instar 1.3.830 → 1.3.832

package/dist/data/provenanceCoverage.js

@@ -0,0 +1,674 @@
+/**
+ * PROVENANCE_COVERAGE census — LLM-Decision Quality Meter ratchet (G5).
+ *
+ * Spec: docs/specs/llm-decision-quality-meter.md §5.6 (census) + §5.4.2 (rule
+ * registry). Same declare-or-fail pattern as LLM_BENCH_COVERAGE
+ * (src/data/llmBenchCoverage.ts precedent), tightened per review: a declaration
+ * per DECISION POINT (a component may hold several distinct decision points
+ * with different prompts/outcomes), each entry one of:
+ *   - status 'wired'                — the callsite carries `options.provenance`
+ *     enrollment (typed import of the decision-point id exported HERE; the
+ *     settlement write additionally validates decisionPoint ∈ census at
+ *     runtime and counts unknowns). Wired entries declare their volumeClass —
+ *     the PROVENANCE store's volume valve (§5.6; the ~250-byte decision_quality
+ *     row is written for every enrolled settlement regardless of class).
+ *   - status 'pending:<ACT-ref>'    — the retrofit backlog, format-validated +
+ *     PINNED shrink-only in tests/unit/provenance-coverage-ratchet.test.ts
+ *     (count can only go down; re-pointing an entry to a different ACT is a
+ *     reviewed baseline change — shrink-only covers count, not identity). The
+ *     runtime half of the two-layer check (§5.6 pending-ref-dead) lives on
+ *     GET /decision-quality, where the evolution queue exists.
+ *   - status 'exempt:<taxonomy>'    — a CLOSED taxonomy (an exemption is a
+ *     classification, not an essay): 'deterministic-only' (no LLM verdict at
+ *     this point) | 'no-decision-content' (nothing reconstructable beyond what
+ *     feature_metrics already records) | 'operator-ratified:<resolvable-ref>'.
+ *     Free-text exemptions are refused by the ratchet; the exempt baseline is
+ *     pinned shrink-only exactly like pending (ADV r5).
+ *
+ * ENROLLMENT KEY CONVENTION (§5.6, a census-test ASSERTION, not prose): each
+ * decision point uses a 1:1 `attribution.component` key (the existing
+ * `CompletionEvaluator` vs `CompletionEvaluator/P13` suffix pattern); the key
+ * is UNIQUE across census entries REGARDLESS of status (ADV r7 — uniform
+ * uniqueness closes the pending-absorbs-wired-activity and exempt-false-flags
+ * carve-out attacks). Multi-call compositions get one unique suffixed key PER
+ * point with linkage ONLY via the `composition` field (§5.1.1 — key sharing
+ * would re-open the same-key blind spot).
+ *
+ * WHY (operator goal, 2026-07-10 topic 11960): a new LLM decision point that
+ * skips provenance must fail CI — an unlogged decision-maker cannot be graded,
+ * and "does this gate need a bigger model or a prompt change?" is unanswerable
+ * without provenance + outcomes. Structure > Willpower.
+ *
+ * Companion chain: componentCategories keeps COMPONENT_CATEGORY exhaustive
+ * over LLM callsites; LLM_BENCH_COVERAGE keeps bench coverage exhaustive over
+ * COMPONENT_CATEGORY; THIS census keeps provenance posture exhaustive over
+ * COMPONENT_CATEGORY, per decision point.
+ */
+export const EXEMPT_TAXONOMY_KEYS = [
+    'deterministic-only',
+    'no-decision-content',
+];
+// ───────────────────────────────────────────────────────────────────────────
+// Wired decision-point id constants (typed registration, §5.6).
+//
+// NAMING CONVENTION (ratchet-enforced): a wired decision point 'a-b-c' exports
+// `DP_A_B_C`. Enrolling callsites import the constant — never restate the
+// string (the census is the single source of truth).
+// ───────────────────────────────────────────────────────────────────────────
+/** External-hog kill/leave decision point — the classifier verdict inside the
+ * scan-tick decision loop (ExternalHogScanTick; spec §5.3 first customer). */
+export const DP_EXTERNAL_HOG_KILL_LEAVE = 'external-hog-kill-leave';
+/** Autonomous completion judge — CompletionEvaluator.evaluate() (spec §5.3). */
+export const DP_COMPLETION_EVALUATE = 'completion-evaluate';
+/** P13 stop-rationale judge — CompletionEvaluator.evaluateStopRationale()
+ * (component 'CompletionEvaluator/P13'; spec §5.3). */
+export const DP_COMPLETION_STOP_RATIONALE = 'completion-stop-rationale';
+/** Outbound tone/leak verdict — MessagingToneGate.review() (component
+ * 'MessagingToneGate'; spec §5.6). An ALWAYS-ON HIGH-VOLUME gate: it declares a
+ * `budget:<rows/day>` volume valve (NEVER `full`) and stores content as IDENTITY
+ * only (candidate hash + bounds + code-derived features), never the outbound
+ * body or any plaintext slice of it. */
+export const DP_MESSAGING_TONE_GATE = 'messaging-tone-gate';
+// ───────────────────────────────────────────────────────────────────────────
+// The census
+// ───────────────────────────────────────────────────────────────────────────
+export const PROVENANCE_COVERAGE = [
+    // ── Wired first customers (§5.3 — genuinely low-frequency + high-stakes;
+    //    volumeClass 'full' is RESERVED for this class) ──────────────────────
+    {
+        decisionPoint: DP_EXTERNAL_HOG_KILL_LEAVE,
+        component: 'ExternalHogClassifier',
+        status: 'wired',
+        volumeClass: 'full',
+        // §5.3: judges an attacker-controllable process name (argv is HASHED at the
+        // envelope — the floor needs argv, the provenance row does not); context =
+        // commandHash/ledgerKey/classId, floor booleans, CPU numbers, identity tuples.
+        contentClass: 'content-bearing',
+        reason: 'First customer (spec §5.3): the kill/leave verdict is the highest-consequence LLM decision on the host; enacted disposition + evidence rules grade it.',
+    },
+    {
+        decisionPoint: DP_COMPLETION_EVALUATE,
+        component: 'CompletionEvaluator',
+        status: 'wired',
+        volumeClass: 'full',
+        // §5.3: context carries transcript-slice IDENTITY (hash + bounds) + the
+        // StopSignals corroboration block — never transcript text.
+        contentClass: 'content-bearing',
+        reason: 'First customer (spec §5.3): the autonomous continue/stop judge gates whether a run keeps burning budget; realcheck gives it deterministic ground truth.',
+    },
+    {
+        decisionPoint: DP_COMPLETION_STOP_RATIONALE,
+        component: 'CompletionEvaluator/P13',
+        status: 'wired',
+        volumeClass: 'full',
+        contentClass: 'content-bearing',
+        reason: 'First customer (spec §5.3): the P13 stop-rationale judge decides whether a stop-attempt is EARNED; same transcript-slice-identity envelope as evaluate().',
+    },
+    // ── Wired high-volume gate (§5.6 — NOT full-class; the third enrolled
+    //    customer, which required the grading-pass per-point sub-budget FIRST —
+    //    SUBBUDGET_IMPLEMENTED is now true) ─────────────────────────────────────
+    {
+        decisionPoint: DP_MESSAGING_TONE_GATE,
+        component: 'MessagingToneGate',
+        status: 'wired',
+        // §5.6 volume valve: an ALWAYS-ON high-volume gate MUST NOT be `full`. A
+        // per-UTC-day COUNT budget gives a hard, count-enforced ceiling on the
+        // provenance JSONL archive (loud droppedByBudget counter when hit) — a
+        // deterministic bound preferable to probabilistic sampling for a gate that
+        // fires on every drafted outbound message. 500/day = a representative daily
+        // sample without unbounded growth; the ~250-byte decision_quality row is
+        // ALWAYS written regardless (counts stay complete).
+        volumeClass: 'budget:500',
+        // Content-bearing: the gate judges an agent-authored outbound message. It
+        // enters the row as IDENTITY ONLY — a sha256 of the candidate + byte/char
+        // bounds + code-derived features — never the full body or any plaintext
+        // slice (the provenance store must not become an outbound-message archive;
+        // mirrors the CompletionEvaluator content-bearing sibling, §5.3).
+        contentClass: 'content-bearing',
+        reason: 'The outbound tone/leak authority (spec §5.6 named high-volume point). Enrolled at budget:500/day, identity-only content — never the message body.',
+    },
+    // ── Pending (the ACT-1193 uniform-provenance retrofit backlog — §5.6: "Not
+    //    retrofitting all ~60+ decision points in one PR"; the census makes the
+    //    backlog visible, ratcheted, and re-surfaced as census debt on
+    //    GET /decision-quality). Each enrolls via the §5.1.4 per-callsite
+    //    contract and declares its REAL volume class at enrollment.
+    //    contentClass mirrors the reviewed LLM_UNTRUSTED_INPUT axis
+    //    (src/data/llmBenchCoverage.ts): every point below judges user/model/
+    //    tool-authored text → content-bearing. ─────────────────────────────────
+    // — Sentinels —
+    {
+        decisionPoint: 'input-guard',
+        component: 'InputGuard',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Input-coherence verdict over an inbound prompt; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'session-activity-digest',
+        component: 'SessionActivitySentinel',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Activity digest authored over session tmux output; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'stall-triage-diagnosis',
+        component: 'StallTriageNurse',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Stall-triage diagnosis over session output; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'commitment-detect',
+        component: 'CommitmentSentinel',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Commitment detection over conversation text; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'presence-stall-judge',
+        component: 'PresenceProxy',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Tier-3 stall judgment over session output; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'message-sentinel-classify',
+        component: 'MessageSentinel',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Pause/emergency/normal intent classification over an inbound user message (latency-critical); enrollment queued in the ACT-1193 retrofit backlog.',
+    },
+    {
+        decisionPoint: 'project-drift-check',
+        component: 'ProjectDriftChecker',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Is-work-on-project coherence verdict over session work + files; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'temporal-coherence-check',
+        component: 'TemporalCoherenceChecker',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Temporal-coherence verdict over conversation content; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'watchdog-stuck-judge',
+        component: 'SessionWatchdog',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Stuck-session judgment over live session output; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'resume-sanity-check',
+        component: 'ResumeQueueDrainer',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Resume-sanity verdict before a queued mid-work revival; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'topic-intent-arc-check',
+        component: 'TopicIntentArcCheck',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Arc-check classification of a topic intent over conversation; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'slack-stall-confirm',
+        component: 'SlackAdapter',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Stall-confirm alert-suppression judgment over session output (Slack arm); enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    // — Gates —
+    {
+        decisionPoint: 'prompt-injection-detect',
+        component: 'PromptGate',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Prompt-injection detection over inbound content; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'external-operation-gate',
+        component: 'ExternalOperationGate',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Operation mutability/reversibility classification incl. in-content approval claims; enrollment queued in the ACT-1193 retrofit backlog.',
+    },
+    {
+        decisionPoint: 'warrants-reply-gate',
+        component: 'WarrantsReplyGate',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Should-I-reply verdict over an inbound message; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'unjustified-stop-gate',
+        component: 'UnjustifiedStopGate',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Stop-justified verdict over session state; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'coherence-review',
+        component: 'CoherenceReviewer',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Outbound coherence review — THE measured high-volume point (3,641 of 4,098 llm calls/24h on the dev agent, spec §5.6); MUST declare sampled:<rate> or budget:<rows/day> at enrollment, never full.',
+    },
+    {
+        decisionPoint: 'move-intent-classify',
+        component: 'MoveIntentClassifier',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Move/pin command-vs-discussion intent over an inbound message + context; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'hub-intent-classify',
+        component: 'HubIntentClassifier',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Hub open/tie bind-intent over an inbound hub message; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'profile-intent-classify',
+        component: 'ProfileIntentClassifier',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Topic-profile change intent (framework/model/thinking) over an inbound message; enrollment queued in the ACT-1193 retrofit backlog.',
+    },
+    {
+        decisionPoint: 'llm-sanitize',
+        component: 'LLMSanitizer',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Sanitize verdict over untrusted inbound content (definitionally injection-exposed); enrollment queued in the ACT-1193 retrofit backlog.',
+    },
+    {
+        decisionPoint: 'override-detect',
+        component: 'OverrideDetector',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Override-intent detection over a user turn (uxConfirm pre-routing); enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'task-classify',
+        component: 'TaskClassifier',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Task-type classification over a user task description; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    // — Reflectors —
+    {
+        decisionPoint: 'job-reflect',
+        component: 'JobReflector',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Job-outcome reflection over job output/transcript; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'cross-model-review',
+        component: 'crossModelReviewer',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Cross-model spec-document review over file content; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'self-knowledge-extract',
+        component: 'SelfKnowledgeTree',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Self-knowledge extraction over transcripts; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'tree-triage',
+        component: 'TreeTriage',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Knowledge-tree fragment triage over stored content; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'topic-summarize',
+        component: 'TopicSummarizer',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Topic summary authoring over conversation content; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'contextual-evaluate',
+        component: 'ContextualEvaluator',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Context-relevance evaluation over conversation/session content; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'relationship-extract',
+        component: 'RelationshipManager',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Relationship-fact extraction from conversation (PII-adjacent content); enrollment queued in the ACT-1193 retrofit backlog.',
+    },
+    {
+        decisionPoint: 'standards-conformance-review',
+        component: 'StandardsConformanceReviewer',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Artifact-vs-standard conformance review over file content; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'discovery-evaluate',
+        component: 'DiscoveryEvaluator',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Serendipity-discovery evaluation over subagent output; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'dashboard-insight',
+        component: 'DashboardInsightEngine',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Awareness-only page-data insight authoring (degrades to a deterministic floor); enrollment queued in the ACT-1193 retrofit backlog.',
+    },
+    // — Jobs —
+    {
+        decisionPoint: 'pipe-session-spawn',
+        component: 'PipeSessionSpawner',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Session authoring from (possibly user-authored) task descriptions; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'cartographer-summary-author',
+        component: 'CartographerSweep',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Doc-tree summary authoring over untrusted code; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'standards-coverage-enrich',
+        component: 'StandardsCoverageEnrichment',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Standards-coverage row enrichment over repo content (dark LLM path); enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    // — Previously-uncategorized callsites (LLM Routing Registry audit set) —
+    {
+        decisionPoint: 'input-classify',
+        component: 'InputClassifier',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Auto-approve vs relay classification of inbound input; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'session-summary-extract',
+        component: 'SessionSummarySentinel',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Task/phase/files extraction over tmux output; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'telegram-stall-confirm',
+        component: 'TelegramAdapter',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Stall-confirm alert-suppression judgment over session output (Telegram arm); enrollment queued in the ACT-1193 retrofit backlog.',
+    },
+    {
+        decisionPoint: 'resume-uuid-validate',
+        component: 'ResumeValidator',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Resume-UUID-vs-topic match verdict over session/resume state; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'usher-topic-route',
+        component: 'Usher',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Per-turn topic routing over an inbound user turn; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'topic-intent-extract',
+        component: 'TopicIntentExtractor',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Topic-intent extraction from a user turn; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'pre-compaction-flush',
+        component: 'PreCompactionFlush',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Durable-fact extraction over a transcript before compaction; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'tree-synthesize',
+        component: 'TreeSynthesis',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Knowledge-fragment synthesis into an answer; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'llm-conflict-resolve',
+        component: 'LLMConflictResolver',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Divergent multi-machine state resolution over untrusted peer data; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'open-conversation-brief',
+        component: 'openConversationBrief',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'A2A conversation-brief authoring over peer content; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'a2a-checkin-summarize',
+        component: 'a2a-checkin',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'A2A check-in thread summarization over peer-authored threads; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'correction-distill',
+        component: 'correction-learning',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Recurring-correction distillation into a durable preference; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    {
+        decisionPoint: 'mentor-stage-b-classify',
+        component: 'mentor-stage-b',
+        status: 'pending:ACT-1193',
+        contentClass: 'content-bearing',
+        reason: 'Mentor-signal classification over mentee output; enrollment queued in the ACT-1193 uniform-provenance retrofit backlog.',
+    },
+    // ── Argued exemptions (closed taxonomy; pinned shrink-only) ──────────────
+    {
+        decisionPoint: 'input-detector-alias',
+        component: 'InputDetector',
+        status: 'exempt:deterministic-only',
+        contentClass: 'metadata',
+        reason: 'Attribution-manifest alias only (a legacy prompt-pattern matcher) — no LLM verdict at this point; the live matcher calls with attribution PromptGate, declared there.',
+    },
+    {
+        decisionPoint: 'auto-approve-injection',
+        component: 'AutoApprover',
+        status: 'exempt:deterministic-only',
+        contentClass: 'metadata',
+        reason: 'Mechanical key injection + audit logging — no LLM verdict at this point; the upstream judgment is InputClassifier.classify(), declared as input-classify.',
+    },
+    {
+        decisionPoint: 'integration-gate-delegate',
+        component: 'IntegrationGate',
+        status: 'exempt:deterministic-only',
+        contentClass: 'metadata',
+        reason: 'No LLM prompt of its own — delegates to JobReflector.reflect() (attribution JobReflector, declared as job-reflect); zero LLM-provider callsites of its own.',
+    },
+    {
+        decisionPoint: 'coherence-gate-delegate',
+        component: 'CoherenceGate',
+        status: 'exempt:deterministic-only',
+        contentClass: 'metadata',
+        reason: 'No callsite carries attribution CoherenceGate — all LLM calls flow through CoherenceReviewer.callApi(), declared as coherence-review.',
+    },
+    {
+        decisionPoint: 'promise-beacon-status-line',
+        component: 'PromiseBeacon',
+        status: 'exempt:deterministic-only',
+        contentClass: 'metadata',
+        reason: 'No live LLM prompt — generateStatusLine/classifyProgress hooks are unwired at the construction site; no LLM verdict exists at this point. Revisit if a generator is wired.',
+    },
+    {
+        decisionPoint: 'interactive-pool-canary-judge',
+        component: 'InteractivePoolCanaryJudge',
+        status: 'exempt:no-decision-content',
+        contentClass: 'metadata',
+        reason: 'Judges a FIXED known-answer canary probe — the input is a constant, so nothing is reconstructable beyond what feature_metrics already records (the canary is its own provenance). NOT deterministic-only: it legitimately emits llm-kind metric rows.',
+    },
+];
+// ───────────────────────────────────────────────────────────────────────────
+// Grading-pass fairness marker (§5.5 / LES r6).
+//
+// The grading endpoint's per-pass bound was GLOBAL, not per-point — safe for the
+// two seeded low-frequency full-class customers, but a third ENROLLED customer
+// could starve sibling points' evidence windows. The census ratchet asserts
+// structurally that enrolling beyond the seeded first-customer set requires the
+// per-point round-robin sub-budget FIRST: this is now true because
+// `runDecisionGradingPass` divides its global budget round-robin across the
+// grade-pass-driven points (src/core/decisionGradingPass.ts — the sub-budget
+// helper `perPointSubBudget`), so no single point can consume a whole pass and
+// starve a sibling's maturing evidence window. Flipped in the PR that
+// implements that sub-budget (MessagingToneGate enrollment, the third customer).
+// ───────────────────────────────────────────────────────────────────────────
+export const SUBBUDGET_IMPLEMENTED = true;
+// ───────────────────────────────────────────────────────────────────────────
+// Evidence-rule registry (§5.4.2) — ruleId → rung + evidence-strength + OWNING
+// component (+ registered window parameter). Co-located with the census by
+// spec: imported by the annotate chokepoint and the grading endpoint; the
+// ratchet pins the enums and the existing rule identities.
+//
+// Rung is DERIVED from this registry, never caller-supplied: an annotation
+// claiming a ruleId whose registered rung disagrees, or an unregistered
+// ruleId, is REJECTED and counted (§5.4.2). The chokepoint also rejects an
+// annotation whose gradedBy.component is not the ruleId's registered owner
+// (ADV r5 — a confused in-process annotator cannot inherit another rule's
+// rung/precedence by claiming its id).
+//
+// Rules are precise predicates with IMMUTABLE, VERSIONED ids (§5.4.5): a
+// predicate change — or a change to a registered parameter like windowMs —
+// mints a new ruleId ('-v2'), never mutates '-v1' in place.
+// ───────────────────────────────────────────────────────────────────────────
+/** Grading-ladder rungs, in PRECEDENCE ORDER (§5.4.3): earlier beats later.
+ * A self-reported outcome NEVER overrides an independent grader. */
+export const EVIDENCE_RUNGS = [
+    'deterministic-ground-truth',
+    'recurrence',
+    'llm-interpreter', // DORMANT this build (FD11) — no rule may register it until ACT-1198's preconditions land.
+    'self-report',
+];
+/** Evidence-strength classes (§5.4.2, codex r3): the read surface splits
+ * proof-like from heuristic grades so aggregates cannot imply stronger
+ * correctness than the evidence supports. */
+export const EVIDENCE_STRENGTHS = [
+    'deterministic-proof',
+    'negative-evidence',
+    'recurrence-proxy',
+    'self-report',
+];
+/** Default hog evidence window (§5.4.5 "bounded window (default 6h)"). */
+const HOG_EVIDENCE_WINDOW_MS = 6 * 60 * 60 * 1000;
+export const RULE_REGISTRY = {
+    // A kill graded `wrong` ONLY IF a same-commandHash candidate respawns in-window
+    // AND the kill-time ordering test re-runs TRUE at evidence time (§5.4.5).
+    // Positive-evidence grading runs in the sentinel's scan ticks + grade-on-supersede.
+    'hog-respawn-wrong-v1': {
+        ruleId: 'hog-respawn-wrong-v1',
+        rung: 'deterministic-ground-truth',
+        evidenceStrength: 'deterministic-proof',
+        owningComponent: 'ExternalHogSentinel',
+        windowMs: HOG_EVIDENCE_WINDOW_MS,
+    },
+    // A kill whose commandHash does NOT re-flag in-window (owner recorded dead at
+    // kill time) grades `right` at window close — negative evidence, never proof
+    // (quiet respawns are invisible to the sustained-CPU sensor; §5.4.5). Window-
+    // close grading runs in the grading job reading the durable hog store.
+    'hog-sustained-right-v1': {
+        ruleId: 'hog-sustained-right-v1',
+        rung: 'deterministic-ground-truth',
+        evidenceStrength: 'negative-evidence',
+        owningComponent: 'DecisionGrading',
+        windowMs: HOG_EVIDENCE_WINDOW_MS,
+    },
+    // Applies ONLY to verdict==='leave' AND enacted==='alert-only-model-spared'
+    // AND floorPermitted===true: the SAME PROCESS (targetTuple pid+start-time)
+    // re-flagging in-window grades the leave `wrong`; a same-commandHash
+    // DIFFERENT process grades `unknown` (§5.4.5). Re-flag detection runs in the
+    // sentinel's scan ticks + grade-on-supersede.
+    'hog-leave-recurrence-v1': {
+        ruleId: 'hog-leave-recurrence-v1',
+        rung: 'recurrence',
+        evidenceStrength: 'recurrence-proxy',
+        owningComponent: 'ExternalHogSentinel',
+        windowMs: HOG_EVIDENCE_WINDOW_MS,
+    },
+    // met:true + realcheck pass → right; met:true + realcheck fail → wrong; no
+    // realcheck configured → unknown, never guessed (§5.3/§5.4.5). The annotator
+    // is the deterministic realcheck arm of the autonomous completion path (P8
+    // wiring binds gradedBy.component to this owner).
+    'completion-realcheck-v1': {
+        ruleId: 'completion-realcheck-v1',
+        rung: 'deterministic-ground-truth',
+        evidenceStrength: 'deterministic-proof',
+        owningComponent: 'AutonomousRealCheck',
+    },
+    // §5.3 enacted-disposition self-reports: the deterministic actor that applied
+    // floors/breakers/governors records what was ACTUALLY enacted, immediately,
+    // as a self-report-rung annotation (never overrides an independent grader).
+    'hog-enacted-disposition-v1': {
+        ruleId: 'hog-enacted-disposition-v1',
+        rung: 'self-report',
+        evidenceStrength: 'self-report',
+        owningComponent: 'ExternalHogSentinel',
+    },
+    'completion-enacted-disposition-v1': {
+        ruleId: 'completion-enacted-disposition-v1',
+        rung: 'self-report',
+        evidenceStrength: 'self-report',
+        owningComponent: 'CompletionChokepoint',
+    },
+};
+// ───────────────────────────────────────────────────────────────────────────
+// Lookup helpers (imported by the settlement seam, the annotate chokepoint,
+// the grading endpoint, and the read surface).
+// ───────────────────────────────────────────────────────────────────────────
+const CENSUS_BY_POINT = new Map(PROVENANCE_COVERAGE.map((e) => [e.decisionPoint, e]));
+const CENSUS_BY_COMPONENT = new Map(PROVENANCE_COVERAGE.map((e) => [e.component, e]));
+/** The census entry for a decision point, or undefined (an unknown decision
+ * point at the settlement write is counted, never thrown — §5.6). */
+export function getCensusEntry(decisionPoint) {
+    return CENSUS_BY_POINT.get(decisionPoint);
+}
+/** The census entry keyed by the 1:1 `attribution.component` enrollment key
+ * (the bridge the wired-but-silent / exempt-but-active runtime flags use). */
+export function getCensusEntryByComponent(componentKey) {
+    return CENSUS_BY_COMPONENT.get(componentKey);
+}
+/** True iff the decision point is declared AND wired (enrolled). Pending and
+ * exempt points are NOT enrolled — the seam writes nothing for them. */
+export function isEnrolled(decisionPoint) {
+    return CENSUS_BY_POINT.get(decisionPoint)?.status === 'wired';
+}
+/** The volume class governing the provenance JSONL row for an ENROLLED point
+ * (undefined for unknown/pending/exempt points — a forward-declared class on a
+ * pending entry is advisory and must not valve anything before enrollment). */
+export function getVolumeClass(decisionPoint) {
+    const e = CENSUS_BY_POINT.get(decisionPoint);
+    return e?.status === 'wired' ? e.volumeClass : undefined;
+}
+/** The registered evidence rule for a ruleId, or undefined (the annotate
+ * chokepoint REJECTS and counts annotations claiming an unregistered id). */
+export function getRule(ruleId) {
+    return RULE_REGISTRY[ruleId];
+}
+//# sourceMappingURL=provenanceCoverage.js.map