instar 1.3.738 → 1.3.739

package/dist/data/llmBenchCoverage.d.ts

@@ -35,4 +35,22 @@ export type UntrustedInputFlag = true | {
     false: string;
 };
 export declare const LLM_UNTRUSTED_INPUT: Readonly<Record<string, UntrustedInputFlag>>;
+/** The kind of claim a judge callsite credits/refuses (spec §2). Different kinds
+ * take different evidence, so each `judgesClaims: true` entry declares its kind:
+ *   - completionClaim: proof of asserted work ("the task is done / tests pass");
+ *   - healthClaim: behavioral-signal sufficiency (stall/stuck/health classifiers);
+ *   - scoredCredit: rubric evaluators that award credit for claimed work. */
+export type ClaimKind = 'completionClaim' | 'healthClaim' | 'scoredCredit';
+export declare const CLAIM_KINDS: ReadonlyArray<ClaimKind>;
+/** The `judgesClaims` classification of one LLM callsite:
+ *   - `{ claimKind }`  → judges a claim (true), of the declared kind;
+ *   - `false`          → does not judge a completion/progress/health claim;
+ *   - `{ false: '<reason>' }` → a judge-SHAPED callsite argued OUT of scope
+ *     (pinned shrink-only + reviewed, like the untrustedInput argued-false set). */
+export type JudgesClaimsFlag = {
+    claimKind: ClaimKind;
+} | false | {
+    false: string;
+};
+export declare const LLM_JUDGES_CLAIMS: Readonly<Record<string, JudgesClaimsFlag>>;
 //# sourceMappingURL=llmBenchCoverage.d.ts.map

package/dist/data/llmBenchCoverage.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"llmBenchCoverage.d.ts","sourceRoot":"","sources":["../../src/data/llmBenchCoverage.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAEH,MAAM,MAAM,aAAa,GACrB;IAAE,IAAI,EAAE,MAAM,CAAA;CAAE,GAChB;IAAE,OAAO,EAAE,QAAQ,GAAG,QAAQ,CAAA;CAAE,GAChC;IAAE,MAAM,EAAE,MAAM,CAAA;CAAE,CAAC;AAEvB,eAAO,MAAM,kBAAkB,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,aAAa,CAAC,CAqFtE,CAAC;AA4BF,MAAM,MAAM,kBAAkB,GAAG,IAAI,GAAG;IAAE,KAAK,EAAE,MAAM,CAAA;CAAE,CAAC;AAE1D,eAAO,MAAM,mBAAmB,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,kBAAkB,CAAC,CAiF5E,CAAC"}
+{"version":3,"file":"llmBenchCoverage.d.ts","sourceRoot":"","sources":["../../src/data/llmBenchCoverage.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAEH,MAAM,MAAM,aAAa,GACrB;IAAE,IAAI,EAAE,MAAM,CAAA;CAAE,GAChB;IAAE,OAAO,EAAE,QAAQ,GAAG,QAAQ,CAAA;CAAE,GAChC;IAAE,MAAM,EAAE,MAAM,CAAA;CAAE,CAAC;AAEvB,eAAO,MAAM,kBAAkB,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,aAAa,CAAC,CAqFtE,CAAC;AA4BF,MAAM,MAAM,kBAAkB,GAAG,IAAI,GAAG;IAAE,KAAK,EAAE,MAAM,CAAA;CAAE,CAAC;AAE1D,eAAO,MAAM,mBAAmB,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,kBAAkB,CAAC,CAiF5E,CAAC;AA+CF;;;;6EAI6E;AAC7E,MAAM,MAAM,SAAS,GAAG,iBAAiB,GAAG,aAAa,GAAG,cAAc,CAAC;AAE3E,eAAO,MAAM,WAAW,EAAE,aAAa,CAAC,SAAS,CAIhD,CAAC;AAEF;;;;oFAIoF;AACpF,MAAM,MAAM,gBAAgB,GAAG;IAAE,SAAS,EAAE,SAAS,CAAA;CAAE,GAAG,KAAK,GAAG;IAAE,KAAK,EAAE,MAAM,CAAA;CAAE,CAAC;AAEpF,eAAO,MAAM,iBAAiB,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,gBAAgB,CAAC,CAmExE,CAAC"}

package/dist/data/llmBenchCoverage.js

@@ -171,4 +171,72 @@ export const LLM_UNTRUSTED_INPUT = {
         false: 'attribution-manifest alias only (a legacy prompt-pattern matcher); the live matcher calls with attribution PromptGate, classified true there',
     },
 };
+export const CLAIM_KINDS = [
+    'completionClaim',
+    'healthClaim',
+    'scoredCredit',
+];
+export const LLM_JUDGES_CLAIMS = {
+    // ── Judges (judgesClaims: true) ────────────────────────────────────────────
+    // Completion evaluators — credit/refuse a claim of asserted work.
+    CompletionEvaluator: { claimKind: 'completionClaim' }, // THE motivating callsite: credited a bare "tests pass"
+    UnjustifiedStopGate: { claimKind: 'completionClaim' }, // judges whether a stop is justified = is the claimed work actually done/blocked
+    JobReflector: { claimKind: 'completionClaim' }, // reflects on job success/completion (spec §2 pending-wave judge; classified now so the asymmetry can't silently reopen)
+    // Stall/stuck/health classifiers — credit/refuse a health claim about a session.
+    SessionWatchdog: { claimKind: 'healthClaim' }, // stuck-judge
+    PresenceProxy: { claimKind: 'healthClaim' }, // tier-3 stall judge
+    StallTriageNurse: { claimKind: 'healthClaim' }, // stall-triage diagnosis
+    TelegramAdapter: { claimKind: 'healthClaim' }, // confirmStallAlert — judges whether a session is genuinely stalled before alerting (inclusion criteria: stall/health classifier)
+    SlackAdapter: { claimKind: 'healthClaim' }, // confirmStallAlert (byte-identical prompt to Telegram's; parity)
+    // Scored evaluators — award credit for claimed work.
+    'mentor-stage-b': { claimKind: 'scoredCredit' }, // classifies mentor signals over mentee output (spec §2 pending-wave judge)
+    // ── Not a completion/progress/health judge (judgesClaims: false) ────────────
+    // Sentinels that classify/summarize but do not credit a completion/health claim.
+    InputDetector: false, // legacy prompt-pattern matcher alias
+    InputGuard: false, // input-coherence, not a completion/health claim
+    SessionActivitySentinel: false, // activity DIGEST — summarizes activity, does not credit a claim (exclusion: pure summarizer)
+    CommitmentSentinel: false, // detects commitments in text
+    PromiseBeacon: false, // no live LLM prompt (hooks unwired at construction — see its bench exemption)
+    MessageSentinel: false, // classifies message intent (pause/emergency), not a completion claim
+    TopicIntentArcCheck: false, // arc-check classification of a topic's intent, not a completion/health claim
+    InputClassifier: false, // input auto-approve vs relay classification
+    SessionSummarySentinel: false, // summarizes tmux output → task/phase/files (exclusion: pure summarizer)
+    ProjectDriftChecker: false, // is-work-on-project coherence, not a completion/health claim
+    TemporalCoherenceChecker: false, // temporal-coherence classification
+    ResumeQueueDrainer: false, // inspects session state before a resume — judges STATE validity, not an agent's claim
+    InteractivePoolCanaryJudge: false, // judges a FIXED known-answer canary constant, not an agent claim
+    // Gates that classify an action/message but do not credit a completion/health claim.
+    PromptGate: false, // detects prompt-injection in content
+    AutoApprover: false, // mechanical key injection, no LLM prompt of its own
+    IntegrationGate: false, // delegates to JobReflector.reflect(), no own callsite
+    ExternalOperationGate: false, // classifies operation mutability/reversibility, not a completion claim
+    WarrantsReplyGate: false, // "should I reply?" — not a completion/health claim
+    CoherenceGate: false, // no own callsite — flows through CoherenceReviewer
+    MessagingToneGate: false, // reviews outbound tone/leaks
+    CoherenceReviewer: false, // reviews outbound coherence
+    LLMSanitizer: false, // sanitizes untrusted inbound content
+    OverrideDetector: false, // detects override intent in a turn
+    TaskClassifier: false, // classifies task type
+    ResumeValidator: false, // matches a resume UUID against a topic — a state match, not a claim
+    // Reflectors/jobs that extract/summarize/route but do not credit a completion/health claim.
+    crossModelReviewer: false, // reviews a SPEC document, not a session's completion claim
+    SelfKnowledgeTree: false, // extracts self-knowledge
+    TreeTriage: false, // triages tree fragments
+    TopicSummarizer: false, // summarizes a topic
+    ContextualEvaluator: false, // evaluates context relevance
+    RelationshipManager: false, // extracts relationship facts
+    StandardsConformanceReviewer: false, // reviews artifact-vs-standard conformance, not a session completion claim
+    DiscoveryEvaluator: false, // evaluates serendipity discoveries
+    Usher: false, // routes a turn to candidate topics
+    TopicIntentExtractor: false, // extracts topic intent from a turn
+    PreCompactionFlush: false, // extracts durable facts before compaction
+    TreeSynthesis: false, // synthesizes knowledge fragments → answer
+    LLMConflictResolver: false, // resolves divergent multi-machine state
+    openConversationBrief: false, // generates an A2A conversation brief
+    'a2a-checkin': false, // summarizes A2A check-in threads
+    'correction-learning': false, // distills recurring corrections → preference
+    PipeSessionSpawner: false, // spawns from task descriptions
+    CartographerSweep: false, // authors doc-tree summaries over code
+    StandardsCoverageEnrichment: false, // enriches standards-coverage rows
+};
 //# sourceMappingURL=llmBenchCoverage.js.map

package/dist/data/llmBenchCoverage.js.map

@@ -1 +1 @@
-{"version":3,"file":"llmBenchCoverage.js","sourceRoot":"","sources":["../../src/data/llmBenchCoverage.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAOH,MAAM,CAAC,MAAM,kBAAkB,GAA4C;IACzE,mEAAmE;IACnE,eAAe,EAAE,EAAE,IAAI,EAAE,mBAAmB,EAAE;IAC9C,iBAAiB,EAAE,EAAE,IAAI,EAAE,WAAW,EAAE;IACxC,6DAA6D;IAC7D,mBAAmB,EAAE,EAAE,IAAI,EAAE,iCAAiC,EAAE;IAChE,qBAAqB,EAAE,EAAE,IAAI,EAAE,kBAAkB,EAAE;IACnD,YAAY,EAAE,EAAE,IAAI,EAAE,qBAAqB,EAAE;IAC7C,iBAAiB,EAAE,EAAE,IAAI,EAAE,gBAAgB,EAAE;IAC7C,eAAe,EAAE,EAAE,IAAI,EAAE,kBAAkB,EAAE;IAC7C,KAAK,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE;IACxB,qBAAqB,EAAE,EAAE,IAAI,EAAE,sBAAsB,EAAE;IACvD,iBAAiB,EAAE,EAAE,IAAI,EAAE,aAAa,EAAE;IAE1C,kEAAkE;IAClE,0BAA0B,EAAE;QAC1B,MAAM,EACJ,0HAA0H;KAC7H;IAED,mFAAmF;IACnF,UAAU,EAAE,EAAE,IAAI,EAAE,uBAAuB,EAAE;IAC7C,uBAAuB,EAAE,EAAE,IAAI,EAAE,iBAAiB,EAAE;IACpD,gBAAgB,EAAE,EAAE,IAAI,EAAE,wBAAwB,EAAE;IACpD,kBAAkB,EAAE,EAAE,IAAI,EAAE,qBAAqB,EAAE;IACnD,aAAa,EAAE,EAAE,IAAI,EAAE,sBAAsB,EAAE;IAC/C,mBAAmB,EAAE,EAAE,IAAI,EAAE,qBAAqB,EAAE;IACpD,wBAAwB,EAAE,EAAE,IAAI,EAAE,oBAAoB,EAAE;IACxD,eAAe,EAAE,EAAE,IAAI,EAAE,sBAAsB,EAAE;IACjD,kBAAkB,EAAE,EAAE,IAAI,EAAE,qBAAqB,EAAE;IACnD,mBAAmB,EAAE,EAAE,IAAI,EAAE,oBAAoB,EAAE;IACnD,eAAe,EAAE,EAAE,IAAI,EAAE,wBAAwB,EAAE;IACnD,mEAAmE;IACnE,sEAAsE;IACtE,iEAAiE;IACjE,YAAY,EAAE,EAAE,IAAI,EAAE,wBAAwB,EAAE;IAChD,UAAU,EAAE,EAAE,IAAI,EAAE,oBAAoB,EAAE;IAC1C,mBAAmB,EAAE,EAAE,IAAI,EAAE,uBAAuB,EAAE;IACtD,gBAAgB,EAAE,EAAE,IAAI,EAAE,mBAAmB,EAAE;IAC/C,cAAc,EAAE,EAAE,IAAI,EAAE,iBAAiB,EAAE;IAC3C,eAAe,EAAE,EAAE,IAAI,EAAE,kBAAkB,EAAE;IAC7C,sBAAsB,EAAE,EAAE,IAAI,EAAE,0BAA0B,EAAE;IAC5D,oBAAoB,EAAE,EAAE,IAAI,EAAE,wBAAwB,EAAE;IAExD,yFAAyF;IACzF,eAAe,EAAE;QACf,MAAM,EACJ,yKAAyK;KAC5K;IACD,aAAa,EAAE;QACb,MAAM,EACJ,qKAAqK;KACxK;IACD,YAAY,EAAE;QACZ,MAAM,EACJ,6IAA6I;KAChJ;IACD,aAAa,EAAE;QACb,MAAM,EACJ,mLAAmL;KACtL;IACD,aAAa,EAAE;QACb,MAAM,EACJ,+LAA+L;KAClM;IAED,mDAAmD;IACnD,YAAY,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACnC,kBAAkB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACzC,iBAAiB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACxC,UAAU,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACjC,eAAe,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACtC,mBAAmB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IAC1C,mBAAmB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IAC1C,4BAA4B,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACnD,kBAAkB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACzC,kBAAkB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACzC,iBAAiB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACxC,2BAA2B,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IAClD,kBAAkB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACzC,aAAa,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACpC,mBAAmB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IAC1C,qBAAqB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IAC5C,aAAa,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACpC,gBAAgB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;CACxC,CAAC;AA8BF,MAAM,CAAC,MAAM,mBAAmB,GAAiD;IAC/E,yEAAyE;IACzE,UAAU,EAAE,IAAI;IAChB,uBAAuB,EAAE,IAAI;IAC7B,gBAAgB,EAAE,IAAI;IACtB,kBAAkB,EAAE,IAAI;IACxB,aAAa,EAAE,IAAI;IACnB,eAAe,EAAE,IAAI;IACrB,mBAAmB,EAAE,IAAI;IACzB,wBAAwB,EAAE,IAAI;IAC9B,mBAAmB,EAAE,IAAI;IACzB,eAAe,EAAE,IAAI;IACrB,kBAAkB,EAAE,IAAI;IACxB,mBAAmB,EAAE,IAAI;IACzB,YAAY,EAAE,IAAI;IAClB,eAAe,EAAE,IAAI;IACrB,sBAAsB,EAAE,IAAI;IAC5B,eAAe,EAAE,IAAI;IAErB,4DAA4D;IAC5D,UAAU,EAAE,IAAI;IAChB,qBAAqB,EAAE,IAAI,EAAE,uEAAuE;IACpG,iBAAiB,EAAE,IAAI;IACvB,mBAAmB,EAAE,IAAI;IACzB,iBAAiB,EAAE,IAAI,EAAE,oEAAoE;IAC7F,iBAAiB,EAAE,IAAI;IACvB,YAAY,EAAE,IAAI,EAAE,kDAAkD;IACtE,gBAAgB,EAAE,IAAI;IACtB,cAAc,EAAE,IAAI;IACpB,eAAe,EAAE,IAAI,EAAE,wEAAwE;IAE/F,oFAAoF;IACpF,YAAY,EAAE,IAAI;IAClB,kBAAkB,EAAE,IAAI;IACxB,iBAAiB,EAAE,IAAI;IACvB,UAAU,EAAE,IAAI;IAChB,eAAe,EAAE,IAAI;IACrB,mBAAmB,EAAE,IAAI;IACzB,mBAAmB,EAAE,IAAI;IACzB,4BAA4B,EAAE,IAAI;IAClC,kBAAkB,EAAE,IAAI;IACxB,KAAK,EAAE,IAAI;IACX,oBAAoB,EAAE,IAAI;IAC1B,kBAAkB,EAAE,IAAI;IACxB,aAAa,EAAE,IAAI;IACnB,mBAAmB,EAAE,IAAI,EAAE,sDAAsD;IACjF,qBAAqB,EAAE,IAAI;IAC3B,aAAa,EAAE,IAAI,EAAE,4BAA4B;IACjD,qBAAqB,EAAE,IAAI;IAC3B,gBAAgB,EAAE,IAAI;IAEtB,+DAA+D;IAC/D,kBAAkB,EAAE,IAAI,EAAE,0DAA0D;IACpF,iBAAiB,EAAE,IAAI,EAAE,gGAAgG;IACzH,2BAA2B,EAAE,IAAI;IAEjC,0FAA0F;IAC1F,aAAa,EAAE;QACb,KAAK,EACH,8KAA8K;KACjL;IACD,0BAA0B,EAAE;QAC1B,KAAK,EACH,iIAAiI;KACpI;IACD,YAAY,EAAE;QACZ,KAAK,EACH,2IAA2I;KAC9I;IACD,eAAe,EAAE;QACf,KAAK,EACH,mIAAmI;KACtI;IACD,aAAa,EAAE;QACb,KAAK,EACH,+HAA+H;KAClI;IACD,aAAa,EAAE;QACb,KAAK,EACH,8IAA8I;KACjJ;CACF,CAAC"}
+{"version":3,"file":"llmBenchCoverage.js","sourceRoot":"","sources":["../../src/data/llmBenchCoverage.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAOH,MAAM,CAAC,MAAM,kBAAkB,GAA4C;IACzE,mEAAmE;IACnE,eAAe,EAAE,EAAE,IAAI,EAAE,mBAAmB,EAAE;IAC9C,iBAAiB,EAAE,EAAE,IAAI,EAAE,WAAW,EAAE;IACxC,6DAA6D;IAC7D,mBAAmB,EAAE,EAAE,IAAI,EAAE,iCAAiC,EAAE;IAChE,qBAAqB,EAAE,EAAE,IAAI,EAAE,kBAAkB,EAAE;IACnD,YAAY,EAAE,EAAE,IAAI,EAAE,qBAAqB,EAAE;IAC7C,iBAAiB,EAAE,EAAE,IAAI,EAAE,gBAAgB,EAAE;IAC7C,eAAe,EAAE,EAAE,IAAI,EAAE,kBAAkB,EAAE;IAC7C,KAAK,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE;IACxB,qBAAqB,EAAE,EAAE,IAAI,EAAE,sBAAsB,EAAE;IACvD,iBAAiB,EAAE,EAAE,IAAI,EAAE,aAAa,EAAE;IAE1C,kEAAkE;IAClE,0BAA0B,EAAE;QAC1B,MAAM,EACJ,0HAA0H;KAC7H;IAED,mFAAmF;IACnF,UAAU,EAAE,EAAE,IAAI,EAAE,uBAAuB,EAAE;IAC7C,uBAAuB,EAAE,EAAE,IAAI,EAAE,iBAAiB,EAAE;IACpD,gBAAgB,EAAE,EAAE,IAAI,EAAE,wBAAwB,EAAE;IACpD,kBAAkB,EAAE,EAAE,IAAI,EAAE,qBAAqB,EAAE;IACnD,aAAa,EAAE,EAAE,IAAI,EAAE,sBAAsB,EAAE;IAC/C,mBAAmB,EAAE,EAAE,IAAI,EAAE,qBAAqB,EAAE;IACpD,wBAAwB,EAAE,EAAE,IAAI,EAAE,oBAAoB,EAAE;IACxD,eAAe,EAAE,EAAE,IAAI,EAAE,sBAAsB,EAAE;IACjD,kBAAkB,EAAE,EAAE,IAAI,EAAE,qBAAqB,EAAE;IACnD,mBAAmB,EAAE,EAAE,IAAI,EAAE,oBAAoB,EAAE;IACnD,eAAe,EAAE,EAAE,IAAI,EAAE,wBAAwB,EAAE;IACnD,mEAAmE;IACnE,sEAAsE;IACtE,iEAAiE;IACjE,YAAY,EAAE,EAAE,IAAI,EAAE,wBAAwB,EAAE;IAChD,UAAU,EAAE,EAAE,IAAI,EAAE,oBAAoB,EAAE;IAC1C,mBAAmB,EAAE,EAAE,IAAI,EAAE,uBAAuB,EAAE;IACtD,gBAAgB,EAAE,EAAE,IAAI,EAAE,mBAAmB,EAAE;IAC/C,cAAc,EAAE,EAAE,IAAI,EAAE,iBAAiB,EAAE;IAC3C,eAAe,EAAE,EAAE,IAAI,EAAE,kBAAkB,EAAE;IAC7C,sBAAsB,EAAE,EAAE,IAAI,EAAE,0BAA0B,EAAE;IAC5D,oBAAoB,EAAE,EAAE,IAAI,EAAE,wBAAwB,EAAE;IAExD,yFAAyF;IACzF,eAAe,EAAE;QACf,MAAM,EACJ,yKAAyK;KAC5K;IACD,aAAa,EAAE;QACb,MAAM,EACJ,qKAAqK;KACxK;IACD,YAAY,EAAE;QACZ,MAAM,EACJ,6IAA6I;KAChJ;IACD,aAAa,EAAE;QACb,MAAM,EACJ,mLAAmL;KACtL;IACD,aAAa,EAAE;QACb,MAAM,EACJ,+LAA+L;KAClM;IAED,mDAAmD;IACnD,YAAY,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACnC,kBAAkB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACzC,iBAAiB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACxC,UAAU,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACjC,eAAe,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACtC,mBAAmB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IAC1C,mBAAmB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IAC1C,4BAA4B,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACnD,kBAAkB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACzC,kBAAkB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACzC,iBAAiB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACxC,2BAA2B,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IAClD,kBAAkB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACzC,aAAa,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACpC,mBAAmB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IAC1C,qBAAqB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IAC5C,aAAa,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;IACpC,gBAAgB,EAAE,EAAE,OAAO,EAAE,QAAQ,EAAE;CACxC,CAAC;AA8BF,MAAM,CAAC,MAAM,mBAAmB,GAAiD;IAC/E,yEAAyE;IACzE,UAAU,EAAE,IAAI;IAChB,uBAAuB,EAAE,IAAI;IAC7B,gBAAgB,EAAE,IAAI;IACtB,kBAAkB,EAAE,IAAI;IACxB,aAAa,EAAE,IAAI;IACnB,eAAe,EAAE,IAAI;IACrB,mBAAmB,EAAE,IAAI;IACzB,wBAAwB,EAAE,IAAI;IAC9B,mBAAmB,EAAE,IAAI;IACzB,eAAe,EAAE,IAAI;IACrB,kBAAkB,EAAE,IAAI;IACxB,mBAAmB,EAAE,IAAI;IACzB,YAAY,EAAE,IAAI;IAClB,eAAe,EAAE,IAAI;IACrB,sBAAsB,EAAE,IAAI;IAC5B,eAAe,EAAE,IAAI;IAErB,4DAA4D;IAC5D,UAAU,EAAE,IAAI;IAChB,qBAAqB,EAAE,IAAI,EAAE,uEAAuE;IACpG,iBAAiB,EAAE,IAAI;IACvB,mBAAmB,EAAE,IAAI;IACzB,iBAAiB,EAAE,IAAI,EAAE,oEAAoE;IAC7F,iBAAiB,EAAE,IAAI;IACvB,YAAY,EAAE,IAAI,EAAE,kDAAkD;IACtE,gBAAgB,EAAE,IAAI;IACtB,cAAc,EAAE,IAAI;IACpB,eAAe,EAAE,IAAI,EAAE,wEAAwE;IAE/F,oFAAoF;IACpF,YAAY,EAAE,IAAI;IAClB,kBAAkB,EAAE,IAAI;IACxB,iBAAiB,EAAE,IAAI;IACvB,UAAU,EAAE,IAAI;IAChB,eAAe,EAAE,IAAI;IACrB,mBAAmB,EAAE,IAAI;IACzB,mBAAmB,EAAE,IAAI;IACzB,4BAA4B,EAAE,IAAI;IAClC,kBAAkB,EAAE,IAAI;IACxB,KAAK,EAAE,IAAI;IACX,oBAAoB,EAAE,IAAI;IAC1B,kBAAkB,EAAE,IAAI;IACxB,aAAa,EAAE,IAAI;IACnB,mBAAmB,EAAE,IAAI,EAAE,sDAAsD;IACjF,qBAAqB,EAAE,IAAI;IAC3B,aAAa,EAAE,IAAI,EAAE,4BAA4B;IACjD,qBAAqB,EAAE,IAAI;IAC3B,gBAAgB,EAAE,IAAI;IAEtB,+DAA+D;IAC/D,kBAAkB,EAAE,IAAI,EAAE,0DAA0D;IACpF,iBAAiB,EAAE,IAAI,EAAE,gGAAgG;IACzH,2BAA2B,EAAE,IAAI;IAEjC,0FAA0F;IAC1F,aAAa,EAAE;QACb,KAAK,EACH,8KAA8K;KACjL;IACD,0BAA0B,EAAE;QAC1B,KAAK,EACH,iIAAiI;KACpI;IACD,YAAY,EAAE;QACZ,KAAK,EACH,2IAA2I;KAC9I;IACD,eAAe,EAAE;QACf,KAAK,EACH,mIAAmI;KACtI;IACD,aAAa,EAAE;QACb,KAAK,EACH,+HAA+H;KAClI;IACD,aAAa,EAAE;QACb,KAAK,EACH,8IAA8I;KACjJ;CACF,CAAC;AAsDF,MAAM,CAAC,MAAM,WAAW,GAA6B;IACnD,iBAAiB;IACjB,aAAa;IACb,cAAc;CACf,CAAC;AASF,MAAM,CAAC,MAAM,iBAAiB,GAA+C;IAC3E,8EAA8E;IAC9E,kEAAkE;IAClE,mBAAmB,EAAE,EAAE,SAAS,EAAE,iBAAiB,EAAE,EAAE,wDAAwD;IAC/G,mBAAmB,EAAE,EAAE,SAAS,EAAE,iBAAiB,EAAE,EAAE,iFAAiF;IACxI,YAAY,EAAE,EAAE,SAAS,EAAE,iBAAiB,EAAE,EAAE,yHAAyH;IAEzK,iFAAiF;IACjF,eAAe,EAAE,EAAE,SAAS,EAAE,aAAa,EAAE,EAAE,cAAc;IAC7D,aAAa,EAAE,EAAE,SAAS,EAAE,aAAa,EAAE,EAAE,qBAAqB;IAClE,gBAAgB,EAAE,EAAE,SAAS,EAAE,aAAa,EAAE,EAAE,yBAAyB;IACzE,eAAe,EAAE,EAAE,SAAS,EAAE,aAAa,EAAE,EAAE,kIAAkI;IACjL,YAAY,EAAE,EAAE,SAAS,EAAE,aAAa,EAAE,EAAE,kEAAkE;IAE9G,qDAAqD;IACrD,gBAAgB,EAAE,EAAE,SAAS,EAAE,cAAc,EAAE,EAAE,4EAA4E;IAE7H,+EAA+E;IAC/E,iFAAiF;IACjF,aAAa,EAAE,KAAK,EAAE,sCAAsC;IAC5D,UAAU,EAAE,KAAK,EAAE,iDAAiD;IACpE,uBAAuB,EAAE,KAAK,EAAE,8FAA8F;IAC9H,kBAAkB,EAAE,KAAK,EAAE,8BAA8B;IACzD,aAAa,EAAE,KAAK,EAAE,+EAA+E;IACrG,eAAe,EAAE,KAAK,EAAE,sEAAsE;IAC9F,mBAAmB,EAAE,KAAK,EAAE,8EAA8E;IAC1G,eAAe,EAAE,KAAK,EAAE,6CAA6C;IACrE,sBAAsB,EAAE,KAAK,EAAE,yEAAyE;IACxG,mBAAmB,EAAE,KAAK,EAAE,8DAA8D;IAC1F,wBAAwB,EAAE,KAAK,EAAE,oCAAoC;IACrE,kBAAkB,EAAE,KAAK,EAAE,uFAAuF;IAClH,0BAA0B,EAAE,KAAK,EAAE,kEAAkE;IAErG,qFAAqF;IACrF,UAAU,EAAE,KAAK,EAAE,sCAAsC;IACzD,YAAY,EAAE,KAAK,EAAE,qDAAqD;IAC1E,eAAe,EAAE,KAAK,EAAE,uDAAuD;IAC/E,qBAAqB,EAAE,KAAK,EAAE,wEAAwE;IACtG,iBAAiB,EAAE,KAAK,EAAE,oDAAoD;IAC9E,aAAa,EAAE,KAAK,EAAE,oDAAoD;IAC1E,iBAAiB,EAAE,KAAK,EAAE,8BAA8B;IACxD,iBAAiB,EAAE,KAAK,EAAE,6BAA6B;IACvD,YAAY,EAAE,KAAK,EAAE,sCAAsC;IAC3D,gBAAgB,EAAE,KAAK,EAAE,oCAAoC;IAC7D,cAAc,EAAE,KAAK,EAAE,uBAAuB;IAC9C,eAAe,EAAE,KAAK,EAAE,qEAAqE;IAE7F,4FAA4F;IAC5F,kBAAkB,EAAE,KAAK,EAAE,4DAA4D;IACvF,iBAAiB,EAAE,KAAK,EAAE,0BAA0B;IACpD,UAAU,EAAE,KAAK,EAAE,yBAAyB;IAC5C,eAAe,EAAE,KAAK,EAAE,qBAAqB;IAC7C,mBAAmB,EAAE,KAAK,EAAE,8BAA8B;IAC1D,mBAAmB,EAAE,KAAK,EAAE,8BAA8B;IAC1D,4BAA4B,EAAE,KAAK,EAAE,2EAA2E;IAChH,kBAAkB,EAAE,KAAK,EAAE,oCAAoC;IAC/D,KAAK,EAAE,KAAK,EAAE,oCAAoC;IAClD,oBAAoB,EAAE,KAAK,EAAE,oCAAoC;IACjE,kBAAkB,EAAE,KAAK,EAAE,2CAA2C;IACtE,aAAa,EAAE,KAAK,EAAE,2CAA2C;IACjE,mBAAmB,EAAE,KAAK,EAAE,yCAAyC;IACrE,qBAAqB,EAAE,KAAK,EAAE,sCAAsC;IACpE,aAAa,EAAE,KAAK,EAAE,kCAAkC;IACxD,qBAAqB,EAAE,KAAK,EAAE,8CAA8C;IAC5E,kBAAkB,EAAE,KAAK,EAAE,gCAAgC;IAC3D,iBAAiB,EAAE,KAAK,EAAE,uCAAuC;IACjE,2BAA2B,EAAE,KAAK,EAAE,mCAAmC;CACxE,CAAC"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "instar",
-  "version": "1.3.738",
+  "version": "1.3.739",
   "description": "Coherence infrastructure for self-evolving AI agents — on the Claude Code or Codex subscription you already have.",
   "type": "module",
   "main": "dist/index.js",

package/src/data/builtin-manifest.json

@@ -1,8 +1,8 @@
 {
   "$schema": "./builtin-manifest.schema.json",
   "schemaVersion": 1,
-  "generatedAt": "2026-07-03T22:30:08.221Z",
-  "instarVersion": "1.3.738",
+  "generatedAt": "2026-07-03T22:57:15.282Z",
+  "instarVersion": "1.3.739",
   "entryCount": 202,
   "entries": {
     "hook:session-start": {

package/src/data/llmBenchCoverage.ts

@@ -226,3 +226,137 @@ export const LLM_UNTRUSTED_INPUT: Readonly<Record<string, UntrustedInputFlag>> =
       'attribution-manifest alias only (a legacy prompt-pattern matcher); the live matcher calls with attribution PromptGate, classified true there',
   },
 };
+
+// ───────────────────────────────────────────────────────────────────────────
+// Evidence-Bar Extension — Judge Prompts (defect class 3 / claim-vs-evidence —
+// docs/specs/evidence-bar-judge-extension.md §2 "the judge-nature classification")
+//
+// The `judgesClaims` axis of the program's shared per-callsite metadata record
+// (class-closure-gate.md §"Program-shared machinery"; co-located in THIS file
+// with the bench-coverage record it extends, exactly like the sibling
+// `untrustedInput` axis). The consolidated axis ratchet derives the required
+// bench axes from all these per-callsite axes together.
+//
+// WHY (earned): on 2026-07-02 the INSTAR-Bench v2 defect-class review found the
+// completion judge (and four other model routes on the same case) credited a
+// BARE assertion ("tests pass," no output shown) as satisfied evidence — the
+// judge prompt never defined what counts as evidence, so models defaulted to
+// crediting the claim. The agent-facing "Bug-Fix Evidence Bar" holds the
+// CLAIMANT ("verify before you claim"); the prompts that JUDGE such claims were
+// never given the same rule. This classification is the structural record of
+// WHICH callsites judge a claim, so the (deferred) bench-axis ratchet can
+// require both a bare-claim (false-accept) AND a real-evidence (false-reject)
+// case for each — the known over-correction hazard (the first fix rejected REAL
+// evidence on 6 routes) makes the false-reject direction mandatory.
+//
+// INCLUSION CRITERIA (spec §2): the callsite's task is to CREDIT or REFUSE an
+// agent/session claim of completion, progress, or health — verdict gates,
+// completion evaluators, stuck/stall/health classifiers, and scored evaluators
+// that award credit for claimed work. EXCLUSION: pure summarizers and
+// extractors are out unless they score or credit a claim.
+//
+// THE FIELD IS REQUIRED AND EXPLICIT FOR EVERY COMPONENT_CATEGORY KEY — there is
+// NO DEFAULT (same polarity rule as `untrustedInput`). A `{ claimKind }` entry
+// is a judge (true) and declares WHICH kind of claim it judges (the axis cases +
+// accepted evidence classes are authored per kind). Plain `false` is a callsite
+// that does not judge a completion/progress/health claim. A judge-SHAPED
+// callsite that argues it does NOT judge claims is written as
+// `{ false: '<argued reason>' }` and pinned shrink-only in
+// tests/unit/judges-claims-classification-ratchet.test.ts — a silent omission
+// is red CI, so the flag can never default toward the un-benched state.
+//
+// DETERMINISTIC ARM (spec "Honest reach"): the real-check `verification_command`
+// verifier is NAMED in the spec seed but is NOT an LLM callsite (it runs the
+// actual command on a met-verdict) — it has no COMPONENT_CATEGORY key and is
+// therefore out of this LLM classification by construction. A prompt bar governs
+// what is SHOWN; verification of what is TRUE belongs to that deterministic arm.
+// ───────────────────────────────────────────────────────────────────────────
+
+/** The kind of claim a judge callsite credits/refuses (spec §2). Different kinds
+ * take different evidence, so each `judgesClaims: true` entry declares its kind:
+ *   - completionClaim: proof of asserted work ("the task is done / tests pass");
+ *   - healthClaim: behavioral-signal sufficiency (stall/stuck/health classifiers);
+ *   - scoredCredit: rubric evaluators that award credit for claimed work. */
+export type ClaimKind = 'completionClaim' | 'healthClaim' | 'scoredCredit';
+
+export const CLAIM_KINDS: ReadonlyArray<ClaimKind> = [
+  'completionClaim',
+  'healthClaim',
+  'scoredCredit',
+];
+
+/** The `judgesClaims` classification of one LLM callsite:
+ *   - `{ claimKind }`  → judges a claim (true), of the declared kind;
+ *   - `false`          → does not judge a completion/progress/health claim;
+ *   - `{ false: '<reason>' }` → a judge-SHAPED callsite argued OUT of scope
+ *     (pinned shrink-only + reviewed, like the untrustedInput argued-false set). */
+export type JudgesClaimsFlag = { claimKind: ClaimKind } | false | { false: string };
+
+export const LLM_JUDGES_CLAIMS: Readonly<Record<string, JudgesClaimsFlag>> = {
+  // ── Judges (judgesClaims: true) ────────────────────────────────────────────
+  // Completion evaluators — credit/refuse a claim of asserted work.
+  CompletionEvaluator: { claimKind: 'completionClaim' }, // THE motivating callsite: credited a bare "tests pass"
+  UnjustifiedStopGate: { claimKind: 'completionClaim' }, // judges whether a stop is justified = is the claimed work actually done/blocked
+  JobReflector: { claimKind: 'completionClaim' }, // reflects on job success/completion (spec §2 pending-wave judge; classified now so the asymmetry can't silently reopen)
+
+  // Stall/stuck/health classifiers — credit/refuse a health claim about a session.
+  SessionWatchdog: { claimKind: 'healthClaim' }, // stuck-judge
+  PresenceProxy: { claimKind: 'healthClaim' }, // tier-3 stall judge
+  StallTriageNurse: { claimKind: 'healthClaim' }, // stall-triage diagnosis
+  TelegramAdapter: { claimKind: 'healthClaim' }, // confirmStallAlert — judges whether a session is genuinely stalled before alerting (inclusion criteria: stall/health classifier)
+  SlackAdapter: { claimKind: 'healthClaim' }, // confirmStallAlert (byte-identical prompt to Telegram's; parity)
+
+  // Scored evaluators — award credit for claimed work.
+  'mentor-stage-b': { claimKind: 'scoredCredit' }, // classifies mentor signals over mentee output (spec §2 pending-wave judge)
+
+  // ── Not a completion/progress/health judge (judgesClaims: false) ────────────
+  // Sentinels that classify/summarize but do not credit a completion/health claim.
+  InputDetector: false, // legacy prompt-pattern matcher alias
+  InputGuard: false, // input-coherence, not a completion/health claim
+  SessionActivitySentinel: false, // activity DIGEST — summarizes activity, does not credit a claim (exclusion: pure summarizer)
+  CommitmentSentinel: false, // detects commitments in text
+  PromiseBeacon: false, // no live LLM prompt (hooks unwired at construction — see its bench exemption)
+  MessageSentinel: false, // classifies message intent (pause/emergency), not a completion claim
+  TopicIntentArcCheck: false, // arc-check classification of a topic's intent, not a completion/health claim
+  InputClassifier: false, // input auto-approve vs relay classification
+  SessionSummarySentinel: false, // summarizes tmux output → task/phase/files (exclusion: pure summarizer)
+  ProjectDriftChecker: false, // is-work-on-project coherence, not a completion/health claim
+  TemporalCoherenceChecker: false, // temporal-coherence classification
+  ResumeQueueDrainer: false, // inspects session state before a resume — judges STATE validity, not an agent's claim
+  InteractivePoolCanaryJudge: false, // judges a FIXED known-answer canary constant, not an agent claim
+
+  // Gates that classify an action/message but do not credit a completion/health claim.
+  PromptGate: false, // detects prompt-injection in content
+  AutoApprover: false, // mechanical key injection, no LLM prompt of its own
+  IntegrationGate: false, // delegates to JobReflector.reflect(), no own callsite
+  ExternalOperationGate: false, // classifies operation mutability/reversibility, not a completion claim
+  WarrantsReplyGate: false, // "should I reply?" — not a completion/health claim
+  CoherenceGate: false, // no own callsite — flows through CoherenceReviewer
+  MessagingToneGate: false, // reviews outbound tone/leaks
+  CoherenceReviewer: false, // reviews outbound coherence
+  LLMSanitizer: false, // sanitizes untrusted inbound content
+  OverrideDetector: false, // detects override intent in a turn
+  TaskClassifier: false, // classifies task type
+  ResumeValidator: false, // matches a resume UUID against a topic — a state match, not a claim
+
+  // Reflectors/jobs that extract/summarize/route but do not credit a completion/health claim.
+  crossModelReviewer: false, // reviews a SPEC document, not a session's completion claim
+  SelfKnowledgeTree: false, // extracts self-knowledge
+  TreeTriage: false, // triages tree fragments
+  TopicSummarizer: false, // summarizes a topic
+  ContextualEvaluator: false, // evaluates context relevance
+  RelationshipManager: false, // extracts relationship facts
+  StandardsConformanceReviewer: false, // reviews artifact-vs-standard conformance, not a session completion claim
+  DiscoveryEvaluator: false, // evaluates serendipity discoveries
+  Usher: false, // routes a turn to candidate topics
+  TopicIntentExtractor: false, // extracts topic intent from a turn
+  PreCompactionFlush: false, // extracts durable facts before compaction
+  TreeSynthesis: false, // synthesizes knowledge fragments → answer
+  LLMConflictResolver: false, // resolves divergent multi-machine state
+  openConversationBrief: false, // generates an A2A conversation brief
+  'a2a-checkin': false, // summarizes A2A check-in threads
+  'correction-learning': false, // distills recurring corrections → preference
+  PipeSessionSpawner: false, // spawns from task descriptions
+  CartographerSweep: false, // authors doc-tree summaries over code
+  StandardsCoverageEnrichment: false, // enriches standards-coverage rows
+};

package/upgrades/1.3.739.md

@@ -0,0 +1,77 @@
+# Upgrade Guide — vNEXT
+
+<!-- assembled-by: assemble-next-md -->
+<!-- bump: minor -->
+
+## What Changed
+
+The mechanical arm of the **Evidence-Bar Extension to Judge Prompts** standard
+(`docs/specs/evidence-bar-judge-extension.md`, defect class 3 / `claim-vs-evidence` closure),
+shipped as a self-contained DARK increment — no runtime wiring, no operator-gated registry
+amendment, no config key, no behavioral prompt change. It is the structural answer to the
+2026-07-02 INSTAR-Bench v2 finding that the completion judge (and four other model routes on
+the same case) credited a BARE assertion ("tests pass," no output shown) as satisfied evidence:
+the agent-facing **Bug-Fix Evidence Bar** holds the CLAIMANT ("verify before you claim"), but
+the prompts that JUDGE such claims were never given the same rule — an asymmetry where we hold
+the claimer to a bar the judge does not know exists.
+
+This increment ships **the judge-nature classification + its CI ratchet only**:
+
+- The **`judgesClaims` classification** (`LLM_JUDGES_CLAIMS` in `src/data/llmBenchCoverage.ts`):
+  the `judgesClaims` axis of the program's ONE shared per-callsite metadata record (sibling of
+  the authority-clause `untrustedInput` axis), required-explicit for every LLM component (no
+  default — a silent omission is red CI). A judge is a `{ claimKind }` entry declaring WHICH
+  kind of claim it credits/refuses — `completionClaim` (proof of asserted work), `healthClaim`
+  (stall/stuck/health signal sufficiency), or `scoredCredit` (rubric evaluators) — because those
+  are NOT one evidence problem and their axis cases are authored per kind. Nine callsites classify
+  as judges (the five measured completion/health judges — CompletionEvaluator, UnjustifiedStopGate,
+  SessionWatchdog, PresenceProxy, StallTriageNurse — plus the two stall-confirm adapters and the
+  two spec-named pending-wave judges JobReflector + mentor-stage-b); the deterministic real-check
+  `verification_command` verifier is seed-named but is NOT an LLM callsite (it runs the actual
+  command), out of the classification by construction.
+- A **classification ratchet** (`tests/unit/judges-claims-classification-ratchet.test.ts`):
+  required-explicit + no-dangling + valid-claimKind-per-judge + the spec-named JUDGE SEED pinned
+  as a judge (the callsites measured crediting a bare claim can't silently slip out of scope) +
+  the argued-false set pinned shrink-only with a real-reason floor. Same pinned-baseline family as
+  `llm-bench-coverage-ratchet` and `untrusted-input-classification-ratchet`.
+
+Deliberately OUT of scope (not orphan deferrals — see `upgrades/side-effects/`):
+- The **registry amendment** (spec §1) — operator-gated; ships ONLY with Justin's explicit
+  sign-off. It is DRAFTED in the spec; this run does not edit the standards registry.
+- The **bench-axis pair ratchet** (spec §3: a bare-claim false-accept case + a real-evidence
+  false-reject case per judge) — blocked on the SAME program-wide "batteries readable by CI"
+  decision the sibling authority-clause axis ratchet deferred on (`research/` is absent from
+  canonical main), owned by `class-closure-gate.md` §"Program-shared machinery" and binding all
+  three axis specs.
+- The **`evidenceBar()` prompt clause** (spec §4) — a sibling of the authority clause in
+  `src/core/promptClauses.ts`. That shared library is introduced by the still-open sibling PR
+  (authority-clause); adding the clause here would collide on the same new file. The clause is
+  only consumed via the A/B-gated per-component migrations (spec §4/rollout §2), which are
+  behavioral and deferred regardless. It lands as a follow-up once the shared library is on main.
+
+## What to Tell Your User
+
+Nothing changes for you right now — this ships **dark**, and it is maintainer-only machinery (a
+no-op on your install unless you develop instar itself). It records WHICH of my AI checks exist to
+JUDGE a completion/health claim, so the same "a claim is not evidence — show the output" bar that
+already binds what I claim can be extended to the checks that judge those claims. The judge-prompt
+wording and the registry text itself change later, and only after an operator sign-off.
+
+## Summary of New Capabilities
+
+None active for end users in this increment — everything ships dark and additive. (For instar
+maintainers: a required per-callsite `judgesClaims` classification with a shrink-only ratchet and
+a pinned JUDGE-SEED floor, so a claim-judging callsite can never silently ship un-benched against
+the bare-claim / real-evidence axis pair.)
+
+## Evidence
+
+- `npx vitest run tests/unit/judges-claims-classification-ratchet.test.ts tests/unit/llm-bench-coverage-ratchet.test.ts`
+  → **2 files, 13 tests, 0 failures** (required-explicit + no-dangling + valid-claimKind + the
+  JUDGE-SEED floor + shrink-only argued-false + the grounding canary; the pre-existing coverage
+  ratchet still green against the additive record). During authoring the ratchet caught three
+  unclassified callsites (TopicIntentArcCheck, InputClassifier, SessionSummarySentinel) and
+  failed the build until they were classified — the required-explicit guard doing its job.
+- `npx tsc --noEmit` → exit 0. `npm run lint` → exit 0.
+- Dark-by-construction: `LLM_JUDGES_CLAIMS` is build-time metadata read only by the new pinned
+  ratchet; it changes NO prompt text and wires NO runtime gate.

package/upgrades/side-effects/evidence-bar-judge-extension.md

@@ -0,0 +1,66 @@
+# Side-Effects Review — Evidence-Bar Extension (defect class 3, dark increment)
+
+**Version / slug:** `evidence-bar-judge-extension`
+**Date:** `2026-07-03`
+**Author:** `echo`
+**Tier:** `1 (dark increment — additive classification + one pinned ratchet; no runtime wiring, no operator-gated text)`
+
+## Summary of the change
+
+The mechanical arm of the "Evidence-Bar Extension to Judge Prompts" standard (defect class 3 / `claim-vs-evidence` closure; `docs/specs/evidence-bar-judge-extension.md`), shipped as a self-contained DARK increment. It ships:
+
+1. **`src/data/llmBenchCoverage.ts`** (additive) — `LLM_JUDGES_CLAIMS`, the `judgesClaims` axis of the program's ONE shared per-callsite metadata record (sibling of the authority-clause `untrustedInput` axis, co-located in the same file). `ClaimKind` = `completionClaim | healthClaim | scoredCredit`; `JudgesClaimsFlag` = `{ claimKind }` (a judge) | `false` (not a judge) | `{ false: reason }` (a judge-shaped callsite argued out of scope). Required-explicit for every `COMPONENT_CATEGORY` key.
+2. **`tests/unit/judges-claims-classification-ratchet.test.ts`** — required-explicit + no-dangling + valid-claimKind-per-judge + the spec-named JUDGE-SEED pinned as a judge + argued-false-shrink-only + real-reason floor. Same pinned-baseline family as the sibling ratchets.
+3. **`docs/specs/evidence-bar-judge-extension.md`** + **`.eli16.md`** — the converged spec + plain-English overview (carried in with the closure so the standard's authority is on canonical main).
+
+## What is DELIBERATELY out of scope (not orphan deferrals)
+
+- **The registry amendment (spec §1).** Operator-gated — ships ONLY with Justin's explicit sign-off (spec front-matter `operator-gate`; the amendment is DRAFTED in spec §1). This run does not write `docs/STANDARDS-REGISTRY.md`.
+- **The bench-axis pair ratchet (spec §3).** A bare-claim (false-accept) case + a real-evidence (false-reject) case per `judgesClaims` judge. Blocked on the program-wide "batteries readable by CI" decision — `research/` is absent from canonical main, so main's CI cannot read axis fields (`class-closure-gate.md` §"Program-shared machinery" #2, binding all three sibling axis specs). The sibling authority-clause axis ratchet deferred on the identical blocker; the keystone (#1347) staged its axis-requirements ratchet into its OWN dark increment for the same reason. Landing it under one standard's name would be doing shared cross-program infra as a side effect.
+- **The `evidenceBar()` prompt clause (spec §4).** A sibling of the authority clause, defined by the spec to live in `src/core/promptClauses.ts`. That shared library is introduced by the still-OPEN sibling PR (authority-clause, #1351); creating it here would hard-collide on the same new file and break auto-merge for whichever PR merges second. The clause is consumed only by the A/B-gated per-component migrations (spec §4 / rollout §2), which are behavioral and deferred regardless — so shipping the clause function now would add a caller-less string with a merge hazard and no live value. It lands as a follow-up once the shared library is on main.
+- **Per-component A/B clause migrations (rollout §2).** Behavioral prompt-text changes, each gated by the A/B protocol; the completion judge goes LAST (three wordings have already failed honestly — the named legitimate terminal state is "incumbent stands + routing mitigation + tracked gap"). A downstream task, not this dark increment.
+
+## Decision-point inventory
+
+- `LLM_JUDGES_CLAIMS` (`src/data/llmBenchCoverage.ts`) — **add** — build-time metadata only. Read by the new pinned ratchet; no runtime consumer. A SIGNAL producer (which callsites judge a completion/health claim), never a runtime authority.
+- `tests/unit/judges-claims-classification-ratchet.test.ts` — **add** — a CI-only pinned baseline (same family as `llm-bench-coverage-ratchet`). It gates the build (adding an unclassified LLM callsite / mis-shaping a judge entry is red CI), never a runtime path.
+
+---
+
+## 1. Over-block
+
+**What legitimate inputs does this change reject that it shouldn't?**
+
+None at runtime — nothing is wired to a runtime allow/deny path. The only new red-CI surfaces are build-time and intentional: (a) adding a new LLM component to `COMPONENT_CATEGORY` without an explicit `judgesClaims` classification, (b) shaping a judge entry with an invalid `claimKind`, (c) marking a spec-named JUDGE-SEED callsite as `false`, or (d) drifting the argued-false pinned baseline. Each is a self-describing failure with fix-instructions in the assertion message — the "visible, reviewed act" the standard exists to force, not an over-block of legitimate work.
+
+## 2. Under-block
+
+**What failure modes does this still miss?**
+
+- The classification records WHICH callsites judge claims; it does NOT yet enforce that each judge carries the bare-claim + real-evidence axis PAIR (that is the deferred bench-axis ratchet, spec §3). So a classified judge can still ship with no false-accept/false-reject coverage until that ratchet lands. This is the known staged-rollout gap, tracked in the spec rollout, not a silent miss.
+- Classification accuracy is author-judged (like the sibling `untrustedInput` axis). A judge mislabeled `false` beyond the spec-named seed would pass — the JUDGE-SEED floor + the argued-false shrink-only pin catch the measured judges and every judge-shaped exemption; the seeding PR IS that review.
+- A prompt bar cannot AUTHENTICATE material — a fabricated transcript showing a fake PASS passes any in-prompt bar (spec "Honest reach"). The authoritative verification arm is the deterministic real-check `verification_command`, which is out of this classification by construction and named in the spec as the true-verification owner.
+
+## 3. Level-of-abstraction fit
+
+**Is this at the right layer?**
+
+Yes. The classification extends the ONE shared per-callsite metadata record the program mandates (`src/data/llmBenchCoverage.ts`), not a new parallel registry — exactly where the sibling `untrustedInput` axis lives; the ratchet is in the established pinned-baseline vitest family. No higher gate already owns "which callsites judge a completion claim"; no lower primitive is duplicated.
+
+## 4. Signal vs authority compliance
+
+`LLM_JUDGES_CLAIMS` and the ratchet are SIGNALS (which callsites judge a claim; is each judge classified and axis-eligible) — they inform the build and the reviewer, never grant or deny a runtime action. No model-produced field is wired to satisfy any check. The (future, deferred) prompt clause, once consumed, would make a model REPORT that a bare claim is unverified rather than CREDIT it — it never becomes the authority; the deterministic real-check verifier owns what is TRUE.
+
+## 5. Interactions with existing systems
+
+- **`llm-bench-coverage-ratchet`** — unaffected: `LLM_JUDGES_CLAIMS` is a NEW additive export; the existing `LLM_BENCH_COVERAGE` union and its ratchet are untouched (both green post-change under a bounded run).
+- **`untrusted-input-classification-ratchet` / `promptClauses.ts`** (sibling authority-clause, PR #1351, still open) — this increment adds a PARALLEL axis export in the same file and a parallel ratchet; it does not import or depend on the authority-clause work. Both append to `LLM_BENCH_COVERAGE`'s file region, so a textual rebase against #1351 may be needed depending on merge order — an additive, mechanical resolve, no semantic conflict.
+- **`class-closure-lint.mjs`** — already names `src/data/llmBenchCoverage.ts` as an agent-authored artifact; touching it routes the PR to operator review (the lint is report-only today). The defect class `claim-vs-evidence` already carries `closureStandard: evidence-bar-judge-extension` on master — untouched here.
+
+## 6. Failure modes / rollback
+
+Pure additive TypeScript + one test + spec docs. Rollback = revert the commit; nothing persists state, nothing runs at runtime, no migration, no config key (spec §"Decision points touched": none at runtime; no agent-side config key exists or is wanted — repo posture only, so no Migration Parity work). tsc clean, `npm run lint` exit 0, both ratchets green under bounded single-file runs (machine under intermittent external CPU pressure; verification used bounded vitest per the CI-as-gate pattern, full suite gated by CI).
+
+## 7. Second-pass reviewer
+
+Self-review (bounded machine-pressure build). The change is additive, dark, and test-pinned; no runtime surface. Key self-checks: (a) confirmed the JUDGE-SEED is exactly the spec §2 named LLM judges (real-check verifier correctly excluded as the deterministic arm); (b) the two stall-confirm adapters (Telegram/Slack) classify true by the spec's stated inclusion criteria (stall/health classifiers), the safe over-inclusive direction; (c) the required-explicit ratchet actually fired during authoring (caught 3 unclassified callsites), proving it is not a no-op; (d) argued-false baseline is empty and the type still supports it for future judge-shaped exemptions.