@iworkforces/tracelattice 2.1.2 → 2.2.1

package/dist/core/evaluator/Calibrator.d.ts

@@ -2,21 +2,25 @@
  * Confidence calibration for sequential thinking reasoning.
  *
  * Provides the {@link Calibrator} class — maps raw model confidence values
- * to calibrated probabilities using per-type Beta(2, 2) priors and
+ * to calibrated probabilities using per-type empirical evidence blending and
  * grid-searched temperature scaling. Reports calibration quality through
  * Brier score and 10-bin Expected Calibration Error (ECE).
  *
  * Math summary:
- * - **Per-type prior**: Beta(α=2, β=2), prior mean = 0.5. As outcomes accumulate
- *   for a type, the calibrated value shrinks toward the observed mean of that
- *   type with weight `priorWeight = 1 / (1 + n / 10)` where `n` is the type's
- *   sample count.
+ * - **Per-type evidence blend**: Raw confidence is the prior signal, weighted by
+ *   `priorWeight = 10 / (10 + n)` where `n` is the type's outcome count. The
+ *   remaining weight is applied to the observed per-type mean. An empty type
+ *   defaults its mean to 0.5 but has zero empirical weight, so cold start returns raw.
  * - **Temperature scaling**: Grid search T ∈ {0.5, 0.75, 1.0, 1.25, 1.5, 2.0}
- *   minimizing negative log-likelihood (NLL) over recorded outcomes. Default T=1.0.
- *   Requires ≥10 outcomes; below that threshold only prior shrinkage is applied.
- * - **Brier score**: `mean((predicted - actual)^2)` over recorded outcomes.
+ *   minimizing negative log-likelihood (NLL) over per-type leave-one-out evidence blends.
+ *   Exact ties retain T=1.0, then follow grid order. Requires at least 10 outcomes; below that
+ *   threshold the evidence blend is returned unchanged.
+ * - **Brier score**: `mean((predicted - actual)^2)` over stored raw outcomes.
  * - **ECE (10-bin)**: bucket predictions in 0.1 increments, compute weighted
- *   absolute deviation between bin mean confidence and bin accuracy.
+ *   absolute deviation between raw bin mean confidence and bin accuracy.
+ *
+ * Stored outcomes and metrics remain raw. This transform does not guarantee factuality, better
+ * metrics for every dataset, or monotonic per-example changes.
  *
  * @module core/evaluator/Calibrator
  */

package/dist/core/evaluator/Calibrator.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"Calibrator.d.ts","sourceRoot":"","sources":["../../../src/core/evaluator/Calibrator.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AAEH,OAAO,KAAK,EACX,kBAAkB,EAClB,iBAAiB,EACjB,WAAW,EACX,MAAM,+BAA+B,CAAC;AACvC,OAAO,KAAK,EAAE,gBAAgB,EAAuB,MAAM,+BAA+B,CAAC;AAC3F,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,oCAAoC,CAAC;AACtE,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,wBAAwB,CAAC;AAyHxD;;;;;;;;;;;;;GAaG;AACH,qBAAa,UAAW,YAAW,WAAW;IAC7C,SAAgB,OAAO,EAAE,OAAO,CAAC;IACjC,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAmB;IAC7C,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAgC;gBAElD,eAAe,EAAE,gBAAgB,EAAE,OAAO,EAAE,OAAO;IAKxD,SAAS,CACf,aAAa,EAAE,MAAM,EACrB,IAAI,EAAE,WAAW,EACjB,SAAS,EAAE,SAAS,GAClB,iBAAiB;IAoBb,OAAO,CAAC,SAAS,CAAC,EAAE,SAAS,GAAG,kBAAkB;IAclD,KAAK,CAAC,SAAS,EAAE,SAAS,GAAG,IAAI;IAKjC,YAAY,CAAC,SAAS,EAAE,SAAS,GAAG,IAAI;IAKxC,QAAQ,IAAI,IAAI;CAIvB"}
+{"version":3,"file":"Calibrator.d.ts","sourceRoot":"","sources":["../../../src/core/evaluator/Calibrator.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;GAyBG;AAEH,OAAO,KAAK,EACX,kBAAkB,EAClB,iBAAiB,EACjB,WAAW,EACX,MAAM,+BAA+B,CAAC;AACvC,OAAO,KAAK,EAAE,gBAAgB,EAAuB,MAAM,+BAA+B,CAAC;AAC3F,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,oCAAoC,CAAC;AACtE,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,wBAAwB,CAAC;AAoIxD;;;;;;;;;;;;;GAaG;AACH,qBAAa,UAAW,YAAW,WAAW;IAC7C,SAAgB,OAAO,EAAE,OAAO,CAAC;IACjC,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAmB;IAC7C,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAgC;gBAElD,eAAe,EAAE,gBAAgB,EAAE,OAAO,EAAE,OAAO;IAKxD,SAAS,CACf,aAAa,EAAE,MAAM,EACrB,IAAI,EAAE,WAAW,EACjB,SAAS,EAAE,SAAS,GAClB,iBAAiB;IAmBb,OAAO,CAAC,SAAS,CAAC,EAAE,SAAS,GAAG,kBAAkB;IAclD,KAAK,CAAC,SAAS,EAAE,SAAS,GAAG,IAAI;IAMjC,YAAY,CAAC,SAAS,EAAE,SAAS,GAAG,IAAI;IAKxC,QAAQ,IAAI,IAAI;CAIvB"}

package/dist/core/evaluator/Calibrator.js

@@ -1,4 +1,4 @@
-import { EPSILON, MIN_OUTCOMES_FOR_TEMPERATURE, applyTemperature, fitTemperature } from "./calibration-math.js";
+import { EPSILON, MIN_OUTCOMES_FOR_TEMPERATURE, applyTemperature, blendWithRawPrior, fitTemperature } from "./calibration-math.js";
 import { ALL_THOUGHT_TYPES } from "./internals.js";
 const ECE_BINS = 10;
 function aggregatePerType(outcomes) {
@@ -12,12 +12,26 @@ function aggregatePerType(outcomes) {
         prev.count += 1;
         sums.set(o.type, prev);
     }
-    const result = new Map();
-    for (const [type, { sum, count }] of sums)result.set(type, {
-        mean: sum / count,
-        count
+    return sums;
+}
+function leaveOneOutBlendedOutcomes(outcomes) {
+    const perType = aggregatePerType(outcomes);
+    return outcomes.map((outcome)=>{
+        const typeStats = perType.get(outcome.type) ?? {
+            sum: outcome.actual,
+            count: 1
+        };
+        const sampleCount = typeStats.count - 1;
+        if (0 === sampleCount) return {
+            ...outcome
+        };
+        const empiricalMean = (typeStats.sum - outcome.actual) / sampleCount;
+        const { blended } = blendWithRawPrior(outcome.predicted, empiricalMean, sampleCount);
+        return {
+            ...outcome,
+            predicted: blended
+        };
     });
-    return result;
 }
 function brierScore(outcomes) {
     if (0 === outcomes.length) return null;
@@ -92,11 +106,10 @@ class Calibrator {
         const perType = aggregatePerType(outcomes);
         const typeStats = perType.get(type);
         const n = typeStats?.count ?? 0;
-        const observedMean = typeStats?.mean ?? 0.5;
-        const priorWeight = 1 / (1 + n / 10);
-        const shrunk = priorWeight * observedMean + (1 - priorWeight) * raw;
+        const observedMean = void 0 === typeStats ? 0.5 : typeStats.sum / typeStats.count;
+        const { blended, priorWeight } = blendWithRawPrior(raw, observedMean, n);
         const temperature = this._temperatures.get(sessionId) ?? 1.0;
-        const calibrated = outcomes.length >= MIN_OUTCOMES_FOR_TEMPERATURE ? applyTemperature(shrunk, temperature) : shrunk;
+        const calibrated = outcomes.length >= MIN_OUTCOMES_FOR_TEMPERATURE ? applyTemperature(blended, temperature) : blended;
         return {
             raw,
             calibrated,
@@ -116,7 +129,8 @@ class Calibrator {
     }
     refit(sessionId) {
         if (!this.enabled) return;
-        this._temperatures.set(sessionId, fitTemperature(this._recorder.getOutcomes(sessionId)));
+        const outcomes = this._recorder.getOutcomes(sessionId);
+        this._temperatures.set(sessionId, fitTemperature(leaveOneOutBlendedOutcomes(outcomes)));
     }
     clearSession(sessionId) {
         if (!this.enabled) return;

package/dist/core/evaluator/Calibrator.js.map

@@ -1 +1 @@
-{"version":3,"file":"core/evaluator/Calibrator.js","sources":["../../../src/core/evaluator/Calibrator.ts"],"sourcesContent":["/**\n * Confidence calibration for sequential thinking reasoning.\n *\n * Provides the {@link Calibrator} class — maps raw model confidence values\n * to calibrated probabilities using per-type Beta(2, 2) priors and\n * grid-searched temperature scaling. Reports calibration quality through\n * Brier score and 10-bin Expected Calibration Error (ECE).\n *\n * Math summary:\n * - **Per-type prior**: Beta(α=2, β=2), prior mean = 0.5. As outcomes accumulate\n *   for a type, the calibrated value shrinks toward the observed mean of that\n *   type with weight `priorWeight = 1 / (1 + n / 10)` where `n` is the type's\n *   sample count.\n * - **Temperature scaling**: Grid search T ∈ {0.5, 0.75, 1.0, 1.25, 1.5, 2.0}\n *   minimizing negative log-likelihood (NLL) over recorded outcomes. Default T=1.0.\n *   Requires ≥10 outcomes; below that threshold only prior shrinkage is applied.\n * - **Brier score**: `mean((predicted - actual)^2)` over recorded outcomes.\n * - **ECE (10-bin)**: bucket predictions in 0.1 increments, compute weighted\n *   absolute deviation between bin mean confidence and bin accuracy.\n *\n * @module core/evaluator/Calibrator\n */\n\nimport type {\n\tCalibrationMetrics,\n\tCalibrationResult,\n\tICalibrator,\n} from '../../contracts/calibrator.js';\nimport type { IOutcomeRecorder, VerificationOutcome } from '../../contracts/interfaces.js';\nimport type { ThoughtType } from '../../contracts/reasoning-types.js';\nimport type { SessionId } from '../../contracts/ids.js';\nimport {\n\tEPSILON,\n\tMIN_OUTCOMES_FOR_TEMPERATURE,\n\tapplyTemperature,\n\tfitTemperature,\n} from './calibration-math.js';\nimport { ALL_THOUGHT_TYPES } from './internals.js';\n\n/** Number of bins used by ECE (10-bin → 0.1 increments). */\nconst ECE_BINS = 10;\n\n/**\n * Build per-type empirical means + counts from a list of outcomes.\n *\n * @param outcomes - Outcomes to aggregate.\n * @returns Map of thought type to `{ mean, count }`. Types with no outcomes\n *          are absent from the map.\n */\nfunction aggregatePerType(\n\toutcomes: readonly VerificationOutcome[]\n): Map<string, { mean: number; count: number }> {\n\tconst sums = new Map<string, { sum: number; count: number }>();\n\tfor (const o of outcomes) {\n\t\tconst prev = sums.get(o.type) ?? { sum: 0, count: 0 };\n\t\tprev.sum += o.actual;\n\t\tprev.count += 1;\n\t\tsums.set(o.type, prev);\n\t}\n\tconst result = new Map<string, { mean: number; count: number }>();\n\tfor (const [type, { sum, count }] of sums) {\n\t\tresult.set(type, { mean: sum / count, count });\n\t}\n\treturn result;\n}\n\n/**\n * Compute the Brier score over a list of outcomes.\n *\n * @param outcomes - Outcomes to score.\n * @returns Mean squared error between predicted probability and actual label,\n *          or `null` if `outcomes` is empty.\n */\nfunction brierScore(outcomes: readonly VerificationOutcome[]): number | null {\n\tif (outcomes.length === 0) return null;\n\tlet sum = 0;\n\tfor (const o of outcomes) {\n\t\tconst diff = o.predicted - o.actual;\n\t\tsum += diff * diff;\n\t}\n\treturn sum / outcomes.length;\n}\n\n/**\n * Compute 10-bin Expected Calibration Error.\n *\n * Buckets predictions in 0.1-wide bins, then sums weighted absolute\n * differences between bin mean confidence and bin accuracy.\n *\n * @param outcomes - Outcomes to evaluate.\n * @returns ECE in `[0, 1]`, or `null` if `outcomes` is empty.\n */\nfunction expectedCalibrationError(outcomes: readonly VerificationOutcome[]): number | null {\n\tif (outcomes.length === 0) return null;\n\tconst binConfSum = new Array<number>(ECE_BINS).fill(0);\n\tconst binAccSum = new Array<number>(ECE_BINS).fill(0);\n\tconst binCount = new Array<number>(ECE_BINS).fill(0);\n\tfor (const o of outcomes) {\n\t\tconst p = Math.min(1 - EPSILON, Math.max(0, o.predicted));\n\t\tconst idx = Math.min(ECE_BINS - 1, Math.floor(p * ECE_BINS));\n\t\tbinConfSum[idx] = (binConfSum[idx] ?? 0) + p;\n\t\tbinAccSum[idx] = (binAccSum[idx] ?? 0) + o.actual;\n\t\tbinCount[idx] = (binCount[idx] ?? 0) + 1;\n\t}\n\tconst total = outcomes.length;\n\tlet ece = 0;\n\tfor (let i = 0; i < ECE_BINS; i++) {\n\t\tconst n = binCount[i] ?? 0;\n\t\tif (n === 0) continue;\n\t\tconst meanConf = (binConfSum[i] ?? 0) / n;\n\t\tconst meanAcc = (binAccSum[i] ?? 0) / n;\n\t\tece += (n / total) * Math.abs(meanConf - meanAcc);\n\t}\n\treturn ece;\n}\n\n/**\n * Build the per-type Brier score breakdown for {@link CalibrationMetrics}.\n *\n * @param outcomes - Outcomes to bucket by type.\n * @returns Record keyed by every {@link ThoughtType}; `null` for types with\n *          no recorded outcomes.\n */\nfunction perTypeBrier(\n\toutcomes: readonly VerificationOutcome[]\n): Record<ThoughtType, number | null> {\n\tconst buckets = new Map<string, VerificationOutcome[]>();\n\tfor (const o of outcomes) {\n\t\tconst list = buckets.get(o.type) ?? [];\n\t\tlist.push(o);\n\t\tbuckets.set(o.type, list);\n\t}\n\tconst result = {} as Record<ThoughtType, number | null>;\n\tfor (const t of ALL_THOUGHT_TYPES) {\n\t\tresult[t] = brierScore(buckets.get(t) ?? []);\n\t}\n\treturn result;\n}\n\n/** Empty CalibrationMetrics returned when calibration is disabled. */\nfunction emptyMetrics(): CalibrationMetrics {\n\tconst perType = {} as Record<ThoughtType, number | null>;\n\tfor (const t of ALL_THOUGHT_TYPES) perType[t] = null;\n\treturn {\n\t\tbrierScore: null,\n\t\tece: null,\n\t\tsampleCount: 0,\n\t\tperTypeBrier: perType,\n\t};\n}\n\n/**\n * Confidence calibration service.\n *\n * Stateless w.r.t. outcome storage (delegated to {@link IOutcomeRecorder}),\n * but maintains a small per-session temperature cache that is recomputed\n * via {@link Calibrator.refit}.\n *\n * @example\n * ```typescript\n * const calibrator = new Calibrator(outcomeRecorder, true);\n * const result = calibrator.calibrate(0.9, 'hypothesis', 'session-1');\n * console.log(result.calibrated, result.temperature, result.priorWeight);\n * ```\n */\nexport class Calibrator implements ICalibrator {\n\tpublic readonly enabled: boolean;\n\tprivate readonly _recorder: IOutcomeRecorder;\n\tprivate readonly _temperatures = new Map<SessionId, number>();\n\n\tconstructor(outcomeRecorder: IOutcomeRecorder, enabled: boolean) {\n\t\tthis._recorder = outcomeRecorder;\n\t\tthis.enabled = enabled;\n\t}\n\n\tpublic calibrate(\n\t\trawConfidence: number,\n\t\ttype: ThoughtType,\n\t\tsessionId: SessionId\n\t): CalibrationResult {\n\t\tconst raw = Math.min(1, Math.max(0, rawConfidence));\n\t\tif (!this.enabled) {\n\t\t\treturn { raw, calibrated: raw, temperature: 1.0, priorWeight: 0 };\n\t\t}\n\t\tconst outcomes = this._recorder.getOutcomes(sessionId);\n\t\tconst perType = aggregatePerType(outcomes);\n\t\tconst typeStats = perType.get(type);\n\t\tconst n = typeStats?.count ?? 0;\n\t\tconst observedMean = typeStats?.mean ?? 0.5;\n\t\tconst priorWeight = 1 / (1 + n / 10);\n\t\tconst shrunk = priorWeight * observedMean + (1 - priorWeight) * raw;\n\t\tconst temperature = this._temperatures.get(sessionId) ?? 1.0;\n\t\tconst calibrated =\n\t\t\toutcomes.length >= MIN_OUTCOMES_FOR_TEMPERATURE\n\t\t\t\t? applyTemperature(shrunk, temperature)\n\t\t\t\t: shrunk;\n\t\treturn { raw, calibrated, temperature, priorWeight };\n\t}\n\n\tpublic metrics(sessionId?: SessionId): CalibrationMetrics {\n\t\tif (!this.enabled) return emptyMetrics();\n\t\tconst outcomes =\n\t\t\tsessionId === undefined\n\t\t\t\t? this._recorder.getAllOutcomes()\n\t\t\t\t: this._recorder.getOutcomes(sessionId);\n\t\treturn {\n\t\t\tbrierScore: brierScore(outcomes),\n\t\t\tece: expectedCalibrationError(outcomes),\n\t\t\tsampleCount: outcomes.length,\n\t\t\tperTypeBrier: perTypeBrier(outcomes),\n\t\t};\n\t}\n\n\tpublic refit(sessionId: SessionId): void {\n\t\tif (!this.enabled) return;\n\t\tthis._temperatures.set(sessionId, fitTemperature(this._recorder.getOutcomes(sessionId)));\n\t}\n\n\tpublic clearSession(sessionId: SessionId): void {\n\t\tif (!this.enabled) return;\n\t\tthis._temperatures.delete(sessionId);\n\t}\n\n\tpublic clearAll(): void {\n\t\tif (!this.enabled) return;\n\t\tthis._temperatures.clear();\n\t}\n}\n"],"names":["ECE_BINS","aggregatePerType","outcomes","sums","Map","o","prev","result","type","sum","count","brierScore","diff","expectedCalibrationError","binConfSum","Array","binAccSum","binCount","p","Math","EPSILON","idx","total","ece","i","n","meanConf","meanAcc","perTypeBrier","buckets","list","t","ALL_THOUGHT_TYPES","emptyMetrics","perType","Calibrator","outcomeRecorder","enabled","rawConfidence","sessionId","raw","typeStats","observedMean","priorWeight","shrunk","temperature","calibrated","MIN_OUTCOMES_FOR_TEMPERATURE","applyTemperature","undefined","fitTemperature"],"mappings":";;AAwCA,MAAMA,WAAW;AASjB,SAASC,iBACRC,QAAwC;IAExC,MAAMC,OAAO,IAAIC;IACjB,KAAK,MAAMC,KAAKH,SAAU;QACzB,MAAMI,OAAOH,KAAK,GAAG,CAACE,EAAE,IAAI,KAAK;YAAE,KAAK;YAAG,OAAO;QAAE;QACpDC,KAAK,GAAG,IAAID,EAAE,MAAM;QACpBC,KAAK,KAAK,IAAI;QACdH,KAAK,GAAG,CAACE,EAAE,IAAI,EAAEC;IAClB;IACA,MAAMC,SAAS,IAAIH;IACnB,KAAK,MAAM,CAACI,MAAM,EAAEC,GAAG,EAAEC,KAAK,EAAE,CAAC,IAAIP,KACpCI,OAAO,GAAG,CAACC,MAAM;QAAE,MAAMC,MAAMC;QAAOA;IAAM;IAE7C,OAAOH;AACR;AASA,SAASI,WAAWT,QAAwC;IAC3D,IAAIA,AAAoB,MAApBA,SAAS,MAAM,EAAQ,OAAO;IAClC,IAAIO,MAAM;IACV,KAAK,MAAMJ,KAAKH,SAAU;QACzB,MAAMU,OAAOP,EAAE,SAAS,GAAGA,EAAE,MAAM;QACnCI,OAAOG,OAAOA;IACf;IACA,OAAOH,MAAMP,SAAS,MAAM;AAC7B;AAWA,SAASW,yBAAyBX,QAAwC;IACzE,IAAIA,AAAoB,MAApBA,SAAS,MAAM,EAAQ,OAAO;IAClC,MAAMY,aAAa,IAAIC,MAAcf,UAAU,IAAI,CAAC;IACpD,MAAMgB,YAAY,IAAID,MAAcf,UAAU,IAAI,CAAC;IACnD,MAAMiB,WAAW,IAAIF,MAAcf,UAAU,IAAI,CAAC;IAClD,KAAK,MAAMK,KAAKH,SAAU;QACzB,MAAMgB,IAAIC,KAAK,GAAG,CAAC,IAAIC,SAASD,KAAK,GAAG,CAAC,GAAGd,EAAE,SAAS;QACvD,MAAMgB,MAAMF,KAAK,GAAG,CAACnB,WAAW,GAAGmB,KAAK,KAAK,CAACD,IAAIlB;QAClDc,UAAU,CAACO,IAAI,GAAIP,AAAAA,CAAAA,UAAU,CAACO,IAAI,IAAI,KAAKH;QAC3CF,SAAS,CAACK,IAAI,GAAIL,AAAAA,CAAAA,SAAS,CAACK,IAAI,IAAI,KAAKhB,EAAE,MAAM;QACjDY,QAAQ,CAACI,IAAI,GAAIJ,AAAAA,CAAAA,QAAQ,CAACI,IAAI,IAAI,KAAK;IACxC;IACA,MAAMC,QAAQpB,SAAS,MAAM;IAC7B,IAAIqB,MAAM;IACV,IAAK,IAAIC,IAAI,GAAGA,IAAIxB,UAAUwB,IAAK;QAClC,MAAMC,IAAIR,QAAQ,CAACO,EAAE,IAAI;QACzB,IAAIC,AAAM,MAANA,GAAS;QACb,MAAMC,WAAYZ,AAAAA,CAAAA,UAAU,CAACU,EAAE,IAAI,KAAKC;QACxC,MAAME,UAAWX,AAAAA,CAAAA,SAAS,CAACQ,EAAE,IAAI,KAAKC;QACtCF,OAAQE,IAAIH,QAASH,KAAK,GAAG,CAACO,WAAWC;IAC1C;IACA,OAAOJ;AACR;AASA,SAASK,aACR1B,QAAwC;IAExC,MAAM2B,UAAU,IAAIzB;IACpB,KAAK,MAAMC,KAAKH,SAAU;QACzB,MAAM4B,OAAOD,QAAQ,GAAG,CAACxB,EAAE,IAAI,KAAK,EAAE;QACtCyB,KAAK,IAAI,CAACzB;QACVwB,QAAQ,GAAG,CAACxB,EAAE,IAAI,EAAEyB;IACrB;IACA,MAAMvB,SAAS,CAAC;IAChB,KAAK,MAAMwB,KAAKC,kBACfzB,MAAM,CAACwB,EAAE,GAAGpB,WAAWkB,QAAQ,GAAG,CAACE,MAAM,EAAE;IAE5C,OAAOxB;AACR;AAGA,SAAS0B;IACR,MAAMC,UAAU,CAAC;IACjB,KAAK,MAAMH,KAAKC,kBAAmBE,OAAO,CAACH,EAAE,GAAG;IAChD,OAAO;QACN,YAAY;QACZ,KAAK;QACL,aAAa;QACb,cAAcG;IACf;AACD;AAgBO,MAAMC;IACI,QAAiB;IAChB,UAA4B;IAC5B,gBAAgB,IAAI/B,MAAyB;IAE9D,YAAYgC,eAAiC,EAAEC,OAAgB,CAAE;QAChE,IAAI,CAAC,SAAS,GAAGD;QACjB,IAAI,CAAC,OAAO,GAAGC;IAChB;IAEO,UACNC,aAAqB,EACrB9B,IAAiB,EACjB+B,SAAoB,EACA;QACpB,MAAMC,MAAMrB,KAAK,GAAG,CAAC,GAAGA,KAAK,GAAG,CAAC,GAAGmB;QACpC,IAAI,CAAC,IAAI,CAAC,OAAO,EAChB,OAAO;YAAEE;YAAK,YAAYA;YAAK,aAAa;YAAK,aAAa;QAAE;QAEjE,MAAMtC,WAAW,IAAI,CAAC,SAAS,CAAC,WAAW,CAACqC;QAC5C,MAAML,UAAUjC,iBAAiBC;QACjC,MAAMuC,YAAYP,QAAQ,GAAG,CAAC1B;QAC9B,MAAMiB,IAAIgB,WAAW,SAAS;QAC9B,MAAMC,eAAeD,WAAW,QAAQ;QACxC,MAAME,cAAc,IAAK,KAAIlB,IAAI,EAAC;QAClC,MAAMmB,SAASD,cAAcD,eAAgB,KAAIC,WAAU,IAAKH;QAChE,MAAMK,cAAc,IAAI,CAAC,aAAa,CAAC,GAAG,CAACN,cAAc;QACzD,MAAMO,aACL5C,SAAS,MAAM,IAAI6C,+BAChBC,iBAAiBJ,QAAQC,eACzBD;QACJ,OAAO;YAAEJ;YAAKM;YAAYD;YAAaF;QAAY;IACpD;IAEO,QAAQJ,SAAqB,EAAsB;QACzD,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE,OAAON;QAC1B,MAAM/B,WACLqC,AAAcU,WAAdV,YACG,IAAI,CAAC,SAAS,CAAC,cAAc,KAC7B,IAAI,CAAC,SAAS,CAAC,WAAW,CAACA;QAC/B,OAAO;YACN,YAAY5B,WAAWT;YACvB,KAAKW,yBAAyBX;YAC9B,aAAaA,SAAS,MAAM;YAC5B,cAAc0B,aAAa1B;QAC5B;IACD;IAEO,MAAMqC,SAAoB,EAAQ;QACxC,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE;QACnB,IAAI,CAAC,aAAa,CAAC,GAAG,CAACA,WAAWW,eAAe,IAAI,CAAC,SAAS,CAAC,WAAW,CAACX;IAC7E;IAEO,aAAaA,SAAoB,EAAQ;QAC/C,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE;QACnB,IAAI,CAAC,aAAa,CAAC,MAAM,CAACA;IAC3B;IAEO,WAAiB;QACvB,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE;QACnB,IAAI,CAAC,aAAa,CAAC,KAAK;IACzB;AACD"}
+{"version":3,"file":"core/evaluator/Calibrator.js","sources":["../../../src/core/evaluator/Calibrator.ts"],"sourcesContent":["/**\n * Confidence calibration for sequential thinking reasoning.\n *\n * Provides the {@link Calibrator} class — maps raw model confidence values\n * to calibrated probabilities using per-type empirical evidence blending and\n * grid-searched temperature scaling. Reports calibration quality through\n * Brier score and 10-bin Expected Calibration Error (ECE).\n *\n * Math summary:\n * - **Per-type evidence blend**: Raw confidence is the prior signal, weighted by\n *   `priorWeight = 10 / (10 + n)` where `n` is the type's outcome count. The\n *   remaining weight is applied to the observed per-type mean. An empty type\n *   defaults its mean to 0.5 but has zero empirical weight, so cold start returns raw.\n * - **Temperature scaling**: Grid search T ∈ {0.5, 0.75, 1.0, 1.25, 1.5, 2.0}\n *   minimizing negative log-likelihood (NLL) over per-type leave-one-out evidence blends.\n *   Exact ties retain T=1.0, then follow grid order. Requires at least 10 outcomes; below that\n *   threshold the evidence blend is returned unchanged.\n * - **Brier score**: `mean((predicted - actual)^2)` over stored raw outcomes.\n * - **ECE (10-bin)**: bucket predictions in 0.1 increments, compute weighted\n *   absolute deviation between raw bin mean confidence and bin accuracy.\n *\n * Stored outcomes and metrics remain raw. This transform does not guarantee factuality, better\n * metrics for every dataset, or monotonic per-example changes.\n *\n * @module core/evaluator/Calibrator\n */\n\nimport type {\n\tCalibrationMetrics,\n\tCalibrationResult,\n\tICalibrator,\n} from '../../contracts/calibrator.js';\nimport type { IOutcomeRecorder, VerificationOutcome } from '../../contracts/interfaces.js';\nimport type { ThoughtType } from '../../contracts/reasoning-types.js';\nimport type { SessionId } from '../../contracts/ids.js';\nimport {\n\tEPSILON,\n\tMIN_OUTCOMES_FOR_TEMPERATURE,\n\tapplyTemperature,\n\tblendWithRawPrior,\n\tfitTemperature,\n} from './calibration-math.js';\nimport { ALL_THOUGHT_TYPES } from './internals.js';\n\n/** Number of bins used by ECE (10-bin → 0.1 increments). */\nconst ECE_BINS = 10;\n\n/**\n * Build per-type empirical sums + counts from a list of outcomes.\n *\n * @param outcomes - Outcomes to aggregate.\n * @returns Map of thought type to `{ sum, count }`. Types with no outcomes\n *          are absent from the map.\n */\nfunction aggregatePerType(\n\toutcomes: readonly VerificationOutcome[]\n): Map<ThoughtType, { sum: number; count: number }> {\n\tconst sums = new Map<ThoughtType, { sum: number; count: number }>();\n\tfor (const o of outcomes) {\n\t\tconst prev = sums.get(o.type) ?? { sum: 0, count: 0 };\n\t\tprev.sum += o.actual;\n\t\tprev.count += 1;\n\t\tsums.set(o.type, prev);\n\t}\n\treturn sums;\n}\n\nfunction leaveOneOutBlendedOutcomes(\n\toutcomes: readonly VerificationOutcome[]\n): VerificationOutcome[] {\n\tconst perType = aggregatePerType(outcomes);\n\treturn outcomes.map((outcome) => {\n\t\tconst typeStats = perType.get(outcome.type) ?? { sum: outcome.actual, count: 1 };\n\t\tconst sampleCount = typeStats.count - 1;\n\t\tif (sampleCount === 0) return { ...outcome };\n\t\tconst empiricalMean = (typeStats.sum - outcome.actual) / sampleCount;\n\t\tconst { blended } = blendWithRawPrior(outcome.predicted, empiricalMean, sampleCount);\n\t\treturn { ...outcome, predicted: blended };\n\t});\n}\n\n/**\n * Compute the Brier score over a list of outcomes.\n *\n * @param outcomes - Outcomes to score.\n * @returns Mean squared error between predicted probability and actual label,\n *          or `null` if `outcomes` is empty.\n */\nfunction brierScore(outcomes: readonly VerificationOutcome[]): number | null {\n\tif (outcomes.length === 0) return null;\n\tlet sum = 0;\n\tfor (const o of outcomes) {\n\t\tconst diff = o.predicted - o.actual;\n\t\tsum += diff * diff;\n\t}\n\treturn sum / outcomes.length;\n}\n\n/**\n * Compute 10-bin Expected Calibration Error.\n *\n * Buckets predictions in 0.1-wide bins, then sums weighted absolute\n * differences between bin mean confidence and bin accuracy.\n *\n * @param outcomes - Outcomes to evaluate.\n * @returns ECE in `[0, 1]`, or `null` if `outcomes` is empty.\n */\nfunction expectedCalibrationError(outcomes: readonly VerificationOutcome[]): number | null {\n\tif (outcomes.length === 0) return null;\n\tconst binConfSum = new Array<number>(ECE_BINS).fill(0);\n\tconst binAccSum = new Array<number>(ECE_BINS).fill(0);\n\tconst binCount = new Array<number>(ECE_BINS).fill(0);\n\tfor (const o of outcomes) {\n\t\tconst p = Math.min(1 - EPSILON, Math.max(0, o.predicted));\n\t\tconst idx = Math.min(ECE_BINS - 1, Math.floor(p * ECE_BINS));\n\t\tbinConfSum[idx] = (binConfSum[idx] ?? 0) + p;\n\t\tbinAccSum[idx] = (binAccSum[idx] ?? 0) + o.actual;\n\t\tbinCount[idx] = (binCount[idx] ?? 0) + 1;\n\t}\n\tconst total = outcomes.length;\n\tlet ece = 0;\n\tfor (let i = 0; i < ECE_BINS; i++) {\n\t\tconst n = binCount[i] ?? 0;\n\t\tif (n === 0) continue;\n\t\tconst meanConf = (binConfSum[i] ?? 0) / n;\n\t\tconst meanAcc = (binAccSum[i] ?? 0) / n;\n\t\tece += (n / total) * Math.abs(meanConf - meanAcc);\n\t}\n\treturn ece;\n}\n\n/**\n * Build the per-type Brier score breakdown for {@link CalibrationMetrics}.\n *\n * @param outcomes - Outcomes to bucket by type.\n * @returns Record keyed by every {@link ThoughtType}; `null` for types with\n *          no recorded outcomes.\n */\nfunction perTypeBrier(\n\toutcomes: readonly VerificationOutcome[]\n): Record<ThoughtType, number | null> {\n\tconst buckets = new Map<string, VerificationOutcome[]>();\n\tfor (const o of outcomes) {\n\t\tconst list = buckets.get(o.type) ?? [];\n\t\tlist.push(o);\n\t\tbuckets.set(o.type, list);\n\t}\n\tconst result = {} as Record<ThoughtType, number | null>;\n\tfor (const t of ALL_THOUGHT_TYPES) {\n\t\tresult[t] = brierScore(buckets.get(t) ?? []);\n\t}\n\treturn result;\n}\n\n/** Empty CalibrationMetrics returned when calibration is disabled. */\nfunction emptyMetrics(): CalibrationMetrics {\n\tconst perType = {} as Record<ThoughtType, number | null>;\n\tfor (const t of ALL_THOUGHT_TYPES) perType[t] = null;\n\treturn {\n\t\tbrierScore: null,\n\t\tece: null,\n\t\tsampleCount: 0,\n\t\tperTypeBrier: perType,\n\t};\n}\n\n/**\n * Confidence calibration service.\n *\n * Stateless w.r.t. outcome storage (delegated to {@link IOutcomeRecorder}),\n * but maintains a small per-session temperature cache that is recomputed\n * via {@link Calibrator.refit}.\n *\n * @example\n * ```typescript\n * const calibrator = new Calibrator(outcomeRecorder, true);\n * const result = calibrator.calibrate(0.9, 'hypothesis', 'session-1');\n * console.log(result.calibrated, result.temperature, result.priorWeight);\n * ```\n */\nexport class Calibrator implements ICalibrator {\n\tpublic readonly enabled: boolean;\n\tprivate readonly _recorder: IOutcomeRecorder;\n\tprivate readonly _temperatures = new Map<SessionId, number>();\n\n\tconstructor(outcomeRecorder: IOutcomeRecorder, enabled: boolean) {\n\t\tthis._recorder = outcomeRecorder;\n\t\tthis.enabled = enabled;\n\t}\n\n\tpublic calibrate(\n\t\trawConfidence: number,\n\t\ttype: ThoughtType,\n\t\tsessionId: SessionId\n\t): CalibrationResult {\n\t\tconst raw = Math.min(1, Math.max(0, rawConfidence));\n\t\tif (!this.enabled) {\n\t\t\treturn { raw, calibrated: raw, temperature: 1.0, priorWeight: 0 };\n\t\t}\n\t\tconst outcomes = this._recorder.getOutcomes(sessionId);\n\t\tconst perType = aggregatePerType(outcomes);\n\t\tconst typeStats = perType.get(type);\n\t\tconst n = typeStats?.count ?? 0;\n\t\tconst observedMean = typeStats === undefined ? 0.5 : typeStats.sum / typeStats.count;\n\t\tconst { blended, priorWeight } = blendWithRawPrior(raw, observedMean, n);\n\t\tconst temperature = this._temperatures.get(sessionId) ?? 1.0;\n\t\tconst calibrated =\n\t\t\toutcomes.length >= MIN_OUTCOMES_FOR_TEMPERATURE\n\t\t\t\t? applyTemperature(blended, temperature)\n\t\t\t\t: blended;\n\t\treturn { raw, calibrated, temperature, priorWeight };\n\t}\n\n\tpublic metrics(sessionId?: SessionId): CalibrationMetrics {\n\t\tif (!this.enabled) return emptyMetrics();\n\t\tconst outcomes =\n\t\t\tsessionId === undefined\n\t\t\t\t? this._recorder.getAllOutcomes()\n\t\t\t\t: this._recorder.getOutcomes(sessionId);\n\t\treturn {\n\t\t\tbrierScore: brierScore(outcomes),\n\t\t\tece: expectedCalibrationError(outcomes),\n\t\t\tsampleCount: outcomes.length,\n\t\t\tperTypeBrier: perTypeBrier(outcomes),\n\t\t};\n\t}\n\n\tpublic refit(sessionId: SessionId): void {\n\t\tif (!this.enabled) return;\n\t\tconst outcomes = this._recorder.getOutcomes(sessionId);\n\t\tthis._temperatures.set(sessionId, fitTemperature(leaveOneOutBlendedOutcomes(outcomes)));\n\t}\n\n\tpublic clearSession(sessionId: SessionId): void {\n\t\tif (!this.enabled) return;\n\t\tthis._temperatures.delete(sessionId);\n\t}\n\n\tpublic clearAll(): void {\n\t\tif (!this.enabled) return;\n\t\tthis._temperatures.clear();\n\t}\n}\n"],"names":["ECE_BINS","aggregatePerType","outcomes","sums","Map","o","prev","leaveOneOutBlendedOutcomes","perType","outcome","typeStats","sampleCount","empiricalMean","blended","blendWithRawPrior","brierScore","sum","diff","expectedCalibrationError","binConfSum","Array","binAccSum","binCount","p","Math","EPSILON","idx","total","ece","i","n","meanConf","meanAcc","perTypeBrier","buckets","list","result","t","ALL_THOUGHT_TYPES","emptyMetrics","Calibrator","outcomeRecorder","enabled","rawConfidence","type","sessionId","raw","observedMean","undefined","priorWeight","temperature","calibrated","MIN_OUTCOMES_FOR_TEMPERATURE","applyTemperature","fitTemperature"],"mappings":";;AA6CA,MAAMA,WAAW;AASjB,SAASC,iBACRC,QAAwC;IAExC,MAAMC,OAAO,IAAIC;IACjB,KAAK,MAAMC,KAAKH,SAAU;QACzB,MAAMI,OAAOH,KAAK,GAAG,CAACE,EAAE,IAAI,KAAK;YAAE,KAAK;YAAG,OAAO;QAAE;QACpDC,KAAK,GAAG,IAAID,EAAE,MAAM;QACpBC,KAAK,KAAK,IAAI;QACdH,KAAK,GAAG,CAACE,EAAE,IAAI,EAAEC;IAClB;IACA,OAAOH;AACR;AAEA,SAASI,2BACRL,QAAwC;IAExC,MAAMM,UAAUP,iBAAiBC;IACjC,OAAOA,SAAS,GAAG,CAAC,CAACO;QACpB,MAAMC,YAAYF,QAAQ,GAAG,CAACC,QAAQ,IAAI,KAAK;YAAE,KAAKA,QAAQ,MAAM;YAAE,OAAO;QAAE;QAC/E,MAAME,cAAcD,UAAU,KAAK,GAAG;QACtC,IAAIC,AAAgB,MAAhBA,aAAmB,OAAO;YAAE,GAAGF,OAAO;QAAC;QAC3C,MAAMG,gBAAiBF,AAAAA,CAAAA,UAAU,GAAG,GAAGD,QAAQ,MAAK,IAAKE;QACzD,MAAM,EAAEE,OAAO,EAAE,GAAGC,kBAAkBL,QAAQ,SAAS,EAAEG,eAAeD;QACxE,OAAO;YAAE,GAAGF,OAAO;YAAE,WAAWI;QAAQ;IACzC;AACD;AASA,SAASE,WAAWb,QAAwC;IAC3D,IAAIA,AAAoB,MAApBA,SAAS,MAAM,EAAQ,OAAO;IAClC,IAAIc,MAAM;IACV,KAAK,MAAMX,KAAKH,SAAU;QACzB,MAAMe,OAAOZ,EAAE,SAAS,GAAGA,EAAE,MAAM;QACnCW,OAAOC,OAAOA;IACf;IACA,OAAOD,MAAMd,SAAS,MAAM;AAC7B;AAWA,SAASgB,yBAAyBhB,QAAwC;IACzE,IAAIA,AAAoB,MAApBA,SAAS,MAAM,EAAQ,OAAO;IAClC,MAAMiB,aAAa,IAAIC,MAAcpB,UAAU,IAAI,CAAC;IACpD,MAAMqB,YAAY,IAAID,MAAcpB,UAAU,IAAI,CAAC;IACnD,MAAMsB,WAAW,IAAIF,MAAcpB,UAAU,IAAI,CAAC;IAClD,KAAK,MAAMK,KAAKH,SAAU;QACzB,MAAMqB,IAAIC,KAAK,GAAG,CAAC,IAAIC,SAASD,KAAK,GAAG,CAAC,GAAGnB,EAAE,SAAS;QACvD,MAAMqB,MAAMF,KAAK,GAAG,CAACxB,WAAW,GAAGwB,KAAK,KAAK,CAACD,IAAIvB;QAClDmB,UAAU,CAACO,IAAI,GAAIP,AAAAA,CAAAA,UAAU,CAACO,IAAI,IAAI,KAAKH;QAC3CF,SAAS,CAACK,IAAI,GAAIL,AAAAA,CAAAA,SAAS,CAACK,IAAI,IAAI,KAAKrB,EAAE,MAAM;QACjDiB,QAAQ,CAACI,IAAI,GAAIJ,AAAAA,CAAAA,QAAQ,CAACI,IAAI,IAAI,KAAK;IACxC;IACA,MAAMC,QAAQzB,SAAS,MAAM;IAC7B,IAAI0B,MAAM;IACV,IAAK,IAAIC,IAAI,GAAGA,IAAI7B,UAAU6B,IAAK;QAClC,MAAMC,IAAIR,QAAQ,CAACO,EAAE,IAAI;QACzB,IAAIC,AAAM,MAANA,GAAS;QACb,MAAMC,WAAYZ,AAAAA,CAAAA,UAAU,CAACU,EAAE,IAAI,KAAKC;QACxC,MAAME,UAAWX,AAAAA,CAAAA,SAAS,CAACQ,EAAE,IAAI,KAAKC;QACtCF,OAAQE,IAAIH,QAASH,KAAK,GAAG,CAACO,WAAWC;IAC1C;IACA,OAAOJ;AACR;AASA,SAASK,aACR/B,QAAwC;IAExC,MAAMgC,UAAU,IAAI9B;IACpB,KAAK,MAAMC,KAAKH,SAAU;QACzB,MAAMiC,OAAOD,QAAQ,GAAG,CAAC7B,EAAE,IAAI,KAAK,EAAE;QACtC8B,KAAK,IAAI,CAAC9B;QACV6B,QAAQ,GAAG,CAAC7B,EAAE,IAAI,EAAE8B;IACrB;IACA,MAAMC,SAAS,CAAC;IAChB,KAAK,MAAMC,KAAKC,kBACfF,MAAM,CAACC,EAAE,GAAGtB,WAAWmB,QAAQ,GAAG,CAACG,MAAM,EAAE;IAE5C,OAAOD;AACR;AAGA,SAASG;IACR,MAAM/B,UAAU,CAAC;IACjB,KAAK,MAAM6B,KAAKC,kBAAmB9B,OAAO,CAAC6B,EAAE,GAAG;IAChD,OAAO;QACN,YAAY;QACZ,KAAK;QACL,aAAa;QACb,cAAc7B;IACf;AACD;AAgBO,MAAMgC;IACI,QAAiB;IAChB,UAA4B;IAC5B,gBAAgB,IAAIpC,MAAyB;IAE9D,YAAYqC,eAAiC,EAAEC,OAAgB,CAAE;QAChE,IAAI,CAAC,SAAS,GAAGD;QACjB,IAAI,CAAC,OAAO,GAAGC;IAChB;IAEO,UACNC,aAAqB,EACrBC,IAAiB,EACjBC,SAAoB,EACA;QACpB,MAAMC,MAAMtB,KAAK,GAAG,CAAC,GAAGA,KAAK,GAAG,CAAC,GAAGmB;QACpC,IAAI,CAAC,IAAI,CAAC,OAAO,EAChB,OAAO;YAAEG;YAAK,YAAYA;YAAK,aAAa;YAAK,aAAa;QAAE;QAEjE,MAAM5C,WAAW,IAAI,CAAC,SAAS,CAAC,WAAW,CAAC2C;QAC5C,MAAMrC,UAAUP,iBAAiBC;QACjC,MAAMQ,YAAYF,QAAQ,GAAG,CAACoC;QAC9B,MAAMd,IAAIpB,WAAW,SAAS;QAC9B,MAAMqC,eAAerC,AAAcsC,WAAdtC,YAA0B,MAAMA,UAAU,GAAG,GAAGA,UAAU,KAAK;QACpF,MAAM,EAAEG,OAAO,EAAEoC,WAAW,EAAE,GAAGnC,kBAAkBgC,KAAKC,cAAcjB;QACtE,MAAMoB,cAAc,IAAI,CAAC,aAAa,CAAC,GAAG,CAACL,cAAc;QACzD,MAAMM,aACLjD,SAAS,MAAM,IAAIkD,+BAChBC,iBAAiBxC,SAASqC,eAC1BrC;QACJ,OAAO;YAAEiC;YAAKK;YAAYD;YAAaD;QAAY;IACpD;IAEO,QAAQJ,SAAqB,EAAsB;QACzD,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE,OAAON;QAC1B,MAAMrC,WACL2C,AAAcG,WAAdH,YACG,IAAI,CAAC,SAAS,CAAC,cAAc,KAC7B,IAAI,CAAC,SAAS,CAAC,WAAW,CAACA;QAC/B,OAAO;YACN,YAAY9B,WAAWb;YACvB,KAAKgB,yBAAyBhB;YAC9B,aAAaA,SAAS,MAAM;YAC5B,cAAc+B,aAAa/B;QAC5B;IACD;IAEO,MAAM2C,SAAoB,EAAQ;QACxC,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE;QACnB,MAAM3C,WAAW,IAAI,CAAC,SAAS,CAAC,WAAW,CAAC2C;QAC5C,IAAI,CAAC,aAAa,CAAC,GAAG,CAACA,WAAWS,eAAe/C,2BAA2BL;IAC7E;IAEO,aAAa2C,SAAoB,EAAQ;QAC/C,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE;QACnB,IAAI,CAAC,aAAa,CAAC,MAAM,CAACA;IAC3B;IAEO,WAAiB;QACvB,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE;QACnB,IAAI,CAAC,aAAa,CAAC,KAAK;IACzB;AACD"}

package/dist/core/evaluator/PatternDetector.d.ts

@@ -22,6 +22,7 @@
  */
 import type { PatternSignal } from '../reasoning.js';
 import type { ThoughtData } from '../thought.js';
+import type { VerificationLinks } from './VerificationLinks.js';
 /**
  * Stateless service that detects reasoning patterns from thought history.
  *
@@ -45,7 +46,7 @@ export declare class PatternDetector {
      * @param branches - Map of branch IDs to their thought arrays
      * @returns Array of detected pattern signals (possibly empty)
      */
-    computePatternSignals(history: ThoughtData[], branches: Record<string, ThoughtData[]>): PatternSignal[];
+    computePatternSignals(history: ThoughtData[], branches: Record<string, ThoughtData[]>, links: VerificationLinks): PatternSignal[];
     /** Detect runs of 3+ consecutive thoughts without verification. */
     private _detectConsecutiveWithoutVerification;
     /** Detect hypothesis thoughts not verified within 3 subsequent thoughts. */

package/dist/core/evaluator/PatternDetector.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"PatternDetector.d.ts","sourceRoot":"","sources":["../../../src/core/evaluator/PatternDetector.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AAEH,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,iBAAiB,CAAC;AACrD,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,eAAe,CAAC;AAEjD;;;;;;;;;;;;;;GAcG;AACH,qBAAa,eAAe;IAC3B;;;;;;OAMG;IACI,qBAAqB,CAC3B,OAAO,EAAE,WAAW,EAAE,EACtB,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,WAAW,EAAE,CAAC,GACrC,aAAa,EAAE;IAalB,mEAAmE;IACnE,OAAO,CAAC,qCAAqC;IAwB7C,4EAA4E;IAC5E,OAAO,CAAC,2BAA2B;IAqBnC,2DAA2D;IAC3D,OAAO,CAAC,6BAA6B;IAmBrC;;;OAGG;IACH,OAAO,CAAC,oBAAoB;IAyB5B,gEAAgE;IAChE,OAAO,CAAC,kBAAkB;IAkB1B,kFAAkF;IAClF,OAAO,CAAC,sBAAsB;IAiC9B,kEAAkE;IAClE,OAAO,CAAC,cAAc;IAuBtB,iFAAiF;IACjF,OAAO,CAAC,0BAA0B;CAyBlC"}
+{"version":3,"file":"PatternDetector.d.ts","sourceRoot":"","sources":["../../../src/core/evaluator/PatternDetector.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AAEH,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,iBAAiB,CAAC;AACrD,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,eAAe,CAAC;AACjD,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,wBAAwB,CAAC;AAEhE;;;;;;;;;;;;;;GAcG;AACH,qBAAa,eAAe;IAC3B;;;;;;OAMG;IACI,qBAAqB,CAC3B,OAAO,EAAE,WAAW,EAAE,EACtB,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,WAAW,EAAE,CAAC,EACvC,KAAK,EAAE,iBAAiB,GACtB,aAAa,EAAE;IAalB,mEAAmE;IACnE,OAAO,CAAC,qCAAqC;IAwB7C,4EAA4E;IAC5E,OAAO,CAAC,2BAA2B;IA+BnC,2DAA2D;IAC3D,OAAO,CAAC,6BAA6B;IAmBrC;;;OAGG;IACH,OAAO,CAAC,oBAAoB;IAyB5B,gEAAgE;IAChE,OAAO,CAAC,kBAAkB;IAkB1B,kFAAkF;IAClF,OAAO,CAAC,sBAAsB;IAiC9B,kEAAkE;IAClE,OAAO,CAAC,cAAc;IAuBtB,iFAAiF;IACjF,OAAO,CAAC,0BAA0B;CA6BlC"}

package/dist/core/evaluator/PatternDetector.js

@@ -1,13 +1,13 @@
 class PatternDetector {
-    computePatternSignals(history, branches) {
+    computePatternSignals(history, branches, links) {
         if (0 === history.length) return [];
         const signals = [];
         signals.push(...this._detectConsecutiveWithoutVerification(history));
-        signals.push(...this._detectUnverifiedHypothesis(history));
+        signals.push(...this._detectUnverifiedHypothesis(history, links));
         signals.push(...this._detectNoAlternativesExplored(history, branches));
         signals.push(...this._detectMonotonicType(history));
         signals.push(...this._detectConfidenceDrift(history));
-        signals.push(...this._detectHealthyVerification(history));
+        signals.push(...this._detectHealthyVerification(history, links));
         return signals;
     }
     _detectConsecutiveWithoutVerification(history) {
@@ -36,16 +36,17 @@ class PatternDetector {
         }
         return signals;
     }
-    _detectUnverifiedHypothesis(history) {
+    _detectUnverifiedHypothesis(history, links) {
         const signals = [];
         for(let i = 0; i < history.length; i++){
-            if ('hypothesis' !== history[i].thought_type) continue;
+            const hypothesis = history[i];
+            if ('hypothesis' !== hypothesis.thought_type || !hypothesis.id || !links.hypothesisIds.has(hypothesis.id)) continue;
             const remaining = history.length - i - 1;
             if (remaining < 3) continue;
             const lookahead = history.slice(i + 1, i + 4);
-            const hasVerification = lookahead.some((t)=>'verification' === t.thought_type);
+            const hasVerification = lookahead.some((verifier)=>links.verifies(verifier, hypothesis));
             if (!hasVerification) {
-                const n = history[i].thought_number ?? i + 1;
+                const n = hypothesis.thought_number ?? i + 1;
                 signals.push({
                     pattern: 'unverified_hypothesis',
                     severity: 'warning',
@@ -163,15 +164,15 @@ class PatternDetector {
             ]
         });
     }
-    _detectHealthyVerification(history) {
+    _detectHealthyVerification(history, links) {
         const signals = [];
         for(let i = 0; i < history.length; i++){
-            if ('hypothesis' !== history[i].thought_type) continue;
-            const hypId = history[i].hypothesis_id;
+            const hypothesis = history[i];
+            if ('hypothesis' !== hypothesis.thought_type || !hypothesis.id || !links.hypothesisIds.has(hypothesis.id)) continue;
             const lookahead = history.slice(i + 1, i + 4);
-            const verifier = lookahead.find((t)=>'verification' === t.thought_type && (t.hypothesis_id === hypId || t.verification_target === (history[i].thought_number ?? i + 1)));
+            const verifier = lookahead.find((candidate)=>links.verifies(candidate, hypothesis));
             if (verifier) {
-                const n = history[i].thought_number ?? i + 1;
+                const n = hypothesis.thought_number ?? i + 1;
                 const m = verifier.thought_number ?? history.indexOf(verifier) + 1;
                 signals.push({
                     pattern: 'healthy_verification',

package/dist/core/evaluator/PatternDetector.js.map

@@ -1 +1 @@
-{"version":3,"file":"core/evaluator/PatternDetector.js","sources":["../../../src/core/evaluator/PatternDetector.ts"],"sourcesContent":["/**\n * Pattern detection for sequential thinking reasoning analytics.\n *\n * Provides the {@link PatternDetector} class — a stateless service that detects\n * reasoning patterns (anti-patterns and positive signals) from thought history.\n * Extracted from {@link ThoughtEvaluator} to keep concerns isolated and reduce\n * cyclomatic complexity per function.\n *\n * Detected patterns:\n * - `consecutive_without_verification` (warning) — 3+ consecutive regular thoughts\n *   without a verification step\n * - `unverified_hypothesis` (warning) — hypothesis not verified within 3 subsequent\n *   thoughts\n * - `no_alternatives_explored` (info) — 5+ thoughts with no critique and no branches\n * - `monotonic_type` (info) — 4+ consecutive thoughts with the same thought_type\n *   (requires ≥1 explicit type and ≥5 thoughts)\n * - `confidence_drift` (warning) — 3+ consecutive thoughts with strictly decreasing\n *   confidence\n * - `healthy_verification` (info) — hypothesis verified within 3 subsequent thoughts\n *\n * @module core/evaluator/PatternDetector\n */\n\nimport type { PatternSignal } from '../reasoning.js';\nimport type { ThoughtData } from '../thought.js';\n\n/**\n * Stateless service that detects reasoning patterns from thought history.\n *\n * @remarks\n * All methods are pure computations — no side effects, no I/O, no internal state.\n * Designed to be registered as transient in the DI container or composed directly\n * into {@link ThoughtEvaluator}.\n *\n * @example\n * ```typescript\n * const detector = new PatternDetector();\n * const signals = detector.computePatternSignals(history, branches);\n * const warnings = signals.filter((s) => s.severity === 'warning');\n * ```\n */\nexport class PatternDetector {\n\t/**\n\t * Detect all reasoning patterns from history and branches.\n\t *\n\t * @param history - All thoughts in the current session\n\t * @param branches - Map of branch IDs to their thought arrays\n\t * @returns Array of detected pattern signals (possibly empty)\n\t */\n\tpublic computePatternSignals(\n\t\thistory: ThoughtData[],\n\t\tbranches: Record<string, ThoughtData[]>\n\t): PatternSignal[] {\n\t\tif (history.length === 0) return [];\n\n\t\tconst signals: PatternSignal[] = [];\n\t\tsignals.push(...this._detectConsecutiveWithoutVerification(history));\n\t\tsignals.push(...this._detectUnverifiedHypothesis(history));\n\t\tsignals.push(...this._detectNoAlternativesExplored(history, branches));\n\t\tsignals.push(...this._detectMonotonicType(history));\n\t\tsignals.push(...this._detectConfidenceDrift(history));\n\t\tsignals.push(...this._detectHealthyVerification(history));\n\t\treturn signals;\n\t}\n\n\t/** Detect runs of 3+ consecutive thoughts without verification. */\n\tprivate _detectConsecutiveWithoutVerification(history: ThoughtData[]): PatternSignal[] {\n\t\tconst signals: PatternSignal[] = [];\n\t\tlet runStart = 0;\n\t\tfor (let i = 0; i < history.length; i++) {\n\t\t\tconst type = history[i]!.thought_type ?? 'regular';\n\t\t\tif (type === 'verification') {\n\t\t\t\trunStart = i + 1;\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tif (i - runStart + 1 >= 3) {\n\t\t\t\tconst start = history[runStart]!.thought_number ?? runStart + 1;\n\t\t\t\tconst end = history[i]!.thought_number ?? i + 1;\n\t\t\t\tsignals.push({\n\t\t\t\t\tpattern: 'consecutive_without_verification',\n\t\t\t\t\tseverity: 'warning',\n\t\t\t\t\tmessage: `3+ consecutive thoughts (${start}-${end}) without verification`,\n\t\t\t\t\tthought_range: [start, end],\n\t\t\t\t});\n\t\t\t\trunStart = i + 1;\n\t\t\t}\n\t\t}\n\t\treturn signals;\n\t}\n\n\t/** Detect hypothesis thoughts not verified within 3 subsequent thoughts. */\n\tprivate _detectUnverifiedHypothesis(history: ThoughtData[]): PatternSignal[] {\n\t\tconst signals: PatternSignal[] = [];\n\t\tfor (let i = 0; i < history.length; i++) {\n\t\t\tif (history[i]!.thought_type !== 'hypothesis') continue;\n\t\t\tconst remaining = history.length - i - 1;\n\t\t\tif (remaining < 3) continue;\n\t\t\tconst lookahead = history.slice(i + 1, i + 4);\n\t\t\tconst hasVerification = lookahead.some((t) => t.thought_type === 'verification');\n\t\t\tif (!hasVerification) {\n\t\t\t\tconst n = history[i]!.thought_number ?? i + 1;\n\t\t\t\tsignals.push({\n\t\t\t\t\tpattern: 'unverified_hypothesis',\n\t\t\t\t\tseverity: 'warning',\n\t\t\t\t\tmessage: `Hypothesis at thought ${n} has not been verified within 3 thoughts`,\n\t\t\t\t\tthought_range: [n, n],\n\t\t\t\t});\n\t\t\t}\n\t\t}\n\t\treturn signals;\n\t}\n\n\t/** Detect 5+ thoughts with no critique and no branches. */\n\tprivate _detectNoAlternativesExplored(\n\t\thistory: ThoughtData[],\n\t\tbranches: Record<string, ThoughtData[]>\n\t): PatternSignal[] {\n\t\tif (history.length < 5) return [];\n\t\tif (history.some((t) => t.thought_type === 'critique')) return [];\n\t\tif (Object.keys(branches).length > 0) return [];\n\t\tconst start = history[0]!.thought_number ?? 1;\n\t\tconst end = history[history.length - 1]!.thought_number ?? history.length;\n\t\treturn [\n\t\t\t{\n\t\t\t\tpattern: 'no_alternatives_explored',\n\t\t\t\tseverity: 'warning',\n\t\t\t\tmessage: '5+ thoughts with no critique or branching — consider exploring alternatives',\n\t\t\t\tthought_range: [start, end],\n\t\t\t},\n\t\t];\n\t}\n\n\t/**\n\t * Detect runs of 4+ consecutive thoughts with the same thought_type.\n\t * Only fires when history has ≥5 thoughts and at least one explicitly set thought_type.\n\t */\n\tprivate _detectMonotonicType(history: ThoughtData[]): PatternSignal[] {\n\t\tif (history.length < 5) return [];\n\t\tconst hasExplicitType = history.some((t) => t.thought_type !== undefined);\n\t\tif (!hasExplicitType) return [];\n\n\t\tconst signals: PatternSignal[] = [];\n\t\tlet runType = history[0]!.thought_type ?? 'regular';\n\t\tlet runStart = 0;\n\t\tlet runLength = 1;\n\n\t\tfor (let i = 1; i < history.length; i++) {\n\t\t\tconst type = history[i]!.thought_type ?? 'regular';\n\t\t\tif (type === runType) {\n\t\t\t\trunLength++;\n\t\t\t} else {\n\t\t\t\tthis._flushMonotonicRun(history, runStart, runLength, runType, signals);\n\t\t\t\trunType = type;\n\t\t\t\trunStart = i;\n\t\t\t\trunLength = 1;\n\t\t\t}\n\t\t}\n\t\tthis._flushMonotonicRun(history, runStart, runLength, runType, signals);\n\t\treturn signals;\n\t}\n\n\t/** Emit a monotonic_type signal if the run length qualifies. */\n\tprivate _flushMonotonicRun(\n\t\thistory: ThoughtData[],\n\t\trunStart: number,\n\t\trunLength: number,\n\t\trunType: string,\n\t\tsignals: PatternSignal[]\n\t): void {\n\t\tif (runLength < 4) return;\n\t\tconst start = history[runStart]!.thought_number ?? runStart + 1;\n\t\tconst end = history[runStart + runLength - 1]!.thought_number ?? runStart + runLength;\n\t\tsignals.push({\n\t\t\tpattern: 'monotonic_type',\n\t\t\tseverity: 'warning',\n\t\t\tmessage: `4+ consecutive '${runType}' thoughts (${start}-${end}) — consider varying approach`,\n\t\t\tthought_range: [start, end],\n\t\t});\n\t}\n\n\t/** Detect runs of 3+ consecutive thoughts with strictly decreasing confidence. */\n\tprivate _detectConfidenceDrift(history: ThoughtData[]): PatternSignal[] {\n\t\tconst signals: PatternSignal[] = [];\n\t\tlet runStart = -1;\n\t\tlet runLength = 0;\n\t\tlet prevConf = -1;\n\n\t\tfor (let i = 0; i < history.length; i++) {\n\t\t\tconst conf = history[i]!.confidence;\n\t\t\tif (conf === undefined) {\n\t\t\t\tthis._flushDriftRun(history, runStart, runLength, signals);\n\t\t\t\trunStart = -1;\n\t\t\t\trunLength = 0;\n\t\t\t\tprevConf = -1;\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tif (runLength === 0) {\n\t\t\t\trunStart = i;\n\t\t\t\trunLength = 1;\n\t\t\t\tprevConf = conf;\n\t\t\t} else if (conf < prevConf) {\n\t\t\t\trunLength++;\n\t\t\t\tprevConf = conf;\n\t\t\t} else {\n\t\t\t\tthis._flushDriftRun(history, runStart, runLength, signals);\n\t\t\t\trunStart = i;\n\t\t\t\trunLength = 1;\n\t\t\t\tprevConf = conf;\n\t\t\t}\n\t\t}\n\t\tthis._flushDriftRun(history, runStart, runLength, signals);\n\t\treturn signals;\n\t}\n\n\t/** Emit a confidence_drift signal if the run length qualifies. */\n\tprivate _flushDriftRun(\n\t\thistory: ThoughtData[],\n\t\trunStart: number,\n\t\trunLength: number,\n\t\tsignals: PatternSignal[]\n\t): void {\n\t\tif (runLength < 3) return;\n\t\tconst first = history[runStart];\n\t\tconst last = history[runStart + runLength - 1];\n\t\tif (!first || !last) return;\n\t\tif (first.confidence === undefined || last.confidence === undefined) return;\n\t\tconst start = first.thought_number ?? runStart + 1;\n\t\tconst end = last.thought_number ?? runStart + runLength;\n\t\tconst firstConf = first.confidence;\n\t\tconst lastConf = last.confidence;\n\t\tsignals.push({\n\t\t\tpattern: 'confidence_drift',\n\t\t\tseverity: 'warning',\n\t\t\tmessage: `Confidence decreasing across thoughts ${start}-${end} (${firstConf} → ${lastConf})`,\n\t\t\tthought_range: [start, end],\n\t\t});\n\t}\n\n\t/** Detect hypothesis verified within 3 subsequent thoughts — positive signal. */\n\tprivate _detectHealthyVerification(history: ThoughtData[]): PatternSignal[] {\n\t\tconst signals: PatternSignal[] = [];\n\t\tfor (let i = 0; i < history.length; i++) {\n\t\t\tif (history[i]!.thought_type !== 'hypothesis') continue;\n\t\t\tconst hypId = history[i]!.hypothesis_id;\n\t\t\tconst lookahead = history.slice(i + 1, i + 4);\n\t\t\tconst verifier = lookahead.find(\n\t\t\t\t(t) =>\n\t\t\t\t\tt.thought_type === 'verification' &&\n\t\t\t\t\t(t.hypothesis_id === hypId ||\n\t\t\t\t\t\tt.verification_target === (history[i]!.thought_number ?? i + 1))\n\t\t\t);\n\t\t\tif (verifier) {\n\t\t\t\tconst n = history[i]!.thought_number ?? i + 1;\n\t\t\t\tconst m = verifier.thought_number ?? history.indexOf(verifier) + 1;\n\t\t\t\tsignals.push({\n\t\t\t\t\tpattern: 'healthy_verification',\n\t\t\t\t\tseverity: 'info',\n\t\t\t\t\tmessage: `Hypothesis at thought ${n} verified at thought ${m} — good practice`,\n\t\t\t\t\tthought_range: [n, m],\n\t\t\t\t});\n\t\t\t}\n\t\t}\n\t\treturn signals;\n\t}\n}\n"],"names":["PatternDetector","history","branches","signals","runStart","i","type","start","end","remaining","lookahead","hasVerification","t","n","Object","hasExplicitType","undefined","runType","runLength","prevConf","conf","first","last","firstConf","lastConf","hypId","verifier","m"],"mappings":"AAyCO,MAAMA;IAQL,sBACNC,OAAsB,EACtBC,QAAuC,EACrB;QAClB,IAAID,AAAmB,MAAnBA,QAAQ,MAAM,EAAQ,OAAO,EAAE;QAEnC,MAAME,UAA2B,EAAE;QACnCA,QAAQ,IAAI,IAAI,IAAI,CAAC,qCAAqC,CAACF;QAC3DE,QAAQ,IAAI,IAAI,IAAI,CAAC,2BAA2B,CAACF;QACjDE,QAAQ,IAAI,IAAI,IAAI,CAAC,6BAA6B,CAACF,SAASC;QAC5DC,QAAQ,IAAI,IAAI,IAAI,CAAC,oBAAoB,CAACF;QAC1CE,QAAQ,IAAI,IAAI,IAAI,CAAC,sBAAsB,CAACF;QAC5CE,QAAQ,IAAI,IAAI,IAAI,CAAC,0BAA0B,CAACF;QAChD,OAAOE;IACR;IAGQ,sCAAsCF,OAAsB,EAAmB;QACtF,MAAME,UAA2B,EAAE;QACnC,IAAIC,WAAW;QACf,IAAK,IAAIC,IAAI,GAAGA,IAAIJ,QAAQ,MAAM,EAAEI,IAAK;YACxC,MAAMC,OAAOL,OAAO,CAACI,EAAE,CAAE,YAAY,IAAI;YACzC,IAAIC,AAAS,mBAATA,MAAyB;gBAC5BF,WAAWC,IAAI;gBACf;YACD;YACA,IAAIA,IAAID,WAAW,KAAK,GAAG;gBAC1B,MAAMG,QAAQN,OAAO,CAACG,SAAS,CAAE,cAAc,IAAIA,WAAW;gBAC9D,MAAMI,MAAMP,OAAO,CAACI,EAAE,CAAE,cAAc,IAAIA,IAAI;gBAC9CF,QAAQ,IAAI,CAAC;oBACZ,SAAS;oBACT,UAAU;oBACV,SAAS,CAAC,yBAAyB,EAAEI,MAAM,CAAC,EAAEC,IAAI,sBAAsB,CAAC;oBACzE,eAAe;wBAACD;wBAAOC;qBAAI;gBAC5B;gBACAJ,WAAWC,IAAI;YAChB;QACD;QACA,OAAOF;IACR;IAGQ,4BAA4BF,OAAsB,EAAmB;QAC5E,MAAME,UAA2B,EAAE;QACnC,IAAK,IAAIE,IAAI,GAAGA,IAAIJ,QAAQ,MAAM,EAAEI,IAAK;YACxC,IAAIJ,AAA6B,iBAA7BA,OAAO,CAACI,EAAE,CAAE,YAAY,EAAmB;YAC/C,MAAMI,YAAYR,QAAQ,MAAM,GAAGI,IAAI;YACvC,IAAII,YAAY,GAAG;YACnB,MAAMC,YAAYT,QAAQ,KAAK,CAACI,IAAI,GAAGA,IAAI;YAC3C,MAAMM,kBAAkBD,UAAU,IAAI,CAAC,CAACE,IAAMA,AAAmB,mBAAnBA,EAAE,YAAY;YAC5D,IAAI,CAACD,iBAAiB;gBACrB,MAAME,IAAIZ,OAAO,CAACI,EAAE,CAAE,cAAc,IAAIA,IAAI;gBAC5CF,QAAQ,IAAI,CAAC;oBACZ,SAAS;oBACT,UAAU;oBACV,SAAS,CAAC,sBAAsB,EAAEU,EAAE,wCAAwC,CAAC;oBAC7E,eAAe;wBAACA;wBAAGA;qBAAE;gBACtB;YACD;QACD;QACA,OAAOV;IACR;IAGQ,8BACPF,OAAsB,EACtBC,QAAuC,EACrB;QAClB,IAAID,QAAQ,MAAM,GAAG,GAAG,OAAO,EAAE;QACjC,IAAIA,QAAQ,IAAI,CAAC,CAACW,IAAMA,AAAmB,eAAnBA,EAAE,YAAY,GAAkB,OAAO,EAAE;QACjE,IAAIE,OAAO,IAAI,CAACZ,UAAU,MAAM,GAAG,GAAG,OAAO,EAAE;QAC/C,MAAMK,QAAQN,OAAO,CAAC,EAAE,CAAE,cAAc,IAAI;QAC5C,MAAMO,MAAMP,OAAO,CAACA,QAAQ,MAAM,GAAG,EAAE,CAAE,cAAc,IAAIA,QAAQ,MAAM;QACzE,OAAO;YACN;gBACC,SAAS;gBACT,UAAU;gBACV,SAAS;gBACT,eAAe;oBAACM;oBAAOC;iBAAI;YAC5B;SACA;IACF;IAMQ,qBAAqBP,OAAsB,EAAmB;QACrE,IAAIA,QAAQ,MAAM,GAAG,GAAG,OAAO,EAAE;QACjC,MAAMc,kBAAkBd,QAAQ,IAAI,CAAC,CAACW,IAAMA,AAAmBI,WAAnBJ,EAAE,YAAY;QAC1D,IAAI,CAACG,iBAAiB,OAAO,EAAE;QAE/B,MAAMZ,UAA2B,EAAE;QACnC,IAAIc,UAAUhB,OAAO,CAAC,EAAE,CAAE,YAAY,IAAI;QAC1C,IAAIG,WAAW;QACf,IAAIc,YAAY;QAEhB,IAAK,IAAIb,IAAI,GAAGA,IAAIJ,QAAQ,MAAM,EAAEI,IAAK;YACxC,MAAMC,OAAOL,OAAO,CAACI,EAAE,CAAE,YAAY,IAAI;YACzC,IAAIC,SAASW,SACZC;iBACM;gBACN,IAAI,CAAC,kBAAkB,CAACjB,SAASG,UAAUc,WAAWD,SAASd;gBAC/Dc,UAAUX;gBACVF,WAAWC;gBACXa,YAAY;YACb;QACD;QACA,IAAI,CAAC,kBAAkB,CAACjB,SAASG,UAAUc,WAAWD,SAASd;QAC/D,OAAOA;IACR;IAGQ,mBACPF,OAAsB,EACtBG,QAAgB,EAChBc,SAAiB,EACjBD,OAAe,EACfd,OAAwB,EACjB;QACP,IAAIe,YAAY,GAAG;QACnB,MAAMX,QAAQN,OAAO,CAACG,SAAS,CAAE,cAAc,IAAIA,WAAW;QAC9D,MAAMI,MAAMP,OAAO,CAACG,WAAWc,YAAY,EAAE,CAAE,cAAc,IAAId,WAAWc;QAC5Ef,QAAQ,IAAI,CAAC;YACZ,SAAS;YACT,UAAU;YACV,SAAS,CAAC,gBAAgB,EAAEc,QAAQ,YAAY,EAAEV,MAAM,CAAC,EAAEC,IAAI,6BAA6B,CAAC;YAC7F,eAAe;gBAACD;gBAAOC;aAAI;QAC5B;IACD;IAGQ,uBAAuBP,OAAsB,EAAmB;QACvE,MAAME,UAA2B,EAAE;QACnC,IAAIC,WAAW;QACf,IAAIc,YAAY;QAChB,IAAIC,WAAW;QAEf,IAAK,IAAId,IAAI,GAAGA,IAAIJ,QAAQ,MAAM,EAAEI,IAAK;YACxC,MAAMe,OAAOnB,OAAO,CAACI,EAAE,CAAE,UAAU;YACnC,IAAIe,AAASJ,WAATI,MAAoB;gBACvB,IAAI,CAAC,cAAc,CAACnB,SAASG,UAAUc,WAAWf;gBAClDC,WAAW;gBACXc,YAAY;gBACZC,WAAW;gBACX;YACD;YACA,IAAID,AAAc,MAAdA,WAAiB;gBACpBd,WAAWC;gBACXa,YAAY;gBACZC,WAAWC;YACZ,OAAO,IAAIA,OAAOD,UAAU;gBAC3BD;gBACAC,WAAWC;YACZ,OAAO;gBACN,IAAI,CAAC,cAAc,CAACnB,SAASG,UAAUc,WAAWf;gBAClDC,WAAWC;gBACXa,YAAY;gBACZC,WAAWC;YACZ;QACD;QACA,IAAI,CAAC,cAAc,CAACnB,SAASG,UAAUc,WAAWf;QAClD,OAAOA;IACR;IAGQ,eACPF,OAAsB,EACtBG,QAAgB,EAChBc,SAAiB,EACjBf,OAAwB,EACjB;QACP,IAAIe,YAAY,GAAG;QACnB,MAAMG,QAAQpB,OAAO,CAACG,SAAS;QAC/B,MAAMkB,OAAOrB,OAAO,CAACG,WAAWc,YAAY,EAAE;QAC9C,IAAI,CAACG,SAAS,CAACC,MAAM;QACrB,IAAID,AAAqBL,WAArBK,MAAM,UAAU,IAAkBC,AAAoBN,WAApBM,KAAK,UAAU,EAAgB;QACrE,MAAMf,QAAQc,MAAM,cAAc,IAAIjB,WAAW;QACjD,MAAMI,MAAMc,KAAK,cAAc,IAAIlB,WAAWc;QAC9C,MAAMK,YAAYF,MAAM,UAAU;QAClC,MAAMG,WAAWF,KAAK,UAAU;QAChCnB,QAAQ,IAAI,CAAC;YACZ,SAAS;YACT,UAAU;YACV,SAAS,CAAC,sCAAsC,EAAEI,MAAM,CAAC,EAAEC,IAAI,EAAE,EAAEe,UAAU,GAAG,EAAEC,SAAS,CAAC,CAAC;YAC7F,eAAe;gBAACjB;gBAAOC;aAAI;QAC5B;IACD;IAGQ,2BAA2BP,OAAsB,EAAmB;QAC3E,MAAME,UAA2B,EAAE;QACnC,IAAK,IAAIE,IAAI,GAAGA,IAAIJ,QAAQ,MAAM,EAAEI,IAAK;YACxC,IAAIJ,AAA6B,iBAA7BA,OAAO,CAACI,EAAE,CAAE,YAAY,EAAmB;YAC/C,MAAMoB,QAAQxB,OAAO,CAACI,EAAE,CAAE,aAAa;YACvC,MAAMK,YAAYT,QAAQ,KAAK,CAACI,IAAI,GAAGA,IAAI;YAC3C,MAAMqB,WAAWhB,UAAU,IAAI,CAC9B,CAACE,IACAA,AAAmB,mBAAnBA,EAAE,YAAY,IACbA,CAAAA,EAAE,aAAa,KAAKa,SACpBb,EAAE,mBAAmB,KAAMX,CAAAA,OAAO,CAACI,EAAE,CAAE,cAAc,IAAIA,IAAI,EAAC;YAEjE,IAAIqB,UAAU;gBACb,MAAMb,IAAIZ,OAAO,CAACI,EAAE,CAAE,cAAc,IAAIA,IAAI;gBAC5C,MAAMsB,IAAID,SAAS,cAAc,IAAIzB,QAAQ,OAAO,CAACyB,YAAY;gBACjEvB,QAAQ,IAAI,CAAC;oBACZ,SAAS;oBACT,UAAU;oBACV,SAAS,CAAC,sBAAsB,EAAEU,EAAE,qBAAqB,EAAEc,EAAE,gBAAgB,CAAC;oBAC9E,eAAe;wBAACd;wBAAGc;qBAAE;gBACtB;YACD;QACD;QACA,OAAOxB;IACR;AACD"}
+{"version":3,"file":"core/evaluator/PatternDetector.js","sources":["../../../src/core/evaluator/PatternDetector.ts"],"sourcesContent":["/**\n * Pattern detection for sequential thinking reasoning analytics.\n *\n * Provides the {@link PatternDetector} class — a stateless service that detects\n * reasoning patterns (anti-patterns and positive signals) from thought history.\n * Extracted from {@link ThoughtEvaluator} to keep concerns isolated and reduce\n * cyclomatic complexity per function.\n *\n * Detected patterns:\n * - `consecutive_without_verification` (warning) — 3+ consecutive regular thoughts\n *   without a verification step\n * - `unverified_hypothesis` (warning) — hypothesis not verified within 3 subsequent\n *   thoughts\n * - `no_alternatives_explored` (info) — 5+ thoughts with no critique and no branches\n * - `monotonic_type` (info) — 4+ consecutive thoughts with the same thought_type\n *   (requires ≥1 explicit type and ≥5 thoughts)\n * - `confidence_drift` (warning) — 3+ consecutive thoughts with strictly decreasing\n *   confidence\n * - `healthy_verification` (info) — hypothesis verified within 3 subsequent thoughts\n *\n * @module core/evaluator/PatternDetector\n */\n\nimport type { PatternSignal } from '../reasoning.js';\nimport type { ThoughtData } from '../thought.js';\nimport type { VerificationLinks } from './VerificationLinks.js';\n\n/**\n * Stateless service that detects reasoning patterns from thought history.\n *\n * @remarks\n * All methods are pure computations — no side effects, no I/O, no internal state.\n * Designed to be registered as transient in the DI container or composed directly\n * into {@link ThoughtEvaluator}.\n *\n * @example\n * ```typescript\n * const detector = new PatternDetector();\n * const signals = detector.computePatternSignals(history, branches);\n * const warnings = signals.filter((s) => s.severity === 'warning');\n * ```\n */\nexport class PatternDetector {\n\t/**\n\t * Detect all reasoning patterns from history and branches.\n\t *\n\t * @param history - All thoughts in the current session\n\t * @param branches - Map of branch IDs to their thought arrays\n\t * @returns Array of detected pattern signals (possibly empty)\n\t */\n\tpublic computePatternSignals(\n\t\thistory: ThoughtData[],\n\t\tbranches: Record<string, ThoughtData[]>,\n\t\tlinks: VerificationLinks\n\t): PatternSignal[] {\n\t\tif (history.length === 0) return [];\n\n\t\tconst signals: PatternSignal[] = [];\n\t\tsignals.push(...this._detectConsecutiveWithoutVerification(history));\n\t\tsignals.push(...this._detectUnverifiedHypothesis(history, links));\n\t\tsignals.push(...this._detectNoAlternativesExplored(history, branches));\n\t\tsignals.push(...this._detectMonotonicType(history));\n\t\tsignals.push(...this._detectConfidenceDrift(history));\n\t\tsignals.push(...this._detectHealthyVerification(history, links));\n\t\treturn signals;\n\t}\n\n\t/** Detect runs of 3+ consecutive thoughts without verification. */\n\tprivate _detectConsecutiveWithoutVerification(history: ThoughtData[]): PatternSignal[] {\n\t\tconst signals: PatternSignal[] = [];\n\t\tlet runStart = 0;\n\t\tfor (let i = 0; i < history.length; i++) {\n\t\t\tconst type = history[i]!.thought_type ?? 'regular';\n\t\t\tif (type === 'verification') {\n\t\t\t\trunStart = i + 1;\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tif (i - runStart + 1 >= 3) {\n\t\t\t\tconst start = history[runStart]!.thought_number ?? runStart + 1;\n\t\t\t\tconst end = history[i]!.thought_number ?? i + 1;\n\t\t\t\tsignals.push({\n\t\t\t\t\tpattern: 'consecutive_without_verification',\n\t\t\t\t\tseverity: 'warning',\n\t\t\t\t\tmessage: `3+ consecutive thoughts (${start}-${end}) without verification`,\n\t\t\t\t\tthought_range: [start, end],\n\t\t\t\t});\n\t\t\t\trunStart = i + 1;\n\t\t\t}\n\t\t}\n\t\treturn signals;\n\t}\n\n\t/** Detect hypothesis thoughts not verified within 3 subsequent thoughts. */\n\tprivate _detectUnverifiedHypothesis(\n\t\thistory: ThoughtData[],\n\t\tlinks: VerificationLinks\n\t): PatternSignal[] {\n\t\tconst signals: PatternSignal[] = [];\n\t\tfor (let i = 0; i < history.length; i++) {\n\t\t\tconst hypothesis = history[i]!;\n\t\t\tif (\n\t\t\t\thypothesis.thought_type !== 'hypothesis' ||\n\t\t\t\t!hypothesis.id ||\n\t\t\t\t!links.hypothesisIds.has(hypothesis.id)\n\t\t\t) {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tconst remaining = history.length - i - 1;\n\t\t\tif (remaining < 3) continue;\n\t\t\tconst lookahead = history.slice(i + 1, i + 4);\n\t\t\tconst hasVerification = lookahead.some((verifier) => links.verifies(verifier, hypothesis));\n\t\t\tif (!hasVerification) {\n\t\t\t\tconst n = hypothesis.thought_number ?? i + 1;\n\t\t\t\tsignals.push({\n\t\t\t\t\tpattern: 'unverified_hypothesis',\n\t\t\t\t\tseverity: 'warning',\n\t\t\t\t\tmessage: `Hypothesis at thought ${n} has not been verified within 3 thoughts`,\n\t\t\t\t\tthought_range: [n, n],\n\t\t\t\t});\n\t\t\t}\n\t\t}\n\t\treturn signals;\n\t}\n\n\t/** Detect 5+ thoughts with no critique and no branches. */\n\tprivate _detectNoAlternativesExplored(\n\t\thistory: ThoughtData[],\n\t\tbranches: Record<string, ThoughtData[]>\n\t): PatternSignal[] {\n\t\tif (history.length < 5) return [];\n\t\tif (history.some((t) => t.thought_type === 'critique')) return [];\n\t\tif (Object.keys(branches).length > 0) return [];\n\t\tconst start = history[0]!.thought_number ?? 1;\n\t\tconst end = history[history.length - 1]!.thought_number ?? history.length;\n\t\treturn [\n\t\t\t{\n\t\t\t\tpattern: 'no_alternatives_explored',\n\t\t\t\tseverity: 'warning',\n\t\t\t\tmessage: '5+ thoughts with no critique or branching — consider exploring alternatives',\n\t\t\t\tthought_range: [start, end],\n\t\t\t},\n\t\t];\n\t}\n\n\t/**\n\t * Detect runs of 4+ consecutive thoughts with the same thought_type.\n\t * Only fires when history has ≥5 thoughts and at least one explicitly set thought_type.\n\t */\n\tprivate _detectMonotonicType(history: ThoughtData[]): PatternSignal[] {\n\t\tif (history.length < 5) return [];\n\t\tconst hasExplicitType = history.some((t) => t.thought_type !== undefined);\n\t\tif (!hasExplicitType) return [];\n\n\t\tconst signals: PatternSignal[] = [];\n\t\tlet runType = history[0]!.thought_type ?? 'regular';\n\t\tlet runStart = 0;\n\t\tlet runLength = 1;\n\n\t\tfor (let i = 1; i < history.length; i++) {\n\t\t\tconst type = history[i]!.thought_type ?? 'regular';\n\t\t\tif (type === runType) {\n\t\t\t\trunLength++;\n\t\t\t} else {\n\t\t\t\tthis._flushMonotonicRun(history, runStart, runLength, runType, signals);\n\t\t\t\trunType = type;\n\t\t\t\trunStart = i;\n\t\t\t\trunLength = 1;\n\t\t\t}\n\t\t}\n\t\tthis._flushMonotonicRun(history, runStart, runLength, runType, signals);\n\t\treturn signals;\n\t}\n\n\t/** Emit a monotonic_type signal if the run length qualifies. */\n\tprivate _flushMonotonicRun(\n\t\thistory: ThoughtData[],\n\t\trunStart: number,\n\t\trunLength: number,\n\t\trunType: string,\n\t\tsignals: PatternSignal[]\n\t): void {\n\t\tif (runLength < 4) return;\n\t\tconst start = history[runStart]!.thought_number ?? runStart + 1;\n\t\tconst end = history[runStart + runLength - 1]!.thought_number ?? runStart + runLength;\n\t\tsignals.push({\n\t\t\tpattern: 'monotonic_type',\n\t\t\tseverity: 'warning',\n\t\t\tmessage: `4+ consecutive '${runType}' thoughts (${start}-${end}) — consider varying approach`,\n\t\t\tthought_range: [start, end],\n\t\t});\n\t}\n\n\t/** Detect runs of 3+ consecutive thoughts with strictly decreasing confidence. */\n\tprivate _detectConfidenceDrift(history: ThoughtData[]): PatternSignal[] {\n\t\tconst signals: PatternSignal[] = [];\n\t\tlet runStart = -1;\n\t\tlet runLength = 0;\n\t\tlet prevConf = -1;\n\n\t\tfor (let i = 0; i < history.length; i++) {\n\t\t\tconst conf = history[i]!.confidence;\n\t\t\tif (conf === undefined) {\n\t\t\t\tthis._flushDriftRun(history, runStart, runLength, signals);\n\t\t\t\trunStart = -1;\n\t\t\t\trunLength = 0;\n\t\t\t\tprevConf = -1;\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tif (runLength === 0) {\n\t\t\t\trunStart = i;\n\t\t\t\trunLength = 1;\n\t\t\t\tprevConf = conf;\n\t\t\t} else if (conf < prevConf) {\n\t\t\t\trunLength++;\n\t\t\t\tprevConf = conf;\n\t\t\t} else {\n\t\t\t\tthis._flushDriftRun(history, runStart, runLength, signals);\n\t\t\t\trunStart = i;\n\t\t\t\trunLength = 1;\n\t\t\t\tprevConf = conf;\n\t\t\t}\n\t\t}\n\t\tthis._flushDriftRun(history, runStart, runLength, signals);\n\t\treturn signals;\n\t}\n\n\t/** Emit a confidence_drift signal if the run length qualifies. */\n\tprivate _flushDriftRun(\n\t\thistory: ThoughtData[],\n\t\trunStart: number,\n\t\trunLength: number,\n\t\tsignals: PatternSignal[]\n\t): void {\n\t\tif (runLength < 3) return;\n\t\tconst first = history[runStart];\n\t\tconst last = history[runStart + runLength - 1];\n\t\tif (!first || !last) return;\n\t\tif (first.confidence === undefined || last.confidence === undefined) return;\n\t\tconst start = first.thought_number ?? runStart + 1;\n\t\tconst end = last.thought_number ?? runStart + runLength;\n\t\tconst firstConf = first.confidence;\n\t\tconst lastConf = last.confidence;\n\t\tsignals.push({\n\t\t\tpattern: 'confidence_drift',\n\t\t\tseverity: 'warning',\n\t\t\tmessage: `Confidence decreasing across thoughts ${start}-${end} (${firstConf} → ${lastConf})`,\n\t\t\tthought_range: [start, end],\n\t\t});\n\t}\n\n\t/** Detect hypothesis verified within 3 subsequent thoughts — positive signal. */\n\tprivate _detectHealthyVerification(\n\t\thistory: ThoughtData[],\n\t\tlinks: VerificationLinks\n\t): PatternSignal[] {\n\t\tconst signals: PatternSignal[] = [];\n\t\tfor (let i = 0; i < history.length; i++) {\n\t\t\tconst hypothesis = history[i]!;\n\t\t\tif (\n\t\t\t\thypothesis.thought_type !== 'hypothesis' ||\n\t\t\t\t!hypothesis.id ||\n\t\t\t\t!links.hypothesisIds.has(hypothesis.id)\n\t\t\t) {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tconst lookahead = history.slice(i + 1, i + 4);\n\t\t\tconst verifier = lookahead.find((candidate) => links.verifies(candidate, hypothesis));\n\t\t\tif (verifier) {\n\t\t\t\tconst n = hypothesis.thought_number ?? i + 1;\n\t\t\t\tconst m = verifier.thought_number ?? history.indexOf(verifier) + 1;\n\t\t\t\tsignals.push({\n\t\t\t\t\tpattern: 'healthy_verification',\n\t\t\t\t\tseverity: 'info',\n\t\t\t\t\tmessage: `Hypothesis at thought ${n} verified at thought ${m} — good practice`,\n\t\t\t\t\tthought_range: [n, m],\n\t\t\t\t});\n\t\t\t}\n\t\t}\n\t\treturn signals;\n\t}\n}\n"],"names":["PatternDetector","history","branches","links","signals","runStart","i","type","start","end","hypothesis","remaining","lookahead","hasVerification","verifier","n","t","Object","hasExplicitType","undefined","runType","runLength","prevConf","conf","first","last","firstConf","lastConf","candidate","m"],"mappings":"AA0CO,MAAMA;IAQL,sBACNC,OAAsB,EACtBC,QAAuC,EACvCC,KAAwB,EACN;QAClB,IAAIF,AAAmB,MAAnBA,QAAQ,MAAM,EAAQ,OAAO,EAAE;QAEnC,MAAMG,UAA2B,EAAE;QACnCA,QAAQ,IAAI,IAAI,IAAI,CAAC,qCAAqC,CAACH;QAC3DG,QAAQ,IAAI,IAAI,IAAI,CAAC,2BAA2B,CAACH,SAASE;QAC1DC,QAAQ,IAAI,IAAI,IAAI,CAAC,6BAA6B,CAACH,SAASC;QAC5DE,QAAQ,IAAI,IAAI,IAAI,CAAC,oBAAoB,CAACH;QAC1CG,QAAQ,IAAI,IAAI,IAAI,CAAC,sBAAsB,CAACH;QAC5CG,QAAQ,IAAI,IAAI,IAAI,CAAC,0BAA0B,CAACH,SAASE;QACzD,OAAOC;IACR;IAGQ,sCAAsCH,OAAsB,EAAmB;QACtF,MAAMG,UAA2B,EAAE;QACnC,IAAIC,WAAW;QACf,IAAK,IAAIC,IAAI,GAAGA,IAAIL,QAAQ,MAAM,EAAEK,IAAK;YACxC,MAAMC,OAAON,OAAO,CAACK,EAAE,CAAE,YAAY,IAAI;YACzC,IAAIC,AAAS,mBAATA,MAAyB;gBAC5BF,WAAWC,IAAI;gBACf;YACD;YACA,IAAIA,IAAID,WAAW,KAAK,GAAG;gBAC1B,MAAMG,QAAQP,OAAO,CAACI,SAAS,CAAE,cAAc,IAAIA,WAAW;gBAC9D,MAAMI,MAAMR,OAAO,CAACK,EAAE,CAAE,cAAc,IAAIA,IAAI;gBAC9CF,QAAQ,IAAI,CAAC;oBACZ,SAAS;oBACT,UAAU;oBACV,SAAS,CAAC,yBAAyB,EAAEI,MAAM,CAAC,EAAEC,IAAI,sBAAsB,CAAC;oBACzE,eAAe;wBAACD;wBAAOC;qBAAI;gBAC5B;gBACAJ,WAAWC,IAAI;YAChB;QACD;QACA,OAAOF;IACR;IAGQ,4BACPH,OAAsB,EACtBE,KAAwB,EACN;QAClB,MAAMC,UAA2B,EAAE;QACnC,IAAK,IAAIE,IAAI,GAAGA,IAAIL,QAAQ,MAAM,EAAEK,IAAK;YACxC,MAAMI,aAAaT,OAAO,CAACK,EAAE;YAC7B,IACCI,AAA4B,iBAA5BA,WAAW,YAAY,IACvB,CAACA,WAAW,EAAE,IACd,CAACP,MAAM,aAAa,CAAC,GAAG,CAACO,WAAW,EAAE,GAEtC;YAED,MAAMC,YAAYV,QAAQ,MAAM,GAAGK,IAAI;YACvC,IAAIK,YAAY,GAAG;YACnB,MAAMC,YAAYX,QAAQ,KAAK,CAACK,IAAI,GAAGA,IAAI;YAC3C,MAAMO,kBAAkBD,UAAU,IAAI,CAAC,CAACE,WAAaX,MAAM,QAAQ,CAACW,UAAUJ;YAC9E,IAAI,CAACG,iBAAiB;gBACrB,MAAME,IAAIL,WAAW,cAAc,IAAIJ,IAAI;gBAC3CF,QAAQ,IAAI,CAAC;oBACZ,SAAS;oBACT,UAAU;oBACV,SAAS,CAAC,sBAAsB,EAAEW,EAAE,wCAAwC,CAAC;oBAC7E,eAAe;wBAACA;wBAAGA;qBAAE;gBACtB;YACD;QACD;QACA,OAAOX;IACR;IAGQ,8BACPH,OAAsB,EACtBC,QAAuC,EACrB;QAClB,IAAID,QAAQ,MAAM,GAAG,GAAG,OAAO,EAAE;QACjC,IAAIA,QAAQ,IAAI,CAAC,CAACe,IAAMA,AAAmB,eAAnBA,EAAE,YAAY,GAAkB,OAAO,EAAE;QACjE,IAAIC,OAAO,IAAI,CAACf,UAAU,MAAM,GAAG,GAAG,OAAO,EAAE;QAC/C,MAAMM,QAAQP,OAAO,CAAC,EAAE,CAAE,cAAc,IAAI;QAC5C,MAAMQ,MAAMR,OAAO,CAACA,QAAQ,MAAM,GAAG,EAAE,CAAE,cAAc,IAAIA,QAAQ,MAAM;QACzE,OAAO;YACN;gBACC,SAAS;gBACT,UAAU;gBACV,SAAS;gBACT,eAAe;oBAACO;oBAAOC;iBAAI;YAC5B;SACA;IACF;IAMQ,qBAAqBR,OAAsB,EAAmB;QACrE,IAAIA,QAAQ,MAAM,GAAG,GAAG,OAAO,EAAE;QACjC,MAAMiB,kBAAkBjB,QAAQ,IAAI,CAAC,CAACe,IAAMA,AAAmBG,WAAnBH,EAAE,YAAY;QAC1D,IAAI,CAACE,iBAAiB,OAAO,EAAE;QAE/B,MAAMd,UAA2B,EAAE;QACnC,IAAIgB,UAAUnB,OAAO,CAAC,EAAE,CAAE,YAAY,IAAI;QAC1C,IAAII,WAAW;QACf,IAAIgB,YAAY;QAEhB,IAAK,IAAIf,IAAI,GAAGA,IAAIL,QAAQ,MAAM,EAAEK,IAAK;YACxC,MAAMC,OAAON,OAAO,CAACK,EAAE,CAAE,YAAY,IAAI;YACzC,IAAIC,SAASa,SACZC;iBACM;gBACN,IAAI,CAAC,kBAAkB,CAACpB,SAASI,UAAUgB,WAAWD,SAAShB;gBAC/DgB,UAAUb;gBACVF,WAAWC;gBACXe,YAAY;YACb;QACD;QACA,IAAI,CAAC,kBAAkB,CAACpB,SAASI,UAAUgB,WAAWD,SAAShB;QAC/D,OAAOA;IACR;IAGQ,mBACPH,OAAsB,EACtBI,QAAgB,EAChBgB,SAAiB,EACjBD,OAAe,EACfhB,OAAwB,EACjB;QACP,IAAIiB,YAAY,GAAG;QACnB,MAAMb,QAAQP,OAAO,CAACI,SAAS,CAAE,cAAc,IAAIA,WAAW;QAC9D,MAAMI,MAAMR,OAAO,CAACI,WAAWgB,YAAY,EAAE,CAAE,cAAc,IAAIhB,WAAWgB;QAC5EjB,QAAQ,IAAI,CAAC;YACZ,SAAS;YACT,UAAU;YACV,SAAS,CAAC,gBAAgB,EAAEgB,QAAQ,YAAY,EAAEZ,MAAM,CAAC,EAAEC,IAAI,6BAA6B,CAAC;YAC7F,eAAe;gBAACD;gBAAOC;aAAI;QAC5B;IACD;IAGQ,uBAAuBR,OAAsB,EAAmB;QACvE,MAAMG,UAA2B,EAAE;QACnC,IAAIC,WAAW;QACf,IAAIgB,YAAY;QAChB,IAAIC,WAAW;QAEf,IAAK,IAAIhB,IAAI,GAAGA,IAAIL,QAAQ,MAAM,EAAEK,IAAK;YACxC,MAAMiB,OAAOtB,OAAO,CAACK,EAAE,CAAE,UAAU;YACnC,IAAIiB,AAASJ,WAATI,MAAoB;gBACvB,IAAI,CAAC,cAAc,CAACtB,SAASI,UAAUgB,WAAWjB;gBAClDC,WAAW;gBACXgB,YAAY;gBACZC,WAAW;gBACX;YACD;YACA,IAAID,AAAc,MAAdA,WAAiB;gBACpBhB,WAAWC;gBACXe,YAAY;gBACZC,WAAWC;YACZ,OAAO,IAAIA,OAAOD,UAAU;gBAC3BD;gBACAC,WAAWC;YACZ,OAAO;gBACN,IAAI,CAAC,cAAc,CAACtB,SAASI,UAAUgB,WAAWjB;gBAClDC,WAAWC;gBACXe,YAAY;gBACZC,WAAWC;YACZ;QACD;QACA,IAAI,CAAC,cAAc,CAACtB,SAASI,UAAUgB,WAAWjB;QAClD,OAAOA;IACR;IAGQ,eACPH,OAAsB,EACtBI,QAAgB,EAChBgB,SAAiB,EACjBjB,OAAwB,EACjB;QACP,IAAIiB,YAAY,GAAG;QACnB,MAAMG,QAAQvB,OAAO,CAACI,SAAS;QAC/B,MAAMoB,OAAOxB,OAAO,CAACI,WAAWgB,YAAY,EAAE;QAC9C,IAAI,CAACG,SAAS,CAACC,MAAM;QACrB,IAAID,AAAqBL,WAArBK,MAAM,UAAU,IAAkBC,AAAoBN,WAApBM,KAAK,UAAU,EAAgB;QACrE,MAAMjB,QAAQgB,MAAM,cAAc,IAAInB,WAAW;QACjD,MAAMI,MAAMgB,KAAK,cAAc,IAAIpB,WAAWgB;QAC9C,MAAMK,YAAYF,MAAM,UAAU;QAClC,MAAMG,WAAWF,KAAK,UAAU;QAChCrB,QAAQ,IAAI,CAAC;YACZ,SAAS;YACT,UAAU;YACV,SAAS,CAAC,sCAAsC,EAAEI,MAAM,CAAC,EAAEC,IAAI,EAAE,EAAEiB,UAAU,GAAG,EAAEC,SAAS,CAAC,CAAC;YAC7F,eAAe;gBAACnB;gBAAOC;aAAI;QAC5B;IACD;IAGQ,2BACPR,OAAsB,EACtBE,KAAwB,EACN;QAClB,MAAMC,UAA2B,EAAE;QACnC,IAAK,IAAIE,IAAI,GAAGA,IAAIL,QAAQ,MAAM,EAAEK,IAAK;YACxC,MAAMI,aAAaT,OAAO,CAACK,EAAE;YAC7B,IACCI,AAA4B,iBAA5BA,WAAW,YAAY,IACvB,CAACA,WAAW,EAAE,IACd,CAACP,MAAM,aAAa,CAAC,GAAG,CAACO,WAAW,EAAE,GAEtC;YAED,MAAME,YAAYX,QAAQ,KAAK,CAACK,IAAI,GAAGA,IAAI;YAC3C,MAAMQ,WAAWF,UAAU,IAAI,CAAC,CAACgB,YAAczB,MAAM,QAAQ,CAACyB,WAAWlB;YACzE,IAAII,UAAU;gBACb,MAAMC,IAAIL,WAAW,cAAc,IAAIJ,IAAI;gBAC3C,MAAMuB,IAAIf,SAAS,cAAc,IAAIb,QAAQ,OAAO,CAACa,YAAY;gBACjEV,QAAQ,IAAI,CAAC;oBACZ,SAAS;oBACT,UAAU;oBACV,SAAS,CAAC,sBAAsB,EAAEW,EAAE,qBAAqB,EAAEc,EAAE,gBAAgB,CAAC;oBAC9E,eAAe;wBAACd;wBAAGc;qBAAE;gBACtB;YACD;QACD;QACA,OAAOzB;IACR;AACD"}

package/dist/core/evaluator/SignalComputer.d.ts

@@ -13,6 +13,7 @@
 import type { ThoughtType } from '../../contracts/reasoning-types.js';
 import type { ConfidenceSignals } from '../reasoning.js';
 import type { ThoughtData } from '../thought.js';
+import type { VerificationLinks } from './VerificationLinks.js';
 /** Result of {@link SignalComputer.computeStructuralQuality}. */
 interface StructuralQualityResult {
     score: number;
@@ -78,7 +79,7 @@ export declare class SignalComputer {
      * @param branches - Map of branch IDs to their thought arrays
      * @returns Computed confidence signals reflecting reasoning quality
      */
-    computeConfidenceSignals(history: ThoughtData[], branches: Record<string, ThoughtData[]>): ConfidenceSignals;
+    computeConfidenceSignals(history: ThoughtData[], branches: Record<string, ThoughtData[]>, links: VerificationLinks): ConfidenceSignals;
     /**
      * Compute the composite structural quality score and its components.
      *
@@ -95,7 +96,7 @@ export declare class SignalComputer {
      *
      * @returns Score + components, or `null` when history is empty
      */
-    computeStructuralQuality(history: ThoughtData[], branches: Record<string, ThoughtData[]>, typeDistribution: Record<ThoughtType, number>, confidences: number[]): StructuralQualityResult | null;
+    computeStructuralQuality(history: ThoughtData[], branches: Record<string, ThoughtData[]>, typeDistribution: Record<ThoughtType, number>, confidences: number[], links: VerificationLinks): StructuralQualityResult | null;
     /**
      * Compute confidence stability: 1 - stddev(confidences).
      *

package/dist/core/evaluator/SignalComputer.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"SignalComputer.d.ts","sourceRoot":"","sources":["../../../src/core/evaluator/SignalComputer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAEH,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,oCAAoC,CAAC;AACtE,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,iBAAiB,CAAC;AACzD,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,eAAe,CAAC;AAgBjD,iEAAiE;AACjE,UAAU,uBAAuB;IAChC,KAAK,EAAE,MAAM,CAAC;IACd,UAAU,EAAE;QACX,cAAc,EAAE,MAAM,CAAC;QACvB,qBAAqB,EAAE,MAAM,CAAC;QAC9B,gBAAgB,EAAE,MAAM,CAAC;QACzB,oBAAoB,EAAE,MAAM,GAAG,IAAI,CAAC;KACpC,CAAC;IACF,cAAc,EAAE;QACf,cAAc,EAAE,MAAM,CAAC;QACvB,qBAAqB,EAAE,MAAM,CAAC;QAC9B,gBAAgB,EAAE,MAAM,CAAC;QACzB,oBAAoB,EAAE,MAAM,GAAG,IAAI,CAAC;KACpC,CAAC;CACF;AAED;;;;;;GAMG;AACH,qBAAa,cAAc;IAC1B;;;;;;;;;;;;;;;;;;;;;;;;;;;OA2BG;IACH,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,eAAe,CAK5B;IAEX;;;OAGG;IACH,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,qBAAqB,CAIlC;IAEX;;;;;;OAMG;IACI,wBAAwB,CAC9B,OAAO,EAAE,WAAW,EAAE,EACtB,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,WAAW,EAAE,CAAC,GACrC,iBAAiB;IAiCpB;;;;;;;;;;;;;;;OAeG;IACI,wBAAwB,CAC9B,OAAO,EAAE,WAAW,EAAE,EACtB,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,WAAW,EAAE,CAAC,EACvC,gBAAgB,EAAE,MAAM,CAAC,WAAW,EAAE,MAAM,CAAC,EAC7C,WAAW,EAAE,MAAM,EAAE,GACnB,uBAAuB,GAAG,IAAI;IA+EjC;;;;;;OAMG;IACI,0BAA0B,CAAC,WAAW,EAAE,MAAM,EAAE,GAAG,MAAM,GAAG,IAAI;IAUvE;;;;;;OAMG;IACI,6BAA6B,CAAC,WAAW,EAAE,MAAM,EAAE,GAAG,MAAM,GAAG,IAAI;CAS1E"}
+{"version":3,"file":"SignalComputer.d.ts","sourceRoot":"","sources":["../../../src/core/evaluator/SignalComputer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAEH,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,oCAAoC,CAAC;AACtE,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,iBAAiB,CAAC;AACzD,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,eAAe,CAAC;AAEjD,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,wBAAwB,CAAC;AAehE,iEAAiE;AACjE,UAAU,uBAAuB;IAChC,KAAK,EAAE,MAAM,CAAC;IACd,UAAU,EAAE;QACX,cAAc,EAAE,MAAM,CAAC;QACvB,qBAAqB,EAAE,MAAM,CAAC;QAC9B,gBAAgB,EAAE,MAAM,CAAC;QACzB,oBAAoB,EAAE,MAAM,GAAG,IAAI,CAAC;KACpC,CAAC;IACF,cAAc,EAAE;QACf,cAAc,EAAE,MAAM,CAAC;QACvB,qBAAqB,EAAE,MAAM,CAAC;QAC9B,gBAAgB,EAAE,MAAM,CAAC;QACzB,oBAAoB,EAAE,MAAM,GAAG,IAAI,CAAC;KACpC,CAAC;CACF;AAED;;;;;;GAMG;AACH,qBAAa,cAAc;IAC1B;;;;;;;;;;;;;;;;;;;;;;;;;;;OA2BG;IACH,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,eAAe,CAK5B;IAEX;;;OAGG;IACH,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,qBAAqB,CAIlC;IAEX;;;;;;OAMG;IACI,wBAAwB,CAC9B,OAAO,EAAE,WAAW,EAAE,EACtB,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,WAAW,EAAE,CAAC,EACvC,KAAK,EAAE,iBAAiB,GACtB,iBAAiB;IAkCpB;;;;;;;;;;;;;;;OAeG;IACI,wBAAwB,CAC9B,OAAO,EAAE,WAAW,EAAE,EACtB,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,WAAW,EAAE,CAAC,EACvC,gBAAgB,EAAE,MAAM,CAAC,WAAW,EAAE,MAAM,CAAC,EAC7C,WAAW,EAAE,MAAM,EAAE,EACrB,KAAK,EAAE,iBAAiB,GACtB,uBAAuB,GAAG,IAAI;IAsEjC;;;;;;OAMG;IACI,0BAA0B,CAAC,WAAW,EAAE,MAAM,EAAE,GAAG,MAAM,GAAG,IAAI;IAUvE;;;;;;OAMG;IACI,6BAA6B,CAAC,WAAW,EAAE,MAAM,EAAE,GAAG,MAAM,GAAG,IAAI;CAS1E"}

package/dist/core/evaluator/SignalComputer.js

@@ -16,10 +16,10 @@ class SignalComputer {
         verificationCoverage: 0.375,
         depthEfficiency: 0.25
     };
-    computeConfidenceSignals(history, branches) {
+    computeConfidenceSignals(history, branches, links) {
         const typeDistribution = _countByType(history);
         const allConfidences = history.map((t)=>t.confidence).filter((c)=>void 0 !== c);
-        const structuralResult = this.computeStructuralQuality(history, branches, typeDistribution, allConfidences);
+        const structuralResult = this.computeStructuralQuality(history, branches, typeDistribution, allConfidences, links);
         return {
             reasoning_depth: history.length,
             revision_count: history.filter((t)=>t.is_revision).length,
@@ -35,7 +35,7 @@ class SignalComputer {
             }
         };
     }
-    computeStructuralQuality(history, branches, typeDistribution, confidences) {
+    computeStructuralQuality(history, branches, typeDistribution, confidences, links) {
         if (0 === history.length) return null;
         const total = history.length;
         let entropy = 0;
@@ -48,12 +48,7 @@ class SignalComputer {
         }
         const rawTypeDiversity = entropy / Math.log2(6);
         const typeDiversity = Math.max(rawTypeDiversity, FLOOR);
-        const hypotheses = history.filter((t)=>'hypothesis' === t.thought_type);
-        const hypothesisIds = new Set(hypotheses.map((t)=>t.hypothesis_id).filter(Boolean));
-        const verifiedIds = new Set(history.filter((t)=>'verification' === t.thought_type && t.hypothesis_id).map((t)=>t.hypothesis_id));
-        const rawVerificationCoverage = 0 === hypothesisIds.size ? 1.0 : [
-            ...hypothesisIds
-        ].filter((id)=>verifiedIds.has(id)).length / hypothesisIds.size;
+        const rawVerificationCoverage = 0 === links.hypotheses.length ? 1.0 : links.verifiedHypothesisCount / links.hypotheses.length;
         const verificationCoverage = Math.max(rawVerificationCoverage, FLOOR);
         const chainDepth = _computeChainDepth(history);
         const branchCount = Object.keys(branches).length;

package/dist/core/evaluator/SignalComputer.js.map

@@ -1 +1 @@
-{"version":3,"file":"core/evaluator/SignalComputer.js","sources":["../../../src/core/evaluator/SignalComputer.ts"],"sourcesContent":["/**\n * Confidence signal computation for sequential thinking.\n *\n * Provides {@link SignalComputer} — a pure stateless service that computes\n * {@link ConfidenceSignals} (including the composite `structural_quality`\n * score and its components) from thought history and branch data.\n *\n * Extracted from `ThoughtEvaluator` as part of the evaluator decomposition.\n * All methods are pure: no side effects, no I/O, no internal state.\n *\n * @module core/evaluator/SignalComputer\n */\n\nimport type { ThoughtType } from '../../contracts/reasoning-types.js';\nimport type { ConfidenceSignals } from '../reasoning.js';\nimport type { ThoughtData } from '../thought.js';\nimport { ALL_THOUGHT_TYPES, _computeChainDepth, _countByType } from './internals.js';\n\n/** Floor value applied to each quality component to prevent geometric mean collapse. */\nconst FLOOR = 0.01;\n\n/**\n * Round a numeric value to a fixed number of decimal places to mitigate\n * IEEE 754 floating-point accumulation errors (e.g. 0.9 + 0.8 averaging to\n * 0.8500000000000001 instead of 0.85).\n */\nfunction roundToPrecision(value: number, decimals: number = 10): number {\n\tconst factor = Math.pow(10, decimals);\n\treturn Math.round(value * factor) / factor;\n}\n\n/** Result of {@link SignalComputer.computeStructuralQuality}. */\ninterface StructuralQualityResult {\n\tscore: number;\n\tcomponents: {\n\t\ttype_diversity: number;\n\t\tverification_coverage: number;\n\t\tdepth_efficiency: number;\n\t\tconfidence_stability: number | null;\n\t};\n\traw_components: {\n\t\ttype_diversity: number;\n\t\tverification_coverage: number;\n\t\tdepth_efficiency: number;\n\t\tconfidence_stability: number | null;\n\t};\n}\n\n/**\n * Stateless service that computes {@link ConfidenceSignals} from thought history.\n *\n * @remarks\n * Pure computation — no side effects, no I/O, no internal state.\n * Safe to register as singleton or transient in DI.\n */\nexport class SignalComputer {\n\t/**\n\t * Structural quality weight configuration.\n\t *\n\t * Weights reflect the relative importance of each quality dimension:\n\t * - type_diversity (0.3): Shannon entropy of thought types — rewards reasoning that uses\n\t *   multiple cognitive modes (hypothesis, verification, critique, synthesis) rather than\n\t *   monotonic sequences. Based on information theory: higher entropy = more information.\n\t *\n\t * - verification_coverage (0.3): Ratio of verified to total hypotheses — rewards scientific\n\t *   rigor. Hypotheses without verification are speculation; verification grounds reasoning.\n\t *   Equal weight with diversity because unverified reasoning is a common failure mode.\n\t *\n\t * - depth_efficiency (0.2): Ratio of structural depth to total thoughts — rewards efficient\n\t *   reasoning. Deep chains (many revisions/branches) are good; but only if proportionally\n\t *   dense. Padding with low-value thoughts penalizes this metric.\n\t *\n\t * - confidence_stability (0.2): Low variance in self-assessed confidence — rewards calibrated\n\t *   reasoning. Wildly fluctuating confidence suggests uncertainty about the approach.\n\t *   Lower weight because confidence values are LLM self-reports (inherently noisy).\n\t *\n\t * Design rationale:\n\t * - Diversity + Verification (0.6) > Depth + Stability (0.4): Structural properties of the\n\t *   reasoning DAG are more important than behavioral signals.\n\t * - All weights are positive (no dimension is penalizing).\n\t * - Weights sum to 1.0 for normalized scoring.\n\t * - When confidence_stability is unavailable (n<2), remaining weights redistribute\n\t *   proportionally: td→0.375, vc→0.375, de→0.25 (preserving relative ratios).\n\t */\n\tprivate static readonly QUALITY_WEIGHTS = {\n\t\ttypeDiversity: 0.3,\n\t\tverificationCoverage: 0.3,\n\t\tdepthEfficiency: 0.2,\n\t\tconfidenceStability: 0.2,\n\t} as const;\n\n\t/**\n\t * Redistributed weights when confidence_stability is excluded (n<2).\n\t * Preserves relative ratios of remaining components (sum = 1.0).\n\t */\n\tprivate static readonly QUALITY_WEIGHTS_NO_CS = {\n\t\ttypeDiversity: 0.375,\n\t\tverificationCoverage: 0.375,\n\t\tdepthEfficiency: 0.25,\n\t} as const;\n\n\t/**\n\t * Compute confidence signals from history context.\n\t *\n\t * @param history - All thoughts in the current session\n\t * @param branches - Map of branch IDs to their thought arrays\n\t * @returns Computed confidence signals reflecting reasoning quality\n\t */\n\tpublic computeConfidenceSignals(\n\t\thistory: ThoughtData[],\n\t\tbranches: Record<string, ThoughtData[]>\n\t): ConfidenceSignals {\n\t\tconst typeDistribution = _countByType(history);\n\t\tconst allConfidences = history\n\t\t\t.map((t) => t.confidence)\n\t\t\t.filter((c): c is number => c !== undefined);\n\n\t\t// Compute structural quality components\n\t\tconst structuralResult = this.computeStructuralQuality(\n\t\t\thistory,\n\t\t\tbranches,\n\t\t\ttypeDistribution,\n\t\t\tallConfidences\n\t\t);\n\n\t\treturn {\n\t\t\treasoning_depth: history.length,\n\t\t\trevision_count: history.filter((t) => t.is_revision).length,\n\t\t\tbranch_count: Object.keys(branches).length,\n\t\t\tthought_type_distribution: typeDistribution,\n\t\t\thas_hypothesis: history.some((t) => t.thought_type === 'hypothesis'),\n\t\t\thas_verification: history.some((t) => t.thought_type === 'verification'),\n\t\t\taverage_confidence:\n\t\t\t\tallConfidences.length > 0\n\t\t\t\t\t? roundToPrecision(allConfidences.reduce((a, b) => a + b, 0) / allConfidences.length)\n\t\t\t\t\t: null,\n\t\t\t...(structuralResult !== null && {\n\t\t\t\tstructural_quality: structuralResult.score,\n\t\t\t\tquality_components: structuralResult.components,\n\t\t\t\tquality_components_raw: structuralResult.raw_components,\n\t\t\t}),\n\t\t};\n\t}\n\n\t/**\n\t * Compute the composite structural quality score and its components.\n\t *\n\t * Algorithm (all components floored at {@link FLOOR} = 0.01):\n\t * - `type_diversity` = Shannon entropy of thought_type distribution / log2(6)\n\t * - `verification_coverage` = verified_hypotheses / max(total_hypotheses, 1) (1.0 if none)\n\t * - `depth_efficiency` = max(chain_depth, branch_count + 1) / total_thoughts, clamped to 1.0\n\t * - `confidence_stability` = 1 - stddev(confidences), default 0.5 when empty, null when single value\n\t * - `structural_quality` = td^0.3 * vc^0.3 * de^0.2 * cs^0.2 (weighted geometric mean)\n\t *\n\t * Note: When confidence_stability is null (fewer than 2 confidence values), it is\n\t * excluded from the geometric mean and the remaining weights are redistributed\n\t * proportionally (td→0.375, vc→0.375, de→0.25).\n\t *\n\t * @returns Score + components, or `null` when history is empty\n\t */\n\tpublic computeStructuralQuality(\n\t\thistory: ThoughtData[],\n\t\tbranches: Record<string, ThoughtData[]>,\n\t\ttypeDistribution: Record<ThoughtType, number>,\n\t\tconfidences: number[]\n\t): StructuralQualityResult | null {\n\t\tif (history.length === 0) return null;\n\n\t\t// 1. type_diversity: Shannon entropy / log2(6)\n\t\tconst total = history.length;\n\t\tlet entropy = 0;\n\t\tfor (const type of ALL_THOUGHT_TYPES) {\n\t\t\tconst count = typeDistribution[type];\n\t\t\tif (count > 0) {\n\t\t\t\tconst pk = count / total;\n\t\t\t\tentropy -= pk * Math.log2(pk);\n\t\t\t}\n\t\t}\n\t\tconst rawTypeDiversity = entropy / Math.log2(6);\n\t\tconst typeDiversity = Math.max(rawTypeDiversity, FLOOR);\n\n\t\t// 2. verification_coverage: verified / total hypotheses (1.0 if none)\n\t\tconst hypotheses = history.filter((t) => t.thought_type === 'hypothesis');\n\t\tconst hypothesisIds = new Set(hypotheses.map((t) => t.hypothesis_id).filter(Boolean));\n\t\tconst verifiedIds = new Set(\n\t\t\thistory\n\t\t\t\t.filter((t) => t.thought_type === 'verification' && t.hypothesis_id)\n\t\t\t\t.map((t) => t.hypothesis_id)\n\t\t);\n\t\tconst rawVerificationCoverage =\n\t\t\thypothesisIds.size === 0\n\t\t\t\t? 1.0\n\t\t\t\t: [...hypothesisIds].filter((id) => verifiedIds.has(id)).length / hypothesisIds.size;\n\t\tconst verificationCoverage = Math.max(rawVerificationCoverage, FLOOR);\n\n\t\t// 3. depth_efficiency: max(chain_depth, branch_count + 1) / total_thoughts, clamped to 1.0\n\t\t// NOTE (Metis H3): Branching is desirable — treat branches as depth-equivalent.\n\t\tconst chainDepth = _computeChainDepth(history);\n\t\tconst branchCount = Object.keys(branches).length;\n\t\tconst effectiveDepth = Math.max(chainDepth, branchCount + 1);\n\t\tconst rawDepthEfficiency = Math.min(effectiveDepth / total, 1.0);\n\t\tconst depthEfficiency = Math.max(rawDepthEfficiency, FLOOR);\n\n\t\t// 4. confidence_stability: 1 - stddev(confidences), default 0.5\n\t\tconst confidenceStability = this.computeConfidenceStability(confidences);\n\t\tconst rawConfidenceStability = this.computeRawConfidenceStability(confidences);\n\n\t\tconst components: StructuralQualityResult['components'] = {\n\t\t\ttype_diversity: typeDiversity,\n\t\t\tverification_coverage: verificationCoverage,\n\t\t\tdepth_efficiency: depthEfficiency,\n\t\t\tconfidence_stability: confidenceStability,\n\t\t};\n\n\t\tconst rawComponents: StructuralQualityResult['raw_components'] = {\n\t\t\ttype_diversity: rawTypeDiversity,\n\t\t\tverification_coverage: rawVerificationCoverage,\n\t\t\tdepth_efficiency: rawDepthEfficiency,\n\t\t\tconfidence_stability: rawConfidenceStability,\n\t\t};\n\n\t\t// Weighted geometric mean: td^0.3 * vc^0.3 * de^0.2 * cs^0.2\n\t\t// When confidence_stability is null (n<2), exclude it from the geomean\n\t\t// and redistribute its weight proportionally to remaining components\n\t\t// (normalized td=0.375, vc=0.375, de=0.25)\n\t\tlet score: number;\n\t\tif (confidenceStability !== null) {\n\t\t\tconst w = SignalComputer.QUALITY_WEIGHTS;\n\t\t\tscore =\n\t\t\t\tMath.pow(typeDiversity, w.typeDiversity) *\n\t\t\t\tMath.pow(verificationCoverage, w.verificationCoverage) *\n\t\t\t\tMath.pow(depthEfficiency, w.depthEfficiency) *\n\t\t\t\tMath.pow(confidenceStability, w.confidenceStability);\n\t\t} else {\n\t\t\tconst w3 = SignalComputer.QUALITY_WEIGHTS_NO_CS;\n\t\t\tscore =\n\t\t\t\tMath.pow(typeDiversity, w3.typeDiversity) *\n\t\t\t\tMath.pow(verificationCoverage, w3.verificationCoverage) *\n\t\t\t\tMath.pow(depthEfficiency, w3.depthEfficiency);\n\t\t}\n\n\t\treturn { score, components, raw_components: rawComponents };\n\t}\n\n\t/**\n\t * Compute confidence stability: 1 - stddev(confidences).\n\t *\n\t * - Empty input → 0.5 (neutral default)\n\t * - Single value → null (insufficient data, excluded from structural quality)\n\t * - Otherwise → max(1 - stddev, FLOOR)\n\t */\n\tpublic computeConfidenceStability(confidences: number[]): number | null {\n\t\tif (confidences.length === 0) return 0.5;\n\t\tif (confidences.length === 1) return null;\n\n\t\tconst mean = confidences.reduce((a, b) => a + b, 0) / confidences.length;\n\t\tconst variance = confidences.reduce((sum, c) => sum + (c - mean) ** 2, 0) / confidences.length;\n\t\tconst stddev = Math.sqrt(variance);\n\t\treturn Math.max(1 - stddev, FLOOR);\n\t}\n\n\t/**\n\t * Compute raw (unfloored) confidence stability: 1 - stddev(confidences).\n\t *\n\t * - Empty input → 0.5 (neutral default)\n\t * - Single value → null (insufficient data)\n\t * - Otherwise → 1 - stddev (may be below FLOOR)\n\t */\n\tpublic computeRawConfidenceStability(confidences: number[]): number | null {\n\t\tif (confidences.length === 0) return 0.5;\n\t\tif (confidences.length === 1) return null;\n\n\t\tconst mean = confidences.reduce((a, b) => a + b, 0) / confidences.length;\n\t\tconst variance = confidences.reduce((sum, c) => sum + (c - mean) ** 2, 0) / confidences.length;\n\t\tconst stddev = Math.sqrt(variance);\n\t\treturn 1 - stddev;\n\t}\n}\n"],"names":["FLOOR","roundToPrecision","value","decimals","factor","Math","SignalComputer","history","branches","typeDistribution","_countByType","allConfidences","t","c","undefined","structuralResult","Object","a","b","confidences","total","entropy","type","ALL_THOUGHT_TYPES","count","pk","rawTypeDiversity","typeDiversity","hypotheses","hypothesisIds","Set","Boolean","verifiedIds","rawVerificationCoverage","id","verificationCoverage","chainDepth","_computeChainDepth","branchCount","effectiveDepth","rawDepthEfficiency","depthEfficiency","confidenceStability","rawConfidenceStability","components","rawComponents","score","w","w3","mean","variance","sum","stddev"],"mappings":";AAmBA,MAAMA,QAAQ;AAOd,SAASC,iBAAiBC,KAAa,EAAEC,WAAmB,EAAE;IAC7D,MAAMC,SAASC,KAAK,GAAG,CAAC,IAAIF;IAC5B,OAAOE,KAAK,KAAK,CAACH,QAAQE,UAAUA;AACrC;AA0BO,MAAME;IA6BZ,OAAwB,kBAAkB;QACzC,eAAe;QACf,sBAAsB;QACtB,iBAAiB;QACjB,qBAAqB;IACtB,EAAW;IAMX,OAAwB,wBAAwB;QAC/C,eAAe;QACf,sBAAsB;QACtB,iBAAiB;IAClB,EAAW;IASJ,yBACNC,OAAsB,EACtBC,QAAuC,EACnB;QACpB,MAAMC,mBAAmBC,aAAaH;QACtC,MAAMI,iBAAiBJ,QACrB,GAAG,CAAC,CAACK,IAAMA,EAAE,UAAU,EACvB,MAAM,CAAC,CAACC,IAAmBA,AAAMC,WAAND;QAG7B,MAAME,mBAAmB,IAAI,CAAC,wBAAwB,CACrDR,SACAC,UACAC,kBACAE;QAGD,OAAO;YACN,iBAAiBJ,QAAQ,MAAM;YAC/B,gBAAgBA,QAAQ,MAAM,CAAC,CAACK,IAAMA,EAAE,WAAW,EAAE,MAAM;YAC3D,cAAcI,OAAO,IAAI,CAACR,UAAU,MAAM;YAC1C,2BAA2BC;YAC3B,gBAAgBF,QAAQ,IAAI,CAAC,CAACK,IAAMA,AAAmB,iBAAnBA,EAAE,YAAY;YAClD,kBAAkBL,QAAQ,IAAI,CAAC,CAACK,IAAMA,AAAmB,mBAAnBA,EAAE,YAAY;YACpD,oBACCD,eAAe,MAAM,GAAG,IACrBV,iBAAiBU,eAAe,MAAM,CAAC,CAACM,GAAGC,IAAMD,IAAIC,GAAG,KAAKP,eAAe,MAAM,IAClF;YACJ,GAAII,AAAqB,SAArBA,oBAA6B;gBAChC,oBAAoBA,iBAAiB,KAAK;gBAC1C,oBAAoBA,iBAAiB,UAAU;gBAC/C,wBAAwBA,iBAAiB,cAAc;YACxD,CAAC;QACF;IACD;IAkBO,yBACNR,OAAsB,EACtBC,QAAuC,EACvCC,gBAA6C,EAC7CU,WAAqB,EACY;QACjC,IAAIZ,AAAmB,MAAnBA,QAAQ,MAAM,EAAQ,OAAO;QAGjC,MAAMa,QAAQb,QAAQ,MAAM;QAC5B,IAAIc,UAAU;QACd,KAAK,MAAMC,QAAQC,kBAAmB;YACrC,MAAMC,QAAQf,gBAAgB,CAACa,KAAK;YACpC,IAAIE,QAAQ,GAAG;gBACd,MAAMC,KAAKD,QAAQJ;gBACnBC,WAAWI,KAAKpB,KAAK,IAAI,CAACoB;YAC3B;QACD;QACA,MAAMC,mBAAmBL,UAAUhB,KAAK,IAAI,CAAC;QAC7C,MAAMsB,gBAAgBtB,KAAK,GAAG,CAACqB,kBAAkB1B;QAGjD,MAAM4B,aAAarB,QAAQ,MAAM,CAAC,CAACK,IAAMA,AAAmB,iBAAnBA,EAAE,YAAY;QACvD,MAAMiB,gBAAgB,IAAIC,IAAIF,WAAW,GAAG,CAAC,CAAChB,IAAMA,EAAE,aAAa,EAAE,MAAM,CAACmB;QAC5E,MAAMC,cAAc,IAAIF,IACvBvB,QACE,MAAM,CAAC,CAACK,IAAMA,AAAmB,mBAAnBA,EAAE,YAAY,IAAuBA,EAAE,aAAa,EAClE,GAAG,CAAC,CAACA,IAAMA,EAAE,aAAa;QAE7B,MAAMqB,0BACLJ,AAAuB,MAAvBA,cAAc,IAAI,GACf,MACA;eAAIA;SAAc,CAAC,MAAM,CAAC,CAACK,KAAOF,YAAY,GAAG,CAACE,KAAK,MAAM,GAAGL,cAAc,IAAI;QACtF,MAAMM,uBAAuB9B,KAAK,GAAG,CAAC4B,yBAAyBjC;QAI/D,MAAMoC,aAAaC,mBAAmB9B;QACtC,MAAM+B,cAActB,OAAO,IAAI,CAACR,UAAU,MAAM;QAChD,MAAM+B,iBAAiBlC,KAAK,GAAG,CAAC+B,YAAYE,cAAc;QAC1D,MAAME,qBAAqBnC,KAAK,GAAG,CAACkC,iBAAiBnB,OAAO;QAC5D,MAAMqB,kBAAkBpC,KAAK,GAAG,CAACmC,oBAAoBxC;QAGrD,MAAM0C,sBAAsB,IAAI,CAAC,0BAA0B,CAACvB;QAC5D,MAAMwB,yBAAyB,IAAI,CAAC,6BAA6B,CAACxB;QAElE,MAAMyB,aAAoD;YACzD,gBAAgBjB;YAChB,uBAAuBQ;YACvB,kBAAkBM;YAClB,sBAAsBC;QACvB;QAEA,MAAMG,gBAA2D;YAChE,gBAAgBnB;YAChB,uBAAuBO;YACvB,kBAAkBO;YAClB,sBAAsBG;QACvB;QAMA,IAAIG;QACJ,IAAIJ,AAAwB,SAAxBA,qBAA8B;YACjC,MAAMK,IAAIzC,eAAe,eAAe;YACxCwC,QACCzC,KAAK,GAAG,CAACsB,eAAeoB,EAAE,aAAa,IACvC1C,KAAK,GAAG,CAAC8B,sBAAsBY,EAAE,oBAAoB,IACrD1C,KAAK,GAAG,CAACoC,iBAAiBM,EAAE,eAAe,IAC3C1C,KAAK,GAAG,CAACqC,qBAAqBK,EAAE,mBAAmB;QACrD,OAAO;YACN,MAAMC,KAAK1C,eAAe,qBAAqB;YAC/CwC,QACCzC,KAAK,GAAG,CAACsB,eAAeqB,GAAG,aAAa,IACxC3C,KAAK,GAAG,CAAC8B,sBAAsBa,GAAG,oBAAoB,IACtD3C,KAAK,GAAG,CAACoC,iBAAiBO,GAAG,eAAe;QAC9C;QAEA,OAAO;YAAEF;YAAOF;YAAY,gBAAgBC;QAAc;IAC3D;IASO,2BAA2B1B,WAAqB,EAAiB;QACvE,IAAIA,AAAuB,MAAvBA,YAAY,MAAM,EAAQ,OAAO;QACrC,IAAIA,AAAuB,MAAvBA,YAAY,MAAM,EAAQ,OAAO;QAErC,MAAM8B,OAAO9B,YAAY,MAAM,CAAC,CAACF,GAAGC,IAAMD,IAAIC,GAAG,KAAKC,YAAY,MAAM;QACxE,MAAM+B,WAAW/B,YAAY,MAAM,CAAC,CAACgC,KAAKtC,IAAMsC,MAAOtC,AAAAA,CAAAA,IAAIoC,IAAG,KAAM,GAAG,KAAK9B,YAAY,MAAM;QAC9F,MAAMiC,SAAS/C,KAAK,IAAI,CAAC6C;QACzB,OAAO7C,KAAK,GAAG,CAAC,IAAI+C,QAAQpD;IAC7B;IASO,8BAA8BmB,WAAqB,EAAiB;QAC1E,IAAIA,AAAuB,MAAvBA,YAAY,MAAM,EAAQ,OAAO;QACrC,IAAIA,AAAuB,MAAvBA,YAAY,MAAM,EAAQ,OAAO;QAErC,MAAM8B,OAAO9B,YAAY,MAAM,CAAC,CAACF,GAAGC,IAAMD,IAAIC,GAAG,KAAKC,YAAY,MAAM;QACxE,MAAM+B,WAAW/B,YAAY,MAAM,CAAC,CAACgC,KAAKtC,IAAMsC,MAAOtC,AAAAA,CAAAA,IAAIoC,IAAG,KAAM,GAAG,KAAK9B,YAAY,MAAM;QAC9F,MAAMiC,SAAS/C,KAAK,IAAI,CAAC6C;QACzB,OAAO,IAAIE;IACZ;AACD"}
+{"version":3,"file":"core/evaluator/SignalComputer.js","sources":["../../../src/core/evaluator/SignalComputer.ts"],"sourcesContent":["/**\n * Confidence signal computation for sequential thinking.\n *\n * Provides {@link SignalComputer} — a pure stateless service that computes\n * {@link ConfidenceSignals} (including the composite `structural_quality`\n * score and its components) from thought history and branch data.\n *\n * Extracted from `ThoughtEvaluator` as part of the evaluator decomposition.\n * All methods are pure: no side effects, no I/O, no internal state.\n *\n * @module core/evaluator/SignalComputer\n */\n\nimport type { ThoughtType } from '../../contracts/reasoning-types.js';\nimport type { ConfidenceSignals } from '../reasoning.js';\nimport type { ThoughtData } from '../thought.js';\nimport { ALL_THOUGHT_TYPES, _computeChainDepth, _countByType } from './internals.js';\nimport type { VerificationLinks } from './VerificationLinks.js';\n\n/** Floor value applied to each quality component to prevent geometric mean collapse. */\nconst FLOOR = 0.01;\n\n/**\n * Round a numeric value to a fixed number of decimal places to mitigate\n * IEEE 754 floating-point accumulation errors (e.g. 0.9 + 0.8 averaging to\n * 0.8500000000000001 instead of 0.85).\n */\nfunction roundToPrecision(value: number, decimals: number = 10): number {\n\tconst factor = Math.pow(10, decimals);\n\treturn Math.round(value * factor) / factor;\n}\n\n/** Result of {@link SignalComputer.computeStructuralQuality}. */\ninterface StructuralQualityResult {\n\tscore: number;\n\tcomponents: {\n\t\ttype_diversity: number;\n\t\tverification_coverage: number;\n\t\tdepth_efficiency: number;\n\t\tconfidence_stability: number | null;\n\t};\n\traw_components: {\n\t\ttype_diversity: number;\n\t\tverification_coverage: number;\n\t\tdepth_efficiency: number;\n\t\tconfidence_stability: number | null;\n\t};\n}\n\n/**\n * Stateless service that computes {@link ConfidenceSignals} from thought history.\n *\n * @remarks\n * Pure computation — no side effects, no I/O, no internal state.\n * Safe to register as singleton or transient in DI.\n */\nexport class SignalComputer {\n\t/**\n\t * Structural quality weight configuration.\n\t *\n\t * Weights reflect the relative importance of each quality dimension:\n\t * - type_diversity (0.3): Shannon entropy of thought types — rewards reasoning that uses\n\t *   multiple cognitive modes (hypothesis, verification, critique, synthesis) rather than\n\t *   monotonic sequences. Based on information theory: higher entropy = more information.\n\t *\n\t * - verification_coverage (0.3): Ratio of verified to total hypotheses — rewards scientific\n\t *   rigor. Hypotheses without verification are speculation; verification grounds reasoning.\n\t *   Equal weight with diversity because unverified reasoning is a common failure mode.\n\t *\n\t * - depth_efficiency (0.2): Ratio of structural depth to total thoughts — rewards efficient\n\t *   reasoning. Deep chains (many revisions/branches) are good; but only if proportionally\n\t *   dense. Padding with low-value thoughts penalizes this metric.\n\t *\n\t * - confidence_stability (0.2): Low variance in self-assessed confidence — rewards calibrated\n\t *   reasoning. Wildly fluctuating confidence suggests uncertainty about the approach.\n\t *   Lower weight because confidence values are LLM self-reports (inherently noisy).\n\t *\n\t * Design rationale:\n\t * - Diversity + Verification (0.6) > Depth + Stability (0.4): Structural properties of the\n\t *   reasoning DAG are more important than behavioral signals.\n\t * - All weights are positive (no dimension is penalizing).\n\t * - Weights sum to 1.0 for normalized scoring.\n\t * - When confidence_stability is unavailable (n<2), remaining weights redistribute\n\t *   proportionally: td→0.375, vc→0.375, de→0.25 (preserving relative ratios).\n\t */\n\tprivate static readonly QUALITY_WEIGHTS = {\n\t\ttypeDiversity: 0.3,\n\t\tverificationCoverage: 0.3,\n\t\tdepthEfficiency: 0.2,\n\t\tconfidenceStability: 0.2,\n\t} as const;\n\n\t/**\n\t * Redistributed weights when confidence_stability is excluded (n<2).\n\t * Preserves relative ratios of remaining components (sum = 1.0).\n\t */\n\tprivate static readonly QUALITY_WEIGHTS_NO_CS = {\n\t\ttypeDiversity: 0.375,\n\t\tverificationCoverage: 0.375,\n\t\tdepthEfficiency: 0.25,\n\t} as const;\n\n\t/**\n\t * Compute confidence signals from history context.\n\t *\n\t * @param history - All thoughts in the current session\n\t * @param branches - Map of branch IDs to their thought arrays\n\t * @returns Computed confidence signals reflecting reasoning quality\n\t */\n\tpublic computeConfidenceSignals(\n\t\thistory: ThoughtData[],\n\t\tbranches: Record<string, ThoughtData[]>,\n\t\tlinks: VerificationLinks\n\t): ConfidenceSignals {\n\t\tconst typeDistribution = _countByType(history);\n\t\tconst allConfidences = history\n\t\t\t.map((t) => t.confidence)\n\t\t\t.filter((c): c is number => c !== undefined);\n\n\t\t// Compute structural quality components\n\t\tconst structuralResult = this.computeStructuralQuality(\n\t\t\thistory,\n\t\t\tbranches,\n\t\t\ttypeDistribution,\n\t\t\tallConfidences,\n\t\t\tlinks\n\t\t);\n\n\t\treturn {\n\t\t\treasoning_depth: history.length,\n\t\t\trevision_count: history.filter((t) => t.is_revision).length,\n\t\t\tbranch_count: Object.keys(branches).length,\n\t\t\tthought_type_distribution: typeDistribution,\n\t\t\thas_hypothesis: history.some((t) => t.thought_type === 'hypothesis'),\n\t\t\thas_verification: history.some((t) => t.thought_type === 'verification'),\n\t\t\taverage_confidence:\n\t\t\t\tallConfidences.length > 0\n\t\t\t\t\t? roundToPrecision(allConfidences.reduce((a, b) => a + b, 0) / allConfidences.length)\n\t\t\t\t\t: null,\n\t\t\t...(structuralResult !== null && {\n\t\t\t\tstructural_quality: structuralResult.score,\n\t\t\t\tquality_components: structuralResult.components,\n\t\t\t\tquality_components_raw: structuralResult.raw_components,\n\t\t\t}),\n\t\t};\n\t}\n\n\t/**\n\t * Compute the composite structural quality score and its components.\n\t *\n\t * Algorithm (all components floored at {@link FLOOR} = 0.01):\n\t * - `type_diversity` = Shannon entropy of thought_type distribution / log2(6)\n\t * - `verification_coverage` = verified_hypotheses / max(total_hypotheses, 1) (1.0 if none)\n\t * - `depth_efficiency` = max(chain_depth, branch_count + 1) / total_thoughts, clamped to 1.0\n\t * - `confidence_stability` = 1 - stddev(confidences), default 0.5 when empty, null when single value\n\t * - `structural_quality` = td^0.3 * vc^0.3 * de^0.2 * cs^0.2 (weighted geometric mean)\n\t *\n\t * Note: When confidence_stability is null (fewer than 2 confidence values), it is\n\t * excluded from the geometric mean and the remaining weights are redistributed\n\t * proportionally (td→0.375, vc→0.375, de→0.25).\n\t *\n\t * @returns Score + components, or `null` when history is empty\n\t */\n\tpublic computeStructuralQuality(\n\t\thistory: ThoughtData[],\n\t\tbranches: Record<string, ThoughtData[]>,\n\t\ttypeDistribution: Record<ThoughtType, number>,\n\t\tconfidences: number[],\n\t\tlinks: VerificationLinks\n\t): StructuralQualityResult | null {\n\t\tif (history.length === 0) return null;\n\n\t\t// 1. type_diversity: Shannon entropy / log2(6)\n\t\tconst total = history.length;\n\t\tlet entropy = 0;\n\t\tfor (const type of ALL_THOUGHT_TYPES) {\n\t\t\tconst count = typeDistribution[type];\n\t\t\tif (count > 0) {\n\t\t\t\tconst pk = count / total;\n\t\t\t\tentropy -= pk * Math.log2(pk);\n\t\t\t}\n\t\t}\n\t\tconst rawTypeDiversity = entropy / Math.log2(6);\n\t\tconst typeDiversity = Math.max(rawTypeDiversity, FLOOR);\n\n\t\t// 2. verification_coverage: verified / total hypotheses (1.0 if none)\n\t\tconst rawVerificationCoverage =\n\t\t\tlinks.hypotheses.length === 0 ? 1.0 : links.verifiedHypothesisCount / links.hypotheses.length;\n\t\tconst verificationCoverage = Math.max(rawVerificationCoverage, FLOOR);\n\n\t\t// 3. depth_efficiency: max(chain_depth, branch_count + 1) / total_thoughts, clamped to 1.0\n\t\t// NOTE (Metis H3): Branching is desirable — treat branches as depth-equivalent.\n\t\tconst chainDepth = _computeChainDepth(history);\n\t\tconst branchCount = Object.keys(branches).length;\n\t\tconst effectiveDepth = Math.max(chainDepth, branchCount + 1);\n\t\tconst rawDepthEfficiency = Math.min(effectiveDepth / total, 1.0);\n\t\tconst depthEfficiency = Math.max(rawDepthEfficiency, FLOOR);\n\n\t\t// 4. confidence_stability: 1 - stddev(confidences), default 0.5\n\t\tconst confidenceStability = this.computeConfidenceStability(confidences);\n\t\tconst rawConfidenceStability = this.computeRawConfidenceStability(confidences);\n\n\t\tconst components: StructuralQualityResult['components'] = {\n\t\t\ttype_diversity: typeDiversity,\n\t\t\tverification_coverage: verificationCoverage,\n\t\t\tdepth_efficiency: depthEfficiency,\n\t\t\tconfidence_stability: confidenceStability,\n\t\t};\n\n\t\tconst rawComponents: StructuralQualityResult['raw_components'] = {\n\t\t\ttype_diversity: rawTypeDiversity,\n\t\t\tverification_coverage: rawVerificationCoverage,\n\t\t\tdepth_efficiency: rawDepthEfficiency,\n\t\t\tconfidence_stability: rawConfidenceStability,\n\t\t};\n\n\t\t// Weighted geometric mean: td^0.3 * vc^0.3 * de^0.2 * cs^0.2\n\t\t// When confidence_stability is null (n<2), exclude it from the geomean\n\t\t// and redistribute its weight proportionally to remaining components\n\t\t// (normalized td=0.375, vc=0.375, de=0.25)\n\t\tlet score: number;\n\t\tif (confidenceStability !== null) {\n\t\t\tconst w = SignalComputer.QUALITY_WEIGHTS;\n\t\t\tscore =\n\t\t\t\tMath.pow(typeDiversity, w.typeDiversity) *\n\t\t\t\tMath.pow(verificationCoverage, w.verificationCoverage) *\n\t\t\t\tMath.pow(depthEfficiency, w.depthEfficiency) *\n\t\t\t\tMath.pow(confidenceStability, w.confidenceStability);\n\t\t} else {\n\t\t\tconst w3 = SignalComputer.QUALITY_WEIGHTS_NO_CS;\n\t\t\tscore =\n\t\t\t\tMath.pow(typeDiversity, w3.typeDiversity) *\n\t\t\t\tMath.pow(verificationCoverage, w3.verificationCoverage) *\n\t\t\t\tMath.pow(depthEfficiency, w3.depthEfficiency);\n\t\t}\n\n\t\treturn { score, components, raw_components: rawComponents };\n\t}\n\n\t/**\n\t * Compute confidence stability: 1 - stddev(confidences).\n\t *\n\t * - Empty input → 0.5 (neutral default)\n\t * - Single value → null (insufficient data, excluded from structural quality)\n\t * - Otherwise → max(1 - stddev, FLOOR)\n\t */\n\tpublic computeConfidenceStability(confidences: number[]): number | null {\n\t\tif (confidences.length === 0) return 0.5;\n\t\tif (confidences.length === 1) return null;\n\n\t\tconst mean = confidences.reduce((a, b) => a + b, 0) / confidences.length;\n\t\tconst variance = confidences.reduce((sum, c) => sum + (c - mean) ** 2, 0) / confidences.length;\n\t\tconst stddev = Math.sqrt(variance);\n\t\treturn Math.max(1 - stddev, FLOOR);\n\t}\n\n\t/**\n\t * Compute raw (unfloored) confidence stability: 1 - stddev(confidences).\n\t *\n\t * - Empty input → 0.5 (neutral default)\n\t * - Single value → null (insufficient data)\n\t * - Otherwise → 1 - stddev (may be below FLOOR)\n\t */\n\tpublic computeRawConfidenceStability(confidences: number[]): number | null {\n\t\tif (confidences.length === 0) return 0.5;\n\t\tif (confidences.length === 1) return null;\n\n\t\tconst mean = confidences.reduce((a, b) => a + b, 0) / confidences.length;\n\t\tconst variance = confidences.reduce((sum, c) => sum + (c - mean) ** 2, 0) / confidences.length;\n\t\tconst stddev = Math.sqrt(variance);\n\t\treturn 1 - stddev;\n\t}\n}\n"],"names":["FLOOR","roundToPrecision","value","decimals","factor","Math","SignalComputer","history","branches","links","typeDistribution","_countByType","allConfidences","t","c","undefined","structuralResult","Object","a","b","confidences","total","entropy","type","ALL_THOUGHT_TYPES","count","pk","rawTypeDiversity","typeDiversity","rawVerificationCoverage","verificationCoverage","chainDepth","_computeChainDepth","branchCount","effectiveDepth","rawDepthEfficiency","depthEfficiency","confidenceStability","rawConfidenceStability","components","rawComponents","score","w","w3","mean","variance","sum","stddev"],"mappings":";AAoBA,MAAMA,QAAQ;AAOd,SAASC,iBAAiBC,KAAa,EAAEC,WAAmB,EAAE;IAC7D,MAAMC,SAASC,KAAK,GAAG,CAAC,IAAIF;IAC5B,OAAOE,KAAK,KAAK,CAACH,QAAQE,UAAUA;AACrC;AA0BO,MAAME;IA6BZ,OAAwB,kBAAkB;QACzC,eAAe;QACf,sBAAsB;QACtB,iBAAiB;QACjB,qBAAqB;IACtB,EAAW;IAMX,OAAwB,wBAAwB;QAC/C,eAAe;QACf,sBAAsB;QACtB,iBAAiB;IAClB,EAAW;IASJ,yBACNC,OAAsB,EACtBC,QAAuC,EACvCC,KAAwB,EACJ;QACpB,MAAMC,mBAAmBC,aAAaJ;QACtC,MAAMK,iBAAiBL,QACrB,GAAG,CAAC,CAACM,IAAMA,EAAE,UAAU,EACvB,MAAM,CAAC,CAACC,IAAmBA,AAAMC,WAAND;QAG7B,MAAME,mBAAmB,IAAI,CAAC,wBAAwB,CACrDT,SACAC,UACAE,kBACAE,gBACAH;QAGD,OAAO;YACN,iBAAiBF,QAAQ,MAAM;YAC/B,gBAAgBA,QAAQ,MAAM,CAAC,CAACM,IAAMA,EAAE,WAAW,EAAE,MAAM;YAC3D,cAAcI,OAAO,IAAI,CAACT,UAAU,MAAM;YAC1C,2BAA2BE;YAC3B,gBAAgBH,QAAQ,IAAI,CAAC,CAACM,IAAMA,AAAmB,iBAAnBA,EAAE,YAAY;YAClD,kBAAkBN,QAAQ,IAAI,CAAC,CAACM,IAAMA,AAAmB,mBAAnBA,EAAE,YAAY;YACpD,oBACCD,eAAe,MAAM,GAAG,IACrBX,iBAAiBW,eAAe,MAAM,CAAC,CAACM,GAAGC,IAAMD,IAAIC,GAAG,KAAKP,eAAe,MAAM,IAClF;YACJ,GAAII,AAAqB,SAArBA,oBAA6B;gBAChC,oBAAoBA,iBAAiB,KAAK;gBAC1C,oBAAoBA,iBAAiB,UAAU;gBAC/C,wBAAwBA,iBAAiB,cAAc;YACxD,CAAC;QACF;IACD;IAkBO,yBACNT,OAAsB,EACtBC,QAAuC,EACvCE,gBAA6C,EAC7CU,WAAqB,EACrBX,KAAwB,EACS;QACjC,IAAIF,AAAmB,MAAnBA,QAAQ,MAAM,EAAQ,OAAO;QAGjC,MAAMc,QAAQd,QAAQ,MAAM;QAC5B,IAAIe,UAAU;QACd,KAAK,MAAMC,QAAQC,kBAAmB;YACrC,MAAMC,QAAQf,gBAAgB,CAACa,KAAK;YACpC,IAAIE,QAAQ,GAAG;gBACd,MAAMC,KAAKD,QAAQJ;gBACnBC,WAAWI,KAAKrB,KAAK,IAAI,CAACqB;YAC3B;QACD;QACA,MAAMC,mBAAmBL,UAAUjB,KAAK,IAAI,CAAC;QAC7C,MAAMuB,gBAAgBvB,KAAK,GAAG,CAACsB,kBAAkB3B;QAGjD,MAAM6B,0BACLpB,AAA4B,MAA5BA,MAAM,UAAU,CAAC,MAAM,GAAS,MAAMA,MAAM,uBAAuB,GAAGA,MAAM,UAAU,CAAC,MAAM;QAC9F,MAAMqB,uBAAuBzB,KAAK,GAAG,CAACwB,yBAAyB7B;QAI/D,MAAM+B,aAAaC,mBAAmBzB;QACtC,MAAM0B,cAAchB,OAAO,IAAI,CAACT,UAAU,MAAM;QAChD,MAAM0B,iBAAiB7B,KAAK,GAAG,CAAC0B,YAAYE,cAAc;QAC1D,MAAME,qBAAqB9B,KAAK,GAAG,CAAC6B,iBAAiBb,OAAO;QAC5D,MAAMe,kBAAkB/B,KAAK,GAAG,CAAC8B,oBAAoBnC;QAGrD,MAAMqC,sBAAsB,IAAI,CAAC,0BAA0B,CAACjB;QAC5D,MAAMkB,yBAAyB,IAAI,CAAC,6BAA6B,CAAClB;QAElE,MAAMmB,aAAoD;YACzD,gBAAgBX;YAChB,uBAAuBE;YACvB,kBAAkBM;YAClB,sBAAsBC;QACvB;QAEA,MAAMG,gBAA2D;YAChE,gBAAgBb;YAChB,uBAAuBE;YACvB,kBAAkBM;YAClB,sBAAsBG;QACvB;QAMA,IAAIG;QACJ,IAAIJ,AAAwB,SAAxBA,qBAA8B;YACjC,MAAMK,IAAIpC,eAAe,eAAe;YACxCmC,QACCpC,KAAK,GAAG,CAACuB,eAAec,EAAE,aAAa,IACvCrC,KAAK,GAAG,CAACyB,sBAAsBY,EAAE,oBAAoB,IACrDrC,KAAK,GAAG,CAAC+B,iBAAiBM,EAAE,eAAe,IAC3CrC,KAAK,GAAG,CAACgC,qBAAqBK,EAAE,mBAAmB;QACrD,OAAO;YACN,MAAMC,KAAKrC,eAAe,qBAAqB;YAC/CmC,QACCpC,KAAK,GAAG,CAACuB,eAAee,GAAG,aAAa,IACxCtC,KAAK,GAAG,CAACyB,sBAAsBa,GAAG,oBAAoB,IACtDtC,KAAK,GAAG,CAAC+B,iBAAiBO,GAAG,eAAe;QAC9C;QAEA,OAAO;YAAEF;YAAOF;YAAY,gBAAgBC;QAAc;IAC3D;IASO,2BAA2BpB,WAAqB,EAAiB;QACvE,IAAIA,AAAuB,MAAvBA,YAAY,MAAM,EAAQ,OAAO;QACrC,IAAIA,AAAuB,MAAvBA,YAAY,MAAM,EAAQ,OAAO;QAErC,MAAMwB,OAAOxB,YAAY,MAAM,CAAC,CAACF,GAAGC,IAAMD,IAAIC,GAAG,KAAKC,YAAY,MAAM;QACxE,MAAMyB,WAAWzB,YAAY,MAAM,CAAC,CAAC0B,KAAKhC,IAAMgC,MAAOhC,AAAAA,CAAAA,IAAI8B,IAAG,KAAM,GAAG,KAAKxB,YAAY,MAAM;QAC9F,MAAM2B,SAAS1C,KAAK,IAAI,CAACwC;QACzB,OAAOxC,KAAK,GAAG,CAAC,IAAI0C,QAAQ/C;IAC7B;IASO,8BAA8BoB,WAAqB,EAAiB;QAC1E,IAAIA,AAAuB,MAAvBA,YAAY,MAAM,EAAQ,OAAO;QACrC,IAAIA,AAAuB,MAAvBA,YAAY,MAAM,EAAQ,OAAO;QAErC,MAAMwB,OAAOxB,YAAY,MAAM,CAAC,CAACF,GAAGC,IAAMD,IAAIC,GAAG,KAAKC,YAAY,MAAM;QACxE,MAAMyB,WAAWzB,YAAY,MAAM,CAAC,CAAC0B,KAAKhC,IAAMgC,MAAOhC,AAAAA,CAAAA,IAAI8B,IAAG,KAAM,GAAG,KAAKxB,YAAY,MAAM;QAC9F,MAAM2B,SAAS1C,KAAK,IAAI,CAACwC;QACzB,OAAO,IAAIE;IACZ;AACD"}

package/dist/core/evaluator/VerificationLinks.d.ts

@@ -0,0 +1,19 @@
+import type { ThoughtId } from '../../contracts/ids.js';
+import type { ResolvedThoughtReferences } from '../IHistoryManager.js';
+import type { ThoughtData } from '../thought.js';
+export type ActiveThoughtBranches = Readonly<Record<string, readonly ThoughtData[]>>;
+export interface VerificationLinks {
+    readonly thoughts: readonly ThoughtData[];
+    readonly hypotheses: readonly ThoughtData[];
+    readonly hypothesisIds: ReadonlySet<ThoughtId>;
+    readonly verifiedHypothesisIds: ReadonlySet<ThoughtId>;
+    readonly verifiedHypothesisCount: number;
+    targetFor(verifierId: ThoughtId): ThoughtId | undefined;
+    isVerified(hypothesis: ThoughtData): boolean;
+    verifies(verifier: ThoughtData, hypothesis: ThoughtData): boolean;
+}
+export declare function resolvedVerificationTarget(thought: ThoughtData, references: ResolvedThoughtReferences): ThoughtId | undefined;
+export declare function resolveVerificationLinks(history: readonly ThoughtData[], branches: ActiveThoughtBranches, admissionTargets?: ReadonlyMap<ThoughtId, ThoughtId>): VerificationLinks;
+export declare function collectActiveMainThoughts(history: readonly ThoughtData[], branches: ActiveThoughtBranches): ThoughtData[];
+export declare function collectActiveThoughts(history: readonly ThoughtData[], branches: ActiveThoughtBranches): ThoughtData[];
+//# sourceMappingURL=VerificationLinks.d.ts.map

package/dist/core/evaluator/VerificationLinks.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"VerificationLinks.d.ts","sourceRoot":"","sources":["../../../src/core/evaluator/VerificationLinks.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,wBAAwB,CAAC;AACxD,OAAO,KAAK,EAAE,yBAAyB,EAAE,MAAM,uBAAuB,CAAC;AACvE,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,eAAe,CAAC;AAEjD,MAAM,MAAM,qBAAqB,GAAG,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,SAAS,WAAW,EAAE,CAAC,CAAC,CAAC;AAErF,MAAM,WAAW,iBAAiB;IACjC,QAAQ,CAAC,QAAQ,EAAE,SAAS,WAAW,EAAE,CAAC;IAC1C,QAAQ,CAAC,UAAU,EAAE,SAAS,WAAW,EAAE,CAAC;IAC5C,QAAQ,CAAC,aAAa,EAAE,WAAW,CAAC,SAAS,CAAC,CAAC;IAC/C,QAAQ,CAAC,qBAAqB,EAAE,WAAW,CAAC,SAAS,CAAC,CAAC;IACvD,QAAQ,CAAC,uBAAuB,EAAE,MAAM,CAAC;IACzC,SAAS,CAAC,UAAU,EAAE,SAAS,GAAG,SAAS,GAAG,SAAS,CAAC;IACxD,UAAU,CAAC,UAAU,EAAE,WAAW,GAAG,OAAO,CAAC;IAC7C,QAAQ,CAAC,QAAQ,EAAE,WAAW,EAAE,UAAU,EAAE,WAAW,GAAG,OAAO,CAAC;CAClE;AAED,wBAAgB,0BAA0B,CACzC,OAAO,EAAE,WAAW,EACpB,UAAU,EAAE,yBAAyB,GACnC,SAAS,GAAG,SAAS,CAIvB;AAED,wBAAgB,wBAAwB,CACvC,OAAO,EAAE,SAAS,WAAW,EAAE,EAC/B,QAAQ,EAAE,qBAAqB,EAC/B,gBAAgB,GAAE,WAAW,CAAC,SAAS,EAAE,SAAS,CAAa,GAC7D,iBAAiB,CA0CnB;AAED,wBAAgB,yBAAyB,CACxC,OAAO,EAAE,SAAS,WAAW,EAAE,EAC/B,QAAQ,EAAE,qBAAqB,GAC7B,WAAW,EAAE,CAmBf;AAED,wBAAgB,qBAAqB,CACpC,OAAO,EAAE,SAAS,WAAW,EAAE,EAC/B,QAAQ,EAAE,qBAAqB,GAC7B,WAAW,EAAE,CAef"}