
+    evaluateAskVolunteeredScore(visibleOutput, testCase, context),
     evaluateAskStaleness(visibleOutput, snapshot, testCase),
     evaluateAskToolProvenance(visibleOutput, context, testCase, snapshot),
     evaluateFormulaVersion(visibleOutput, snapshot, testCase),