open-multi-agent-kit 0.96.2 → 0.98.0

package/dist/core/reasoning-router-promotion.js

@@ -0,0 +1,121 @@
+/**
+ * Promotion gate for learned reasoning-router weights.
+ *
+ * The router already produces a feedback ledger (`router-feedback-collector.ts`)
+ * and an offline calibration stack (`scripts/reasoning-router/`: calibration,
+ * held-out evaluation, golden diffing, and McNemar's exact test). What was
+ * missing is the decision policy that makes automatic promotion trustworthy:
+ * without it the loop is open and every weight change is a judgement call.
+ *
+ * This module is that policy, as a pure total function. It deliberately refuses
+ * to promote on anything short of unanimous evidence:
+ *
+ * - outcomes that survived a two-run replay, so a flipped measurement is never
+ *   credited as a win (see `reasoning-router-replay-stability.ts`);
+ * - a comparison against the frozen reference policy rather than an opponent
+ *   the caller picked, which is what stops baseline shopping;
+ * - a statistically significant win on the held-in gold set (McNemar's exact
+ *   test), computed by the caller and passed in;
+ * - enough discordant pairs for that test to mean anything;
+ * - the candidate — not the baseline — winning those pairs;
+ * - no accuracy regression on a held-out split the calibration never saw;
+ * - no unreviewed behavior changes surfaced by golden diffing; and
+ * - an explicit human approval.
+ *
+ * "The agent tried again" is not the same as "the harness improved", so a
+ * candidate that merely fails to be worse is rejected, and malformed evidence
+ * fails closed rather than defaulting to promotion.
+ */
+/**
+ * Conservative defaults. `minDiscordant` of 20 keeps McNemar from ruling on a
+ * handful of pairs; `maxGoldenChanges` of 0 means a reviewer accepts behavior
+ * changes by regenerating the golden set, never by tolerating drift.
+ *
+ * `minHoldout` is 40 because the router gold set freezes 42 holdout rows (6 per
+ * class × 7 classes). A higher bar would be unsatisfiable rather than strict,
+ * and a gate that can never pass is a gate nobody runs.
+ *
+ * `minReplays` of 2 is the two-run rule: one observation cannot be checked for
+ * stability at all. `maxUnstable` of 0 follows from what instability means here
+ * — the routing path is deterministic, so a single disagreeing row is a defect
+ * in the measurement, not an acceptable error rate to average over.
+ */
+export const DEFAULT_ROUTER_PROMOTION_POLICY = {
+    maxGoldenChanges: 0,
+    maxUnstable: 0,
+    minDiscordant: 20,
+    minHoldout: 40,
+    minReplays: 2,
+    requireFrozenBaseline: true,
+    requireHumanApproval: true,
+};
+function isCount(value) {
+    return Number.isFinite(value) && value >= 0;
+}
+function isMalformed(evidence) {
+    const { goldenChanges, heldIn, holdout, stability } = evidence;
+    if (!isCount(heldIn.baselineWins) || !isCount(heldIn.candidateWins))
+        return true;
+    if (!Number.isFinite(heldIn.pValue) || heldIn.pValue < 0 || heldIn.pValue > 1)
+        return true;
+    if (!isCount(holdout.baselineCorrect) || !isCount(holdout.candidateCorrect) || !isCount(holdout.total))
+        return true;
+    if (holdout.baselineCorrect > holdout.total || holdout.candidateCorrect > holdout.total)
+        return true;
+    if (!isCount(stability.evaluated) || !isCount(stability.unstable) || !isCount(stability.replays))
+        return true;
+    if (stability.unstable > stability.evaluated)
+        return true;
+    return !isCount(goldenChanges);
+}
+/**
+ * Decide whether a candidate weight set may replace the active one.
+ *
+ * Returns every independent blocker rather than the first, so one calibration
+ * run reports the full distance to promotion instead of revealing it one
+ * rejection at a time. Malformed evidence short-circuits: the gate never
+ * reasons about numbers it cannot trust.
+ *
+ * Evidence-integrity blockers are reported ahead of statistical ones. A p-value
+ * computed over unstable rows or against the wrong opponent is not a weaker
+ * result, it is a result about nothing, and the report should say so first.
+ */
+export function evaluateRouterPromotion(evidence, policy = DEFAULT_ROUTER_PROMOTION_POLICY) {
+    if (isMalformed(evidence)) {
+        return { blockers: ["malformed_evidence"], promote: false };
+    }
+    const { baselineKind, goldenChanges, heldIn, holdout, humanApproved, stability } = evidence;
+    const blockers = [];
+    if (stability.replays < policy.minReplays) {
+        blockers.push("insufficient_replays");
+    }
+    if (stability.unstable > policy.maxUnstable) {
+        blockers.push("unstable_evidence");
+    }
+    if (policy.requireFrozenBaseline && baselineKind !== "frozen_reference") {
+        blockers.push("baseline_not_frozen");
+    }
+    if (heldIn.baselineWins + heldIn.candidateWins < policy.minDiscordant) {
+        blockers.push("insufficient_discordant_pairs");
+    }
+    if (!heldIn.significant) {
+        blockers.push("mcnemar_not_significant");
+    }
+    if (heldIn.candidateWins <= heldIn.baselineWins) {
+        blockers.push("candidate_not_favored");
+    }
+    if (holdout.total < policy.minHoldout) {
+        blockers.push("insufficient_holdout");
+    }
+    if (holdout.candidateCorrect < holdout.baselineCorrect) {
+        blockers.push("holdout_regression");
+    }
+    if (goldenChanges > policy.maxGoldenChanges) {
+        blockers.push("unreviewed_behavior_changes");
+    }
+    if (policy.requireHumanApproval && !humanApproved) {
+        blockers.push("human_approval_missing");
+    }
+    return { blockers, promote: blockers.length === 0 };
+}
+//# sourceMappingURL=reasoning-router-promotion.js.map

package/dist/core/reasoning-router-promotion.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"reasoning-router-promotion.js","sourceRoot":"","sources":["../../src/core/reasoning-router-promotion.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2BG;AAwFH;;;;;;;;;;;;;GAaG;AACH,MAAM,CAAC,MAAM,+BAA+B,GAA0B;IACrE,gBAAgB,EAAE,CAAC;IACnB,WAAW,EAAE,CAAC;IACd,aAAa,EAAE,EAAE;IACjB,UAAU,EAAE,EAAE;IACd,UAAU,EAAE,CAAC;IACb,qBAAqB,EAAE,IAAI;IAC3B,oBAAoB,EAAE,IAAI;CAC1B,CAAC;AAEF,SAAS,OAAO,CAAC,KAAa,EAAW;IACxC,OAAO,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,IAAI,CAAC,CAAC;AAAA,CAC5C;AAED,SAAS,WAAW,CAAC,QAAiC,EAAW;IAChE,MAAM,EAAE,aAAa,EAAE,MAAM,EAAE,OAAO,EAAE,SAAS,EAAE,GAAG,QAAQ,CAAC;IAC/D,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,YAAY,CAAC,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,aAAa,CAAC;QAAE,OAAO,IAAI,CAAC;IACjF,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,MAAM,CAAC,IAAI,MAAM,CAAC,MAAM,GAAG,CAAC,IAAI,MAAM,CAAC,MAAM,GAAG,CAAC;QAAE,OAAO,IAAI,CAAC;IAC3F,IAAI,CAAC,OAAO,CAAC,OAAO,CAAC,eAAe,CAAC,IAAI,CAAC,OAAO,CAAC,OAAO,CAAC,gBAAgB,CAAC,IAAI,CAAC,OAAO,CAAC,OAAO,CAAC,KAAK,CAAC;QAAE,OAAO,IAAI,CAAC;IACpH,IAAI,OAAO,CAAC,eAAe,GAAG,OAAO,CAAC,KAAK,IAAI,OAAO,CAAC,gBAAgB,GAAG,OAAO,CAAC,KAAK;QAAE,OAAO,IAAI,CAAC;IACrG,IAAI,CAAC,OAAO,CAAC,SAAS,CAAC,SAAS,CAAC,IAAI,CAAC,OAAO,CAAC,SAAS,CAAC,QAAQ,CAAC,IAAI,CAAC,OAAO,CAAC,SAAS,CAAC,OAAO,CAAC;QAAE,OAAO,IAAI,CAAC;IAC9G,IAAI,SAAS,CAAC,QAAQ,GAAG,SAAS,CAAC,SAAS;QAAE,OAAO,IAAI,CAAC;IAC1D,OAAO,CAAC,OAAO,CAAC,aAAa,CAAC,CAAC;AAAA,CAC/B;AAED;;;;;;;;;;;GAWG;AACH,MAAM,UAAU,uBAAuB,CACtC,QAAiC,EACjC,MAAM,GAA0B,+BAA+B,EACtC;IACzB,IAAI,WAAW,CAAC,QAAQ,CAAC,EAAE,CAAC;QAC3B,OAAO,EAAE,QAAQ,EAAE,CAAC,oBAAoB,CAAC,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;IAC7D,CAAC;IAED,MAAM,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,EAAE,OAAO,EAAE,aAAa,EAAE,SAAS,EAAE,GAAG,QAAQ,CAAC;IAC5F,MAAM,QAAQ,GAA6B,EAAE,CAAC;IAE9C,IAAI,SAAS,CAAC,OAAO,GAAG,MAAM,CAAC,UAAU,EAAE,CAAC;QAC3C,QAAQ,CAAC,IAAI,CAAC,sBAAsB,CAAC,CAAC;IACvC,CAAC;IACD,IAAI,SAAS,CAAC,QAAQ,GAAG,MAAM,CAAC,WAAW,EAAE,CAAC;QAC7C,QAAQ,CAAC,IAAI,CAAC,mBAAmB,CAAC,CAAC;IACpC,CAAC;IACD,IAAI,MAAM,CAAC,qBAAqB,IAAI,YAAY,KAAK,kBAAkB,EAAE,CAAC;QACzE,QAAQ,CAAC,IAAI,CAAC,qBAAqB,CAAC,CAAC;IACtC,CAAC;IACD,IAAI,MAAM,CAAC,YAAY,GAAG,MAAM,CAAC,aAAa,GAAG,MAAM,CAAC,aAAa,EAAE,CAAC;QACvE,QAAQ,CAAC,IAAI,CAAC,+BAA+B,CAAC,CAAC;IAChD,CAAC;IACD,IAAI,CAAC,MAAM,CAAC,WAAW,EAAE,CAAC;QACzB,QAAQ,CAAC,IAAI,CAAC,yBAAyB,CAAC,CAAC;IAC1C,CAAC;IACD,IAAI,MAAM,CAAC,aAAa,IAAI,MAAM,CAAC,YAAY,EAAE,CAAC;QACjD,QAAQ,CAAC,IAAI,CAAC,uBAAuB,CAAC,CAAC;IACxC,CAAC;IACD,IAAI,OAAO,CAAC,KAAK,GAAG,MAAM,CAAC,UAAU,EAAE,CAAC;QACvC,QAAQ,CAAC,IAAI,CAAC,sBAAsB,CAAC,CAAC;IACvC,CAAC;IACD,IAAI,OAAO,CAAC,gBAAgB,GAAG,OAAO,CAAC,eAAe,EAAE,CAAC;QACxD,QAAQ,CAAC,IAAI,CAAC,oBAAoB,CAAC,CAAC;IACrC,CAAC;IACD,IAAI,aAAa,GAAG,MAAM,CAAC,gBAAgB,EAAE,CAAC;QAC7C,QAAQ,CAAC,IAAI,CAAC,6BAA6B,CAAC,CAAC;IAC9C,CAAC;IACD,IAAI,MAAM,CAAC,oBAAoB,IAAI,CAAC,aAAa,EAAE,CAAC;QACnD,QAAQ,CAAC,IAAI,CAAC,wBAAwB,CAAC,CAAC;IACzC,CAAC;IAED,OAAO,EAAE,QAAQ,EAAE,OAAO,EAAE,QAAQ,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;AAAA,CACpD","sourcesContent":["/**\n * Promotion gate for learned reasoning-router weights.\n *\n * The router already produces a feedback ledger (`router-feedback-collector.ts`)\n * and an offline calibration stack (`scripts/reasoning-router/`: calibration,\n * held-out evaluation, golden diffing, and McNemar's exact test). What was\n * missing is the decision policy that makes automatic promotion trustworthy:\n * without it the loop is open and every weight change is a judgement call.\n *\n * This module is that policy, as a pure total function. It deliberately refuses\n * to promote on anything short of unanimous evidence:\n *\n * - outcomes that survived a two-run replay, so a flipped measurement is never\n *   credited as a win (see `reasoning-router-replay-stability.ts`);\n * - a comparison against the frozen reference policy rather than an opponent\n *   the caller picked, which is what stops baseline shopping;\n * - a statistically significant win on the held-in gold set (McNemar's exact\n *   test), computed by the caller and passed in;\n * - enough discordant pairs for that test to mean anything;\n * - the candidate — not the baseline — winning those pairs;\n * - no accuracy regression on a held-out split the calibration never saw;\n * - no unreviewed behavior changes surfaced by golden diffing; and\n * - an explicit human approval.\n *\n * \"The agent tried again\" is not the same as \"the harness improved\", so a\n * candidate that merely fails to be worse is rejected, and malformed evidence\n * fails closed rather than defaulting to promotion.\n */\n\n/** Held-in McNemar outcome. Wins are counted in discordant pairs only. */\nexport interface RouterHeldInEvidence {\n\t/** Rows the baseline classified correctly and the candidate got wrong. */\n\treadonly baselineWins: number;\n\t/** Rows the candidate classified correctly and the baseline got wrong. */\n\treadonly candidateWins: number;\n\treadonly pValue: number;\n\treadonly significant: boolean;\n}\n\n/** Accuracy on a split withheld from calibration. */\nexport interface RouterHoldoutEvidence {\n\treadonly baselineCorrect: number;\n\treadonly candidateCorrect: number;\n\treadonly total: number;\n}\n\n/** How much of the evidence survived repeated identical replays. */\nexport interface RouterStabilityEvidence {\n\t/** Rows replayed under both policies, whatever bucket they landed in. */\n\treadonly evaluated: number;\n\t/** Rows withheld because repeated observations disagreed. */\n\treadonly unstable: number;\n\t/** Independent observations per row behind the credited outcomes. */\n\treadonly replays: number;\n}\n\n/**\n * What the candidate was measured against.\n *\n * `frozen_reference` is the shipped policy that learning never updates. A\n * candidate that only beats an opponent the caller chose has demonstrated\n * nothing, so the gate refuses to reason about `ad_hoc` comparisons.\n */\nexport type RouterBaselineKind = \"frozen_reference\" | \"ad_hoc\";\n\n/** Everything the gate is allowed to consider. */\nexport interface RouterPromotionEvidence {\n\treadonly heldIn: RouterHeldInEvidence;\n\treadonly holdout: RouterHoldoutEvidence;\n\t/** Replay agreement behind every credited outcome. */\n\treadonly stability: RouterStabilityEvidence;\n\t/** Which opponent produced `heldIn` and `holdout`. */\n\treadonly baselineKind: RouterBaselineKind;\n\t/** Golden-diff changes still unreviewed at decision time. */\n\treadonly goldenChanges: number;\n\t/** True only when a human explicitly approved this exact candidate. */\n\treadonly humanApproved: boolean;\n}\n\n/** Tunable thresholds. Loosening these can never authorize a regression. */\nexport interface RouterPromotionPolicy {\n\t/** Minimum discordant pairs before McNemar's verdict is trusted. */\n\treadonly minDiscordant: number;\n\t/** Minimum held-out rows before \"no regression\" is meaningful. */\n\treadonly minHoldout: number;\n\t/** Maximum unreviewed golden-diff changes tolerated. */\n\treadonly maxGoldenChanges: number;\n\t/** Independent replays required per row before any outcome carries credit. */\n\treadonly minReplays: number;\n\t/** Maximum rows allowed to disagree across replays. */\n\treadonly maxUnstable: number;\n\t/** Require the frozen reference policy as the comparison opponent. */\n\treadonly requireFrozenBaseline: boolean;\n\treadonly requireHumanApproval: boolean;\n}\n\nexport type RouterPromotionBlocker =\n\t| \"malformed_evidence\"\n\t| \"insufficient_replays\"\n\t| \"unstable_evidence\"\n\t| \"baseline_not_frozen\"\n\t| \"insufficient_discordant_pairs\"\n\t| \"mcnemar_not_significant\"\n\t| \"candidate_not_favored\"\n\t| \"holdout_regression\"\n\t| \"insufficient_holdout\"\n\t| \"unreviewed_behavior_changes\"\n\t| \"human_approval_missing\";\n\nexport interface RouterPromotionVerdict {\n\treadonly promote: boolean;\n\t/** Every independent reason promotion was refused, in evaluation order. */\n\treadonly blockers: readonly RouterPromotionBlocker[];\n}\n\n/**\n * Conservative defaults. `minDiscordant` of 20 keeps McNemar from ruling on a\n * handful of pairs; `maxGoldenChanges` of 0 means a reviewer accepts behavior\n * changes by regenerating the golden set, never by tolerating drift.\n *\n * `minHoldout` is 40 because the router gold set freezes 42 holdout rows (6 per\n * class × 7 classes). A higher bar would be unsatisfiable rather than strict,\n * and a gate that can never pass is a gate nobody runs.\n *\n * `minReplays` of 2 is the two-run rule: one observation cannot be checked for\n * stability at all. `maxUnstable` of 0 follows from what instability means here\n * — the routing path is deterministic, so a single disagreeing row is a defect\n * in the measurement, not an acceptable error rate to average over.\n */\nexport const DEFAULT_ROUTER_PROMOTION_POLICY: RouterPromotionPolicy = {\n\tmaxGoldenChanges: 0,\n\tmaxUnstable: 0,\n\tminDiscordant: 20,\n\tminHoldout: 40,\n\tminReplays: 2,\n\trequireFrozenBaseline: true,\n\trequireHumanApproval: true,\n};\n\nfunction isCount(value: number): boolean {\n\treturn Number.isFinite(value) && value >= 0;\n}\n\nfunction isMalformed(evidence: RouterPromotionEvidence): boolean {\n\tconst { goldenChanges, heldIn, holdout, stability } = evidence;\n\tif (!isCount(heldIn.baselineWins) || !isCount(heldIn.candidateWins)) return true;\n\tif (!Number.isFinite(heldIn.pValue) || heldIn.pValue < 0 || heldIn.pValue > 1) return true;\n\tif (!isCount(holdout.baselineCorrect) || !isCount(holdout.candidateCorrect) || !isCount(holdout.total)) return true;\n\tif (holdout.baselineCorrect > holdout.total || holdout.candidateCorrect > holdout.total) return true;\n\tif (!isCount(stability.evaluated) || !isCount(stability.unstable) || !isCount(stability.replays)) return true;\n\tif (stability.unstable > stability.evaluated) return true;\n\treturn !isCount(goldenChanges);\n}\n\n/**\n * Decide whether a candidate weight set may replace the active one.\n *\n * Returns every independent blocker rather than the first, so one calibration\n * run reports the full distance to promotion instead of revealing it one\n * rejection at a time. Malformed evidence short-circuits: the gate never\n * reasons about numbers it cannot trust.\n *\n * Evidence-integrity blockers are reported ahead of statistical ones. A p-value\n * computed over unstable rows or against the wrong opponent is not a weaker\n * result, it is a result about nothing, and the report should say so first.\n */\nexport function evaluateRouterPromotion(\n\tevidence: RouterPromotionEvidence,\n\tpolicy: RouterPromotionPolicy = DEFAULT_ROUTER_PROMOTION_POLICY,\n): RouterPromotionVerdict {\n\tif (isMalformed(evidence)) {\n\t\treturn { blockers: [\"malformed_evidence\"], promote: false };\n\t}\n\n\tconst { baselineKind, goldenChanges, heldIn, holdout, humanApproved, stability } = evidence;\n\tconst blockers: RouterPromotionBlocker[] = [];\n\n\tif (stability.replays < policy.minReplays) {\n\t\tblockers.push(\"insufficient_replays\");\n\t}\n\tif (stability.unstable > policy.maxUnstable) {\n\t\tblockers.push(\"unstable_evidence\");\n\t}\n\tif (policy.requireFrozenBaseline && baselineKind !== \"frozen_reference\") {\n\t\tblockers.push(\"baseline_not_frozen\");\n\t}\n\tif (heldIn.baselineWins + heldIn.candidateWins < policy.minDiscordant) {\n\t\tblockers.push(\"insufficient_discordant_pairs\");\n\t}\n\tif (!heldIn.significant) {\n\t\tblockers.push(\"mcnemar_not_significant\");\n\t}\n\tif (heldIn.candidateWins <= heldIn.baselineWins) {\n\t\tblockers.push(\"candidate_not_favored\");\n\t}\n\tif (holdout.total < policy.minHoldout) {\n\t\tblockers.push(\"insufficient_holdout\");\n\t}\n\tif (holdout.candidateCorrect < holdout.baselineCorrect) {\n\t\tblockers.push(\"holdout_regression\");\n\t}\n\tif (goldenChanges > policy.maxGoldenChanges) {\n\t\tblockers.push(\"unreviewed_behavior_changes\");\n\t}\n\tif (policy.requireHumanApproval && !humanApproved) {\n\t\tblockers.push(\"human_approval_missing\");\n\t}\n\n\treturn { blockers, promote: blockers.length === 0 };\n}\n"]}

package/dist/core/reasoning-router-replay-stability.d.ts

@@ -0,0 +1,65 @@
+/**
+ * Two-run stability filter for reasoning-router promotion evidence.
+ *
+ * A promotion gate is only as trustworthy as the outcomes it credits. Routing
+ * evaluations are paired comparisons, and paired comparisons are cheap to fool:
+ * if the measured outcome of a row can flip between two identical replays, then
+ * a "win" on that row may be measurement noise rather than a better policy.
+ * Published routing evaluations report exactly this — a non-trivial share of
+ * model/question pairs score differently when the identical matrix is re-run —
+ * so a credit rule that ignores rerun disagreement will happily promote noise.
+ *
+ * This module implements the rule that closes that hole: a row carries credit
+ * only when it was observed at least `minReplays` times under each policy and
+ * every one of those observations agreed. Rows that disagree are withheld, not
+ * majority-voted — a flipped observation means the measurement is unreliable,
+ * not that the majority is the truth.
+ *
+ * For a deterministic classifier every row is stable by construction, so the
+ * filter is also a determinism attestation: if nondeterminism ever leaks into
+ * the routing path (iteration order, clock, randomness), rows start landing in
+ * the unstable bucket and the gate refuses to promote instead of crediting a
+ * coin flip.
+ *
+ * Pure and total: no I/O, no global state, no prompt text.
+ */
+/** Independent observations required per side before a row may carry credit. */
+export declare const DEFAULT_REPLAY_MIN = 2;
+/** One evaluated row, replayed independently under each policy. */
+export interface RouterReplayRow {
+    readonly rowId: string;
+    /** True when the row belongs to the frozen held-out split. */
+    readonly holdout: boolean;
+    /** Correctness observed on each independent replay under the baseline. */
+    readonly baselineReplays: readonly boolean[];
+    /** Correctness observed on each independent replay under the candidate. */
+    readonly candidateReplays: readonly boolean[];
+}
+/** A row whose replays agreed, so its outcome may carry promotion credit. */
+export interface RouterStableOutcome {
+    readonly rowId: string;
+    readonly holdout: boolean;
+    readonly baselineCorrect: boolean;
+    readonly candidateCorrect: boolean;
+}
+/** Credit-eligible outcomes plus why every other row was withheld. */
+export interface RouterReplayStability {
+    readonly stable: readonly RouterStableOutcome[];
+    /** Rows considered, whatever bucket they landed in. */
+    readonly evaluated: number;
+    /** Rows withheld because repeated observations disagreed. */
+    readonly unstable: number;
+    /** Rows withheld because one side was observed too few times to judge. */
+    readonly insufficientReplays: number;
+    readonly minReplays: number;
+}
+/**
+ * Partition replayed rows into credit-eligible outcomes and withheld rows.
+ *
+ * Insufficient replays outrank disagreement: a row observed once is not stable
+ * evidence that happens to agree with itself, it is evidence that was never
+ * tested for stability at all. Reporting the two buckets separately keeps that
+ * distinction visible to the gate, which fails closed on either.
+ */
+export declare function summarizeReplayStability(rows: readonly RouterReplayRow[], minReplays?: number): RouterReplayStability;
+//# sourceMappingURL=reasoning-router-replay-stability.d.ts.map

package/dist/core/reasoning-router-replay-stability.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"reasoning-router-replay-stability.d.ts","sourceRoot":"","sources":["../../src/core/reasoning-router-replay-stability.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAEH,gFAAgF;AAChF,eAAO,MAAM,kBAAkB,IAAI,CAAC;AAEpC,mEAAmE;AACnE,MAAM,WAAW,eAAe;IAC/B,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,8DAA8D;IAC9D,QAAQ,CAAC,OAAO,EAAE,OAAO,CAAC;IAC1B,0EAA0E;IAC1E,QAAQ,CAAC,eAAe,EAAE,SAAS,OAAO,EAAE,CAAC;IAC7C,2EAA2E;IAC3E,QAAQ,CAAC,gBAAgB,EAAE,SAAS,OAAO,EAAE,CAAC;CAC9C;AAED,6EAA6E;AAC7E,MAAM,WAAW,mBAAmB;IACnC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,OAAO,EAAE,OAAO,CAAC;IAC1B,QAAQ,CAAC,eAAe,EAAE,OAAO,CAAC;IAClC,QAAQ,CAAC,gBAAgB,EAAE,OAAO,CAAC;CACnC;AAED,sEAAsE;AACtE,MAAM,WAAW,qBAAqB;IACrC,QAAQ,CAAC,MAAM,EAAE,SAAS,mBAAmB,EAAE,CAAC;IAChD,uDAAuD;IACvD,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;IAC3B,6DAA6D;IAC7D,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC;IAC1B,0EAA0E;IAC1E,QAAQ,CAAC,mBAAmB,EAAE,MAAM,CAAC;IACrC,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;CAC5B;AAgBD;;;;;;;GAOG;AACH,wBAAgB,wBAAwB,CACvC,IAAI,EAAE,SAAS,eAAe,EAAE,EAChC,UAAU,GAAE,MAA2B,GACrC,qBAAqB,CA2BvB","sourcesContent":["/**\n * Two-run stability filter for reasoning-router promotion evidence.\n *\n * A promotion gate is only as trustworthy as the outcomes it credits. Routing\n * evaluations are paired comparisons, and paired comparisons are cheap to fool:\n * if the measured outcome of a row can flip between two identical replays, then\n * a \"win\" on that row may be measurement noise rather than a better policy.\n * Published routing evaluations report exactly this — a non-trivial share of\n * model/question pairs score differently when the identical matrix is re-run —\n * so a credit rule that ignores rerun disagreement will happily promote noise.\n *\n * This module implements the rule that closes that hole: a row carries credit\n * only when it was observed at least `minReplays` times under each policy and\n * every one of those observations agreed. Rows that disagree are withheld, not\n * majority-voted — a flipped observation means the measurement is unreliable,\n * not that the majority is the truth.\n *\n * For a deterministic classifier every row is stable by construction, so the\n * filter is also a determinism attestation: if nondeterminism ever leaks into\n * the routing path (iteration order, clock, randomness), rows start landing in\n * the unstable bucket and the gate refuses to promote instead of crediting a\n * coin flip.\n *\n * Pure and total: no I/O, no global state, no prompt text.\n */\n\n/** Independent observations required per side before a row may carry credit. */\nexport const DEFAULT_REPLAY_MIN = 2;\n\n/** One evaluated row, replayed independently under each policy. */\nexport interface RouterReplayRow {\n\treadonly rowId: string;\n\t/** True when the row belongs to the frozen held-out split. */\n\treadonly holdout: boolean;\n\t/** Correctness observed on each independent replay under the baseline. */\n\treadonly baselineReplays: readonly boolean[];\n\t/** Correctness observed on each independent replay under the candidate. */\n\treadonly candidateReplays: readonly boolean[];\n}\n\n/** A row whose replays agreed, so its outcome may carry promotion credit. */\nexport interface RouterStableOutcome {\n\treadonly rowId: string;\n\treadonly holdout: boolean;\n\treadonly baselineCorrect: boolean;\n\treadonly candidateCorrect: boolean;\n}\n\n/** Credit-eligible outcomes plus why every other row was withheld. */\nexport interface RouterReplayStability {\n\treadonly stable: readonly RouterStableOutcome[];\n\t/** Rows considered, whatever bucket they landed in. */\n\treadonly evaluated: number;\n\t/** Rows withheld because repeated observations disagreed. */\n\treadonly unstable: number;\n\t/** Rows withheld because one side was observed too few times to judge. */\n\treadonly insufficientReplays: number;\n\treadonly minReplays: number;\n}\n\n/** Verdict for one side of one row. */\ntype ReplayVerdict =\n\t| { readonly kind: \"agreed\"; readonly correct: boolean }\n\t| { readonly kind: \"disagreed\" }\n\t| { readonly kind: \"insufficient\" };\n\nfunction judgeReplays(replays: readonly boolean[], minReplays: number): ReplayVerdict {\n\tconst [first, ...rest] = replays;\n\tif (first === undefined || replays.length < minReplays) {\n\t\treturn { kind: \"insufficient\" };\n\t}\n\treturn rest.every((replay) => replay === first) ? { correct: first, kind: \"agreed\" } : { kind: \"disagreed\" };\n}\n\n/**\n * Partition replayed rows into credit-eligible outcomes and withheld rows.\n *\n * Insufficient replays outrank disagreement: a row observed once is not stable\n * evidence that happens to agree with itself, it is evidence that was never\n * tested for stability at all. Reporting the two buckets separately keeps that\n * distinction visible to the gate, which fails closed on either.\n */\nexport function summarizeReplayStability(\n\trows: readonly RouterReplayRow[],\n\tminReplays: number = DEFAULT_REPLAY_MIN,\n): RouterReplayStability {\n\tconst stable: RouterStableOutcome[] = [];\n\tlet unstable = 0;\n\tlet insufficientReplays = 0;\n\n\tfor (const row of rows) {\n\t\tconst baseline = judgeReplays(row.baselineReplays, minReplays);\n\t\tconst candidate = judgeReplays(row.candidateReplays, minReplays);\n\n\t\tif (baseline.kind === \"insufficient\" || candidate.kind === \"insufficient\") {\n\t\t\tinsufficientReplays++;\n\t\t\tcontinue;\n\t\t}\n\t\tif (baseline.kind === \"disagreed\" || candidate.kind === \"disagreed\") {\n\t\t\tunstable++;\n\t\t\tcontinue;\n\t\t}\n\n\t\tstable.push({\n\t\t\tbaselineCorrect: baseline.correct,\n\t\t\tcandidateCorrect: candidate.correct,\n\t\t\tholdout: row.holdout,\n\t\t\trowId: row.rowId,\n\t\t});\n\t}\n\n\treturn { evaluated: rows.length, insufficientReplays, minReplays, stable, unstable };\n}\n"]}

package/dist/core/reasoning-router-replay-stability.js

@@ -0,0 +1,67 @@
+/**
+ * Two-run stability filter for reasoning-router promotion evidence.
+ *
+ * A promotion gate is only as trustworthy as the outcomes it credits. Routing
+ * evaluations are paired comparisons, and paired comparisons are cheap to fool:
+ * if the measured outcome of a row can flip between two identical replays, then
+ * a "win" on that row may be measurement noise rather than a better policy.
+ * Published routing evaluations report exactly this — a non-trivial share of
+ * model/question pairs score differently when the identical matrix is re-run —
+ * so a credit rule that ignores rerun disagreement will happily promote noise.
+ *
+ * This module implements the rule that closes that hole: a row carries credit
+ * only when it was observed at least `minReplays` times under each policy and
+ * every one of those observations agreed. Rows that disagree are withheld, not
+ * majority-voted — a flipped observation means the measurement is unreliable,
+ * not that the majority is the truth.
+ *
+ * For a deterministic classifier every row is stable by construction, so the
+ * filter is also a determinism attestation: if nondeterminism ever leaks into
+ * the routing path (iteration order, clock, randomness), rows start landing in
+ * the unstable bucket and the gate refuses to promote instead of crediting a
+ * coin flip.
+ *
+ * Pure and total: no I/O, no global state, no prompt text.
+ */
+/** Independent observations required per side before a row may carry credit. */
+export const DEFAULT_REPLAY_MIN = 2;
+function judgeReplays(replays, minReplays) {
+    const [first, ...rest] = replays;
+    if (first === undefined || replays.length < minReplays) {
+        return { kind: "insufficient" };
+    }
+    return rest.every((replay) => replay === first) ? { correct: first, kind: "agreed" } : { kind: "disagreed" };
+}
+/**
+ * Partition replayed rows into credit-eligible outcomes and withheld rows.
+ *
+ * Insufficient replays outrank disagreement: a row observed once is not stable
+ * evidence that happens to agree with itself, it is evidence that was never
+ * tested for stability at all. Reporting the two buckets separately keeps that
+ * distinction visible to the gate, which fails closed on either.
+ */
+export function summarizeReplayStability(rows, minReplays = DEFAULT_REPLAY_MIN) {
+    const stable = [];
+    let unstable = 0;
+    let insufficientReplays = 0;
+    for (const row of rows) {
+        const baseline = judgeReplays(row.baselineReplays, minReplays);
+        const candidate = judgeReplays(row.candidateReplays, minReplays);
+        if (baseline.kind === "insufficient" || candidate.kind === "insufficient") {
+            insufficientReplays++;
+            continue;
+        }
+        if (baseline.kind === "disagreed" || candidate.kind === "disagreed") {
+            unstable++;
+            continue;
+        }
+        stable.push({
+            baselineCorrect: baseline.correct,
+            candidateCorrect: candidate.correct,
+            holdout: row.holdout,
+            rowId: row.rowId,
+        });
+    }
+    return { evaluated: rows.length, insufficientReplays, minReplays, stable, unstable };
+}
+//# sourceMappingURL=reasoning-router-replay-stability.js.map

package/dist/core/reasoning-router-replay-stability.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"reasoning-router-replay-stability.js","sourceRoot":"","sources":["../../src/core/reasoning-router-replay-stability.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAEH,gFAAgF;AAChF,MAAM,CAAC,MAAM,kBAAkB,GAAG,CAAC,CAAC;AAuCpC,SAAS,YAAY,CAAC,OAA2B,EAAE,UAAkB,EAAiB;IACrF,MAAM,CAAC,KAAK,EAAE,GAAG,IAAI,CAAC,GAAG,OAAO,CAAC;IACjC,IAAI,KAAK,KAAK,SAAS,IAAI,OAAO,CAAC,MAAM,GAAG,UAAU,EAAE,CAAC;QACxD,OAAO,EAAE,IAAI,EAAE,cAAc,EAAE,CAAC;IACjC,CAAC;IACD,OAAO,IAAI,CAAC,KAAK,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,MAAM,KAAK,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,EAAE,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,CAAC;AAAA,CAC7G;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,wBAAwB,CACvC,IAAgC,EAChC,UAAU,GAAW,kBAAkB,EACf;IACxB,MAAM,MAAM,GAA0B,EAAE,CAAC;IACzC,IAAI,QAAQ,GAAG,CAAC,CAAC;IACjB,IAAI,mBAAmB,GAAG,CAAC,CAAC;IAE5B,KAAK,MAAM,GAAG,IAAI,IAAI,EAAE,CAAC;QACxB,MAAM,QAAQ,GAAG,YAAY,CAAC,GAAG,CAAC,eAAe,EAAE,UAAU,CAAC,CAAC;QAC/D,MAAM,SAAS,GAAG,YAAY,CAAC,GAAG,CAAC,gBAAgB,EAAE,UAAU,CAAC,CAAC;QAEjE,IAAI,QAAQ,CAAC,IAAI,KAAK,cAAc,IAAI,SAAS,CAAC,IAAI,KAAK,cAAc,EAAE,CAAC;YAC3E,mBAAmB,EAAE,CAAC;YACtB,SAAS;QACV,CAAC;QACD,IAAI,QAAQ,CAAC,IAAI,KAAK,WAAW,IAAI,SAAS,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YACrE,QAAQ,EAAE,CAAC;YACX,SAAS;QACV,CAAC;QAED,MAAM,CAAC,IAAI,CAAC;YACX,eAAe,EAAE,QAAQ,CAAC,OAAO;YACjC,gBAAgB,EAAE,SAAS,CAAC,OAAO;YACnC,OAAO,EAAE,GAAG,CAAC,OAAO;YACpB,KAAK,EAAE,GAAG,CAAC,KAAK;SAChB,CAAC,CAAC;IACJ,CAAC;IAED,OAAO,EAAE,SAAS,EAAE,IAAI,CAAC,MAAM,EAAE,mBAAmB,EAAE,UAAU,EAAE,MAAM,EAAE,QAAQ,EAAE,CAAC;AAAA,CACrF","sourcesContent":["/**\n * Two-run stability filter for reasoning-router promotion evidence.\n *\n * A promotion gate is only as trustworthy as the outcomes it credits. Routing\n * evaluations are paired comparisons, and paired comparisons are cheap to fool:\n * if the measured outcome of a row can flip between two identical replays, then\n * a \"win\" on that row may be measurement noise rather than a better policy.\n * Published routing evaluations report exactly this — a non-trivial share of\n * model/question pairs score differently when the identical matrix is re-run —\n * so a credit rule that ignores rerun disagreement will happily promote noise.\n *\n * This module implements the rule that closes that hole: a row carries credit\n * only when it was observed at least `minReplays` times under each policy and\n * every one of those observations agreed. Rows that disagree are withheld, not\n * majority-voted — a flipped observation means the measurement is unreliable,\n * not that the majority is the truth.\n *\n * For a deterministic classifier every row is stable by construction, so the\n * filter is also a determinism attestation: if nondeterminism ever leaks into\n * the routing path (iteration order, clock, randomness), rows start landing in\n * the unstable bucket and the gate refuses to promote instead of crediting a\n * coin flip.\n *\n * Pure and total: no I/O, no global state, no prompt text.\n */\n\n/** Independent observations required per side before a row may carry credit. */\nexport const DEFAULT_REPLAY_MIN = 2;\n\n/** One evaluated row, replayed independently under each policy. */\nexport interface RouterReplayRow {\n\treadonly rowId: string;\n\t/** True when the row belongs to the frozen held-out split. */\n\treadonly holdout: boolean;\n\t/** Correctness observed on each independent replay under the baseline. */\n\treadonly baselineReplays: readonly boolean[];\n\t/** Correctness observed on each independent replay under the candidate. */\n\treadonly candidateReplays: readonly boolean[];\n}\n\n/** A row whose replays agreed, so its outcome may carry promotion credit. */\nexport interface RouterStableOutcome {\n\treadonly rowId: string;\n\treadonly holdout: boolean;\n\treadonly baselineCorrect: boolean;\n\treadonly candidateCorrect: boolean;\n}\n\n/** Credit-eligible outcomes plus why every other row was withheld. */\nexport interface RouterReplayStability {\n\treadonly stable: readonly RouterStableOutcome[];\n\t/** Rows considered, whatever bucket they landed in. */\n\treadonly evaluated: number;\n\t/** Rows withheld because repeated observations disagreed. */\n\treadonly unstable: number;\n\t/** Rows withheld because one side was observed too few times to judge. */\n\treadonly insufficientReplays: number;\n\treadonly minReplays: number;\n}\n\n/** Verdict for one side of one row. */\ntype ReplayVerdict =\n\t| { readonly kind: \"agreed\"; readonly correct: boolean }\n\t| { readonly kind: \"disagreed\" }\n\t| { readonly kind: \"insufficient\" };\n\nfunction judgeReplays(replays: readonly boolean[], minReplays: number): ReplayVerdict {\n\tconst [first, ...rest] = replays;\n\tif (first === undefined || replays.length < minReplays) {\n\t\treturn { kind: \"insufficient\" };\n\t}\n\treturn rest.every((replay) => replay === first) ? { correct: first, kind: \"agreed\" } : { kind: \"disagreed\" };\n}\n\n/**\n * Partition replayed rows into credit-eligible outcomes and withheld rows.\n *\n * Insufficient replays outrank disagreement: a row observed once is not stable\n * evidence that happens to agree with itself, it is evidence that was never\n * tested for stability at all. Reporting the two buckets separately keeps that\n * distinction visible to the gate, which fails closed on either.\n */\nexport function summarizeReplayStability(\n\trows: readonly RouterReplayRow[],\n\tminReplays: number = DEFAULT_REPLAY_MIN,\n): RouterReplayStability {\n\tconst stable: RouterStableOutcome[] = [];\n\tlet unstable = 0;\n\tlet insufficientReplays = 0;\n\n\tfor (const row of rows) {\n\t\tconst baseline = judgeReplays(row.baselineReplays, minReplays);\n\t\tconst candidate = judgeReplays(row.candidateReplays, minReplays);\n\n\t\tif (baseline.kind === \"insufficient\" || candidate.kind === \"insufficient\") {\n\t\t\tinsufficientReplays++;\n\t\t\tcontinue;\n\t\t}\n\t\tif (baseline.kind === \"disagreed\" || candidate.kind === \"disagreed\") {\n\t\t\tunstable++;\n\t\t\tcontinue;\n\t\t}\n\n\t\tstable.push({\n\t\t\tbaselineCorrect: baseline.correct,\n\t\t\tcandidateCorrect: candidate.correct,\n\t\t\tholdout: row.holdout,\n\t\t\trowId: row.rowId,\n\t\t});\n\t}\n\n\treturn { evaluated: rows.length, insufficientReplays, minReplays, stable, unstable };\n}\n"]}

package/dist/core/reasoning-router-v4-weights.d.ts

@@ -64,11 +64,11 @@ export interface RouterWeightsV4 {
     readonly intentSkeleton: number;
     /** Bump to "code-gen" for a bare "add" keyword, gated by !localEdit. */
     readonly addKeyword: number;
-    /** Bump to the class matching `history[0]`, if supplied (v4-new; 0 under DEFAULT_WEIGHTS_V4, inert until calibrated). */
+    /** Bump to the class matching `history[0]`, if supplied (default 2; gated on existing prompt evidence). */
     readonly multiTurnPrior: number;
-    /** Linear per-bucket bump applied to debug/review/plan under context pressure (v4-new; 0 under DEFAULT_WEIGHTS_V4, inert until calibrated). */
+    /** Linear per-bucket bump applied to debug/review/plan under context pressure (default 1). */
     readonly pressureBucket: number;
-    /** Bump to the class matching an externally supplied judge vote, if any (v4-new; 0 under DEFAULT_WEIGHTS_V4, inert until calibrated). */
+    /** Bump to an externally supplied judge vote, if any (default 2; gated on existing prompt evidence). */
     readonly judgeVote: number;
     /** Bounded look-back window (characters) for negation-cue gating, never crossing a .,;!? boundary. */
     readonly negationWindowChars: number;
@@ -80,9 +80,9 @@ export interface RouterWeightsV4 {
 /**
  * Calibrated default preset. The named weights below are the production v4
  * scorer configuration covered by
- * test/suite/regressions/013-reasoning-router-v4-accuracy.test.ts. Extension
- * fields (`multiTurnPrior`, `pressureBucket`, `judgeVote`) are wired but inert
- * until a future calibration changes them.
+ * test/suite/regressions/013-reasoning-router-v4-accuracy.test.ts. The bounded
+ * extension fields (`multiTurnPrior`, `pressureBucket`, `judgeVote`) are active
+ * only after the base scorer finds prompt evidence; zero-score fallbacks ignore them.
  */
 export declare const DEFAULT_WEIGHTS_V4: RouterWeightsV4;
 //# sourceMappingURL=reasoning-router-v4-weights.d.ts.map

package/dist/core/reasoning-router-v4-weights.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"reasoning-router-v4-weights.d.ts","sourceRoot":"","sources":["../../src/core/reasoning-router-v4-weights.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AAEH,qCAAqC;AACrC,MAAM,MAAM,WAAW,GAAG,SAAS,GAAG,aAAa,GAAG,UAAU,GAAG,OAAO,GAAG,UAAU,GAAG,QAAQ,GAAG,MAAM,CAAC;AAE5G,yEAAyE;AACzE,MAAM,MAAM,mBAAmB,GAAG,SAAS,GAAG,UAAU,GAAG,UAAU,GAAG,OAAO,GAAG,UAAU,GAAG,QAAQ,CAAC;AAExG,yIAAyI;AACzI,eAAO,MAAM,eAAe,EAAE,SAAS,WAAW,EAQjD,CAAC;AAEF;;;;GAIG;AACH,MAAM,WAAW,eAAe;IAC/B,oEAAoE;IACpE,QAAQ,CAAC,eAAe,EAAE,MAAM,CAAC;IACjC,+DAA+D;IAC/D,QAAQ,CAAC,aAAa,EAAE,MAAM,CAAC;IAC/B,6HAA6H;IAC7H,QAAQ,CAAC,kBAAkB,EAAE,MAAM,CAAC;IACpC,yEAAyE;IACzE,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;IAC3B,4EAA4E;IAC5E,QAAQ,CAAC,kBAAkB,EAAE,MAAM,CAAC;IACpC,sEAAsE;IACtE,QAAQ,CAAC,WAAW,EAAE,MAAM,CAAC;IAC7B,kEAAkE;IAClE,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;IAC3B,4GAA4G;IAC5G,QAAQ,CAAC,kBAAkB,EAAE,MAAM,CAAC;IACpC,uFAAuF;IACvF,QAAQ,CAAC,WAAW,EAAE,MAAM,CAAC;IAC7B,wEAAwE;IACxE,QAAQ,CAAC,oBAAoB,EAAE,MAAM,CAAC;IACtC,6GAA6G;IAC7G,QAAQ,CAAC,sBAAsB,EAAE,MAAM,CAAC;IACxC,kHAAkH;IAClH,QAAQ,CAAC,sBAAsB,EAAE,MAAM,CAAC;IACxC,4EAA4E;IAC5E,QAAQ,CAAC,aAAa,EAAE,QAAQ,CAAC,MAAM,CAAC,WAAW,EAAE,MAAM,CAAC,CAAC,CAAC;IAC9D,iGAAiG;IACjG,QAAQ,CAAC,uBAAuB,EAAE,MAAM,CAAC;IACzC,wGAAwG;IACxG,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAC;IAChC,wEAAwE;IACxE,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B,yHAAyH;IACzH,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAC;IAChC,+IAA+I;IAC/I,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAC;IAChC,yIAAyI;IACzI,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;IAC3B,sGAAsG;IACtG,QAAQ,CAAC,mBAAmB,EAAE,MAAM,CAAC;IACrC,4GAA4G;IAC5G,QAAQ,CAAC,sBAAsB,EAAE,MAAM,CAAC;IACxC,gDAAgD;IAChD,QAAQ,CAAC,uBAAuB,EAAE,MAAM,CAAC;CACzC;AAED;;;;;;GAMG;AACH,eAAO,MAAM,kBAAkB,EAAE,eAqChC,CAAC","sourcesContent":["/**\n * Reasoning-router v4 — frozen types and calibratable weight schema.\n *\n * Pure, side-effect-free, erasable TypeScript. No runtime imports. This file\n * owns the closed task/lane unions, named scorer weights, and calibrated default\n * preset used by `/think auto`.\n *\n * `keywordFamily` is the class -> weight table for whole-prompt keyword-family\n * patterns (see reasoning-router-v4.ts). `trivial` carries weight 0 because it\n * is reached only via the zero-score fallback cascade.\n *\n * `secondClauseIntent` applies a bounded bonus when a second, distinct leading\n * verb is found after a short-range conjunction split. `negationWindowChars`\n * bounds negation gating so a nearby \"don't/skip/without/no need\" cue can\n * suppress a whole-prompt signal without crossing a clause boundary.\n *\n * `lowConfidenceThreshold` / `highConfidenceThreshold` bound the confidence\n * bands (`ClassifierVerdictV4.confidenceBand`) computed from score separation\n * plus absolute evidence strength in reasoning-router-v4.ts. They are\n * metadata-only: they never change `taskClass`, only how\n * `resolveThinkingLevelV4WithUncertainty` treats an already-decided class.\n */\n\n/** Closed set of v4 task classes. */\nexport type TaskClassV4 = \"trivial\" | \"simple-edit\" | \"code-gen\" | \"debug\" | \"refactor\" | \"review\" | \"plan\";\n\n/** Subagent lane types recognized by the v4 classifier and resolvers. */\nexport type ReasoningLaneTypeV4 = \"planner\" | \"security\" | \"explorer\" | \"coder\" | \"reviewer\" | \"tester\";\n\n/** Fixed iteration order over TaskClassV4 (matches the type-declaration order; also v4's argmax tie-break precedence, highest first). */\nexport const TASK_CLASSES_V4: readonly TaskClassV4[] = [\n\t\"debug\",\n\t\"refactor\",\n\t\"review\",\n\t\"plan\",\n\t\"simple-edit\",\n\t\"code-gen\",\n\t\"trivial\",\n];\n\n/**\n * Named weight schema for the v4 scorer (reasoning-router-v4.ts). Every field\n * is a plain, deterministic integer/ratio; there is no clock, randomness, or\n * I/O anywhere in this file or its consumer.\n */\nexport interface RouterWeightsV4 {\n\t/** Bump to \"code-gen\" when a code fence or diff hunk is present. */\n\treadonly codeFenceOrDiff: number;\n\t/** Bump to the primary leading-clause intent class, if any. */\n\treadonly leadingIntent: number;\n\t/** Bump to a distinct second-clause intent class in a detected compound prompt (v4-new; default half of `leadingIntent`). */\n\treadonly secondClauseIntent: number;\n\t/** Bump to \"simple-edit\" when the composite local-edit feature fires. */\n\treadonly localEdit: number;\n\t/** Bump to \"debug\" when the composite diagnostic-evidence feature fires. */\n\treadonly diagnosticEvidence: number;\n\t/** Bump to \"review\" when the composite review-scope feature fires. */\n\treadonly reviewScope: number;\n\t/** Bump to \"plan\" when the composite plan-brief feature fires. */\n\treadonly planBrief: number;\n\t/** Bump to \"plan\" for release-bound operational runbooks (commit/push/tag/publish/changelog/CI bundles). */\n\treadonly operationalRunbook: number;\n\t/** Bump to \"refactor\" when the refactor-cue pattern matches anywhere in the prompt. */\n\treadonly refactorCue: number;\n\t/** Bump to \"code-gen\" when an implementation-object noun is present. */\n\treadonly implementationObject: number;\n\t/** Bump to \"code-gen\" when a first-person request names a code artifact (for example, \"I need a script\"). */\n\treadonly codeGenArtifactRequest: number;\n\t/** Bump to \"review\" when a leading review request names an evaluative object such as holes/regressions/issues. */\n\treadonly evaluativeReviewObject: number;\n\t/** Per-class whole-prompt keyword-family bump; trivial has no family, 0. */\n\treadonly keywordFamily: Readonly<Record<TaskClassV4, number>>;\n\t/** Bump to a class when normalized seed->cluster anchors match (medium generalized evidence). */\n\treadonly normalizedIntentCluster: number;\n\t/** Bump to a class when the bounded intent-skeleton extractor matches (medium generalized evidence). */\n\treadonly intentSkeleton: number;\n\t/** Bump to \"code-gen\" for a bare \"add\" keyword, gated by !localEdit. */\n\treadonly addKeyword: number;\n\t/** Bump to the class matching `history[0]`, if supplied (v4-new; 0 under DEFAULT_WEIGHTS_V4, inert until calibrated). */\n\treadonly multiTurnPrior: number;\n\t/** Linear per-bucket bump applied to debug/review/plan under context pressure (v4-new; 0 under DEFAULT_WEIGHTS_V4, inert until calibrated). */\n\treadonly pressureBucket: number;\n\t/** Bump to the class matching an externally supplied judge vote, if any (v4-new; 0 under DEFAULT_WEIGHTS_V4, inert until calibrated). */\n\treadonly judgeVote: number;\n\t/** Bounded look-back window (characters) for negation-cue gating, never crossing a .,;!? boundary. */\n\treadonly negationWindowChars: number;\n\t/** confidence <= this value bands as \"low\" (confidence combines score separation and evidence strength). */\n\treadonly lowConfidenceThreshold: number;\n\t/** confidence >= this value bands as \"high\". */\n\treadonly highConfidenceThreshold: number;\n}\n\n/**\n * Calibrated default preset. The named weights below are the production v4\n * scorer configuration covered by\n * test/suite/regressions/013-reasoning-router-v4-accuracy.test.ts. Extension\n * fields (`multiTurnPrior`, `pressureBucket`, `judgeVote`) are wired but inert\n * until a future calibration changes them.\n */\nexport const DEFAULT_WEIGHTS_V4: RouterWeightsV4 = {\n\tcodeFenceOrDiff: 4,\n\tleadingIntent: 8,\n\tsecondClauseIntent: 4,\n\tlocalEdit: 7,\n\tdiagnosticEvidence: 7,\n\treviewScope: 3,\n\tplanBrief: 8,\n\toperationalRunbook: 8,\n\trefactorCue: 6,\n\timplementationObject: 3,\n\tcodeGenArtifactRequest: 5,\n\tevaluativeReviewObject: 3,\n\tkeywordFamily: {\n\t\ttrivial: 0,\n\t\t\"simple-edit\": 4,\n\t\t\"code-gen\": 4,\n\t\tdebug: 4,\n\t\trefactor: 4,\n\t\treview: 4,\n\t\tplan: 4,\n\t},\n\tnormalizedIntentCluster: 4,\n\tintentSkeleton: 5,\n\taddKeyword: 1,\n\t// Goal 016/L2 calibration (scripts/reasoning-router/calibrate-inert-weights.ts):\n\t// adversarial bounds on the frozen gold set after the topScore>0 vote gate —\n\t// multiTurnPrior/judgeVote flip at 4 (gold-0066), pressureBucket at 2.\n\t// multiTurnPrior=2 / judgeVote=2 sit one step below the bound; pressureBucket=1\n\t// is provably safe because the runtime bucket is hard-clamped to <= 3, so the\n\t// max possible bump (+3) stays under the flip threshold (+6).\n\tmultiTurnPrior: 2,\n\tpressureBucket: 1,\n\tjudgeVote: 2,\n\tnegationWindowChars: 24,\n\tlowConfidenceThreshold: 0.35,\n\thighConfidenceThreshold: 0.7,\n};\n"]}
+{"version":3,"file":"reasoning-router-v4-weights.d.ts","sourceRoot":"","sources":["../../src/core/reasoning-router-v4-weights.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AAEH,qCAAqC;AACrC,MAAM,MAAM,WAAW,GAAG,SAAS,GAAG,aAAa,GAAG,UAAU,GAAG,OAAO,GAAG,UAAU,GAAG,QAAQ,GAAG,MAAM,CAAC;AAE5G,yEAAyE;AACzE,MAAM,MAAM,mBAAmB,GAAG,SAAS,GAAG,UAAU,GAAG,UAAU,GAAG,OAAO,GAAG,UAAU,GAAG,QAAQ,CAAC;AAExG,yIAAyI;AACzI,eAAO,MAAM,eAAe,EAAE,SAAS,WAAW,EAQjD,CAAC;AAEF;;;;GAIG;AACH,MAAM,WAAW,eAAe;IAC/B,oEAAoE;IACpE,QAAQ,CAAC,eAAe,EAAE,MAAM,CAAC;IACjC,+DAA+D;IAC/D,QAAQ,CAAC,aAAa,EAAE,MAAM,CAAC;IAC/B,6HAA6H;IAC7H,QAAQ,CAAC,kBAAkB,EAAE,MAAM,CAAC;IACpC,yEAAyE;IACzE,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;IAC3B,4EAA4E;IAC5E,QAAQ,CAAC,kBAAkB,EAAE,MAAM,CAAC;IACpC,sEAAsE;IACtE,QAAQ,CAAC,WAAW,EAAE,MAAM,CAAC;IAC7B,kEAAkE;IAClE,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;IAC3B,4GAA4G;IAC5G,QAAQ,CAAC,kBAAkB,EAAE,MAAM,CAAC;IACpC,uFAAuF;IACvF,QAAQ,CAAC,WAAW,EAAE,MAAM,CAAC;IAC7B,wEAAwE;IACxE,QAAQ,CAAC,oBAAoB,EAAE,MAAM,CAAC;IACtC,6GAA6G;IAC7G,QAAQ,CAAC,sBAAsB,EAAE,MAAM,CAAC;IACxC,kHAAkH;IAClH,QAAQ,CAAC,sBAAsB,EAAE,MAAM,CAAC;IACxC,4EAA4E;IAC5E,QAAQ,CAAC,aAAa,EAAE,QAAQ,CAAC,MAAM,CAAC,WAAW,EAAE,MAAM,CAAC,CAAC,CAAC;IAC9D,iGAAiG;IACjG,QAAQ,CAAC,uBAAuB,EAAE,MAAM,CAAC;IACzC,wGAAwG;IACxG,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAC;IAChC,wEAAwE;IACxE,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B,2GAA2G;IAC3G,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAC;IAChC,8FAA8F;IAC9F,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAC;IAChC,wGAAwG;IACxG,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;IAC3B,sGAAsG;IACtG,QAAQ,CAAC,mBAAmB,EAAE,MAAM,CAAC;IACrC,4GAA4G;IAC5G,QAAQ,CAAC,sBAAsB,EAAE,MAAM,CAAC;IACxC,gDAAgD;IAChD,QAAQ,CAAC,uBAAuB,EAAE,MAAM,CAAC;CACzC;AAED;;;;;;GAMG;AACH,eAAO,MAAM,kBAAkB,EAAE,eAqChC,CAAC","sourcesContent":["/**\n * Reasoning-router v4 — frozen types and calibratable weight schema.\n *\n * Pure, side-effect-free, erasable TypeScript. No runtime imports. This file\n * owns the closed task/lane unions, named scorer weights, and calibrated default\n * preset used by `/think auto`.\n *\n * `keywordFamily` is the class -> weight table for whole-prompt keyword-family\n * patterns (see reasoning-router-v4.ts). `trivial` carries weight 0 because it\n * is reached only via the zero-score fallback cascade.\n *\n * `secondClauseIntent` applies a bounded bonus when a second, distinct leading\n * verb is found after a short-range conjunction split. `negationWindowChars`\n * bounds negation gating so a nearby \"don't/skip/without/no need\" cue can\n * suppress a whole-prompt signal without crossing a clause boundary.\n *\n * `lowConfidenceThreshold` / `highConfidenceThreshold` bound the confidence\n * bands (`ClassifierVerdictV4.confidenceBand`) computed from score separation\n * plus absolute evidence strength in reasoning-router-v4.ts. They are\n * metadata-only: they never change `taskClass`, only how\n * `resolveThinkingLevelV4WithUncertainty` treats an already-decided class.\n */\n\n/** Closed set of v4 task classes. */\nexport type TaskClassV4 = \"trivial\" | \"simple-edit\" | \"code-gen\" | \"debug\" | \"refactor\" | \"review\" | \"plan\";\n\n/** Subagent lane types recognized by the v4 classifier and resolvers. */\nexport type ReasoningLaneTypeV4 = \"planner\" | \"security\" | \"explorer\" | \"coder\" | \"reviewer\" | \"tester\";\n\n/** Fixed iteration order over TaskClassV4 (matches the type-declaration order; also v4's argmax tie-break precedence, highest first). */\nexport const TASK_CLASSES_V4: readonly TaskClassV4[] = [\n\t\"debug\",\n\t\"refactor\",\n\t\"review\",\n\t\"plan\",\n\t\"simple-edit\",\n\t\"code-gen\",\n\t\"trivial\",\n];\n\n/**\n * Named weight schema for the v4 scorer (reasoning-router-v4.ts). Every field\n * is a plain, deterministic integer/ratio; there is no clock, randomness, or\n * I/O anywhere in this file or its consumer.\n */\nexport interface RouterWeightsV4 {\n\t/** Bump to \"code-gen\" when a code fence or diff hunk is present. */\n\treadonly codeFenceOrDiff: number;\n\t/** Bump to the primary leading-clause intent class, if any. */\n\treadonly leadingIntent: number;\n\t/** Bump to a distinct second-clause intent class in a detected compound prompt (v4-new; default half of `leadingIntent`). */\n\treadonly secondClauseIntent: number;\n\t/** Bump to \"simple-edit\" when the composite local-edit feature fires. */\n\treadonly localEdit: number;\n\t/** Bump to \"debug\" when the composite diagnostic-evidence feature fires. */\n\treadonly diagnosticEvidence: number;\n\t/** Bump to \"review\" when the composite review-scope feature fires. */\n\treadonly reviewScope: number;\n\t/** Bump to \"plan\" when the composite plan-brief feature fires. */\n\treadonly planBrief: number;\n\t/** Bump to \"plan\" for release-bound operational runbooks (commit/push/tag/publish/changelog/CI bundles). */\n\treadonly operationalRunbook: number;\n\t/** Bump to \"refactor\" when the refactor-cue pattern matches anywhere in the prompt. */\n\treadonly refactorCue: number;\n\t/** Bump to \"code-gen\" when an implementation-object noun is present. */\n\treadonly implementationObject: number;\n\t/** Bump to \"code-gen\" when a first-person request names a code artifact (for example, \"I need a script\"). */\n\treadonly codeGenArtifactRequest: number;\n\t/** Bump to \"review\" when a leading review request names an evaluative object such as holes/regressions/issues. */\n\treadonly evaluativeReviewObject: number;\n\t/** Per-class whole-prompt keyword-family bump; trivial has no family, 0. */\n\treadonly keywordFamily: Readonly<Record<TaskClassV4, number>>;\n\t/** Bump to a class when normalized seed->cluster anchors match (medium generalized evidence). */\n\treadonly normalizedIntentCluster: number;\n\t/** Bump to a class when the bounded intent-skeleton extractor matches (medium generalized evidence). */\n\treadonly intentSkeleton: number;\n\t/** Bump to \"code-gen\" for a bare \"add\" keyword, gated by !localEdit. */\n\treadonly addKeyword: number;\n\t/** Bump to the class matching `history[0]`, if supplied (default 2; gated on existing prompt evidence). */\n\treadonly multiTurnPrior: number;\n\t/** Linear per-bucket bump applied to debug/review/plan under context pressure (default 1). */\n\treadonly pressureBucket: number;\n\t/** Bump to an externally supplied judge vote, if any (default 2; gated on existing prompt evidence). */\n\treadonly judgeVote: number;\n\t/** Bounded look-back window (characters) for negation-cue gating, never crossing a .,;!? boundary. */\n\treadonly negationWindowChars: number;\n\t/** confidence <= this value bands as \"low\" (confidence combines score separation and evidence strength). */\n\treadonly lowConfidenceThreshold: number;\n\t/** confidence >= this value bands as \"high\". */\n\treadonly highConfidenceThreshold: number;\n}\n\n/**\n * Calibrated default preset. The named weights below are the production v4\n * scorer configuration covered by\n * test/suite/regressions/013-reasoning-router-v4-accuracy.test.ts. The bounded\n * extension fields (`multiTurnPrior`, `pressureBucket`, `judgeVote`) are active\n * only after the base scorer finds prompt evidence; zero-score fallbacks ignore them.\n */\nexport const DEFAULT_WEIGHTS_V4: RouterWeightsV4 = {\n\tcodeFenceOrDiff: 4,\n\tleadingIntent: 8,\n\tsecondClauseIntent: 4,\n\tlocalEdit: 7,\n\tdiagnosticEvidence: 7,\n\treviewScope: 3,\n\tplanBrief: 8,\n\toperationalRunbook: 8,\n\trefactorCue: 6,\n\timplementationObject: 3,\n\tcodeGenArtifactRequest: 5,\n\tevaluativeReviewObject: 3,\n\tkeywordFamily: {\n\t\ttrivial: 0,\n\t\t\"simple-edit\": 4,\n\t\t\"code-gen\": 4,\n\t\tdebug: 4,\n\t\trefactor: 4,\n\t\treview: 4,\n\t\tplan: 4,\n\t},\n\tnormalizedIntentCluster: 4,\n\tintentSkeleton: 5,\n\taddKeyword: 1,\n\t// Goal 016/L2 calibration (scripts/reasoning-router/calibrate-inert-weights.ts):\n\t// adversarial bounds on the frozen gold set after the topScore>0 vote gate —\n\t// multiTurnPrior/judgeVote flip at 4 (gold-0066), pressureBucket at 2.\n\t// multiTurnPrior=2 / judgeVote=2 sit one step below the bound; pressureBucket=1\n\t// is provably safe because the runtime bucket is hard-clamped to <= 3, so the\n\t// max possible bump (+3) stays under the flip threshold (+6).\n\tmultiTurnPrior: 2,\n\tpressureBucket: 1,\n\tjudgeVote: 2,\n\tnegationWindowChars: 24,\n\tlowConfidenceThreshold: 0.35,\n\thighConfidenceThreshold: 0.7,\n};\n"]}

package/dist/core/reasoning-router-v4-weights.js

@@ -33,9 +33,9 @@ export const TASK_CLASSES_V4 = [
 /**
  * Calibrated default preset. The named weights below are the production v4
  * scorer configuration covered by
- * test/suite/regressions/013-reasoning-router-v4-accuracy.test.ts. Extension
- * fields (`multiTurnPrior`, `pressureBucket`, `judgeVote`) are wired but inert
- * until a future calibration changes them.
+ * test/suite/regressions/013-reasoning-router-v4-accuracy.test.ts. The bounded
+ * extension fields (`multiTurnPrior`, `pressureBucket`, `judgeVote`) are active
+ * only after the base scorer finds prompt evidence; zero-score fallbacks ignore them.
  */
 export const DEFAULT_WEIGHTS_V4 = {
     codeFenceOrDiff: 4,

package/dist/core/reasoning-router-v4-weights.js.map

@@ -1 +1 @@
-{"version":3,"file":"reasoning-router-v4-weights.js","sourceRoot":"","sources":["../../src/core/reasoning-router-v4-weights.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AAQH,yIAAyI;AACzI,MAAM,CAAC,MAAM,eAAe,GAA2B;IACtD,OAAO;IACP,UAAU;IACV,QAAQ;IACR,MAAM;IACN,aAAa;IACb,UAAU;IACV,SAAS;CACT,CAAC;AAsDF;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,kBAAkB,GAAoB;IAClD,eAAe,EAAE,CAAC;IAClB,aAAa,EAAE,CAAC;IAChB,kBAAkB,EAAE,CAAC;IACrB,SAAS,EAAE,CAAC;IACZ,kBAAkB,EAAE,CAAC;IACrB,WAAW,EAAE,CAAC;IACd,SAAS,EAAE,CAAC;IACZ,kBAAkB,EAAE,CAAC;IACrB,WAAW,EAAE,CAAC;IACd,oBAAoB,EAAE,CAAC;IACvB,sBAAsB,EAAE,CAAC;IACzB,sBAAsB,EAAE,CAAC;IACzB,aAAa,EAAE;QACd,OAAO,EAAE,CAAC;QACV,aAAa,EAAE,CAAC;QAChB,UAAU,EAAE,CAAC;QACb,KAAK,EAAE,CAAC;QACR,QAAQ,EAAE,CAAC;QACX,MAAM,EAAE,CAAC;QACT,IAAI,EAAE,CAAC;KACP;IACD,uBAAuB,EAAE,CAAC;IAC1B,cAAc,EAAE,CAAC;IACjB,UAAU,EAAE,CAAC;IACb,iFAAiF;IACjF,+EAA6E;IAC7E,uEAAuE;IACvE,gFAAgF;IAChF,8EAA8E;IAC9E,8DAA8D;IAC9D,cAAc,EAAE,CAAC;IACjB,cAAc,EAAE,CAAC;IACjB,SAAS,EAAE,CAAC;IACZ,mBAAmB,EAAE,EAAE;IACvB,sBAAsB,EAAE,IAAI;IAC5B,uBAAuB,EAAE,GAAG;CAC5B,CAAC","sourcesContent":["/**\n * Reasoning-router v4 — frozen types and calibratable weight schema.\n *\n * Pure, side-effect-free, erasable TypeScript. No runtime imports. This file\n * owns the closed task/lane unions, named scorer weights, and calibrated default\n * preset used by `/think auto`.\n *\n * `keywordFamily` is the class -> weight table for whole-prompt keyword-family\n * patterns (see reasoning-router-v4.ts). `trivial` carries weight 0 because it\n * is reached only via the zero-score fallback cascade.\n *\n * `secondClauseIntent` applies a bounded bonus when a second, distinct leading\n * verb is found after a short-range conjunction split. `negationWindowChars`\n * bounds negation gating so a nearby \"don't/skip/without/no need\" cue can\n * suppress a whole-prompt signal without crossing a clause boundary.\n *\n * `lowConfidenceThreshold` / `highConfidenceThreshold` bound the confidence\n * bands (`ClassifierVerdictV4.confidenceBand`) computed from score separation\n * plus absolute evidence strength in reasoning-router-v4.ts. They are\n * metadata-only: they never change `taskClass`, only how\n * `resolveThinkingLevelV4WithUncertainty` treats an already-decided class.\n */\n\n/** Closed set of v4 task classes. */\nexport type TaskClassV4 = \"trivial\" | \"simple-edit\" | \"code-gen\" | \"debug\" | \"refactor\" | \"review\" | \"plan\";\n\n/** Subagent lane types recognized by the v4 classifier and resolvers. */\nexport type ReasoningLaneTypeV4 = \"planner\" | \"security\" | \"explorer\" | \"coder\" | \"reviewer\" | \"tester\";\n\n/** Fixed iteration order over TaskClassV4 (matches the type-declaration order; also v4's argmax tie-break precedence, highest first). */\nexport const TASK_CLASSES_V4: readonly TaskClassV4[] = [\n\t\"debug\",\n\t\"refactor\",\n\t\"review\",\n\t\"plan\",\n\t\"simple-edit\",\n\t\"code-gen\",\n\t\"trivial\",\n];\n\n/**\n * Named weight schema for the v4 scorer (reasoning-router-v4.ts). Every field\n * is a plain, deterministic integer/ratio; there is no clock, randomness, or\n * I/O anywhere in this file or its consumer.\n */\nexport interface RouterWeightsV4 {\n\t/** Bump to \"code-gen\" when a code fence or diff hunk is present. */\n\treadonly codeFenceOrDiff: number;\n\t/** Bump to the primary leading-clause intent class, if any. */\n\treadonly leadingIntent: number;\n\t/** Bump to a distinct second-clause intent class in a detected compound prompt (v4-new; default half of `leadingIntent`). */\n\treadonly secondClauseIntent: number;\n\t/** Bump to \"simple-edit\" when the composite local-edit feature fires. */\n\treadonly localEdit: number;\n\t/** Bump to \"debug\" when the composite diagnostic-evidence feature fires. */\n\treadonly diagnosticEvidence: number;\n\t/** Bump to \"review\" when the composite review-scope feature fires. */\n\treadonly reviewScope: number;\n\t/** Bump to \"plan\" when the composite plan-brief feature fires. */\n\treadonly planBrief: number;\n\t/** Bump to \"plan\" for release-bound operational runbooks (commit/push/tag/publish/changelog/CI bundles). */\n\treadonly operationalRunbook: number;\n\t/** Bump to \"refactor\" when the refactor-cue pattern matches anywhere in the prompt. */\n\treadonly refactorCue: number;\n\t/** Bump to \"code-gen\" when an implementation-object noun is present. */\n\treadonly implementationObject: number;\n\t/** Bump to \"code-gen\" when a first-person request names a code artifact (for example, \"I need a script\"). */\n\treadonly codeGenArtifactRequest: number;\n\t/** Bump to \"review\" when a leading review request names an evaluative object such as holes/regressions/issues. */\n\treadonly evaluativeReviewObject: number;\n\t/** Per-class whole-prompt keyword-family bump; trivial has no family, 0. */\n\treadonly keywordFamily: Readonly<Record<TaskClassV4, number>>;\n\t/** Bump to a class when normalized seed->cluster anchors match (medium generalized evidence). */\n\treadonly normalizedIntentCluster: number;\n\t/** Bump to a class when the bounded intent-skeleton extractor matches (medium generalized evidence). */\n\treadonly intentSkeleton: number;\n\t/** Bump to \"code-gen\" for a bare \"add\" keyword, gated by !localEdit. */\n\treadonly addKeyword: number;\n\t/** Bump to the class matching `history[0]`, if supplied (v4-new; 0 under DEFAULT_WEIGHTS_V4, inert until calibrated). */\n\treadonly multiTurnPrior: number;\n\t/** Linear per-bucket bump applied to debug/review/plan under context pressure (v4-new; 0 under DEFAULT_WEIGHTS_V4, inert until calibrated). */\n\treadonly pressureBucket: number;\n\t/** Bump to the class matching an externally supplied judge vote, if any (v4-new; 0 under DEFAULT_WEIGHTS_V4, inert until calibrated). */\n\treadonly judgeVote: number;\n\t/** Bounded look-back window (characters) for negation-cue gating, never crossing a .,;!? boundary. */\n\treadonly negationWindowChars: number;\n\t/** confidence <= this value bands as \"low\" (confidence combines score separation and evidence strength). */\n\treadonly lowConfidenceThreshold: number;\n\t/** confidence >= this value bands as \"high\". */\n\treadonly highConfidenceThreshold: number;\n}\n\n/**\n * Calibrated default preset. The named weights below are the production v4\n * scorer configuration covered by\n * test/suite/regressions/013-reasoning-router-v4-accuracy.test.ts. Extension\n * fields (`multiTurnPrior`, `pressureBucket`, `judgeVote`) are wired but inert\n * until a future calibration changes them.\n */\nexport const DEFAULT_WEIGHTS_V4: RouterWeightsV4 = {\n\tcodeFenceOrDiff: 4,\n\tleadingIntent: 8,\n\tsecondClauseIntent: 4,\n\tlocalEdit: 7,\n\tdiagnosticEvidence: 7,\n\treviewScope: 3,\n\tplanBrief: 8,\n\toperationalRunbook: 8,\n\trefactorCue: 6,\n\timplementationObject: 3,\n\tcodeGenArtifactRequest: 5,\n\tevaluativeReviewObject: 3,\n\tkeywordFamily: {\n\t\ttrivial: 0,\n\t\t\"simple-edit\": 4,\n\t\t\"code-gen\": 4,\n\t\tdebug: 4,\n\t\trefactor: 4,\n\t\treview: 4,\n\t\tplan: 4,\n\t},\n\tnormalizedIntentCluster: 4,\n\tintentSkeleton: 5,\n\taddKeyword: 1,\n\t// Goal 016/L2 calibration (scripts/reasoning-router/calibrate-inert-weights.ts):\n\t// adversarial bounds on the frozen gold set after the topScore>0 vote gate —\n\t// multiTurnPrior/judgeVote flip at 4 (gold-0066), pressureBucket at 2.\n\t// multiTurnPrior=2 / judgeVote=2 sit one step below the bound; pressureBucket=1\n\t// is provably safe because the runtime bucket is hard-clamped to <= 3, so the\n\t// max possible bump (+3) stays under the flip threshold (+6).\n\tmultiTurnPrior: 2,\n\tpressureBucket: 1,\n\tjudgeVote: 2,\n\tnegationWindowChars: 24,\n\tlowConfidenceThreshold: 0.35,\n\thighConfidenceThreshold: 0.7,\n};\n"]}
+{"version":3,"file":"reasoning-router-v4-weights.js","sourceRoot":"","sources":["../../src/core/reasoning-router-v4-weights.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AAQH,yIAAyI;AACzI,MAAM,CAAC,MAAM,eAAe,GAA2B;IACtD,OAAO;IACP,UAAU;IACV,QAAQ;IACR,MAAM;IACN,aAAa;IACb,UAAU;IACV,SAAS;CACT,CAAC;AAsDF;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,kBAAkB,GAAoB;IAClD,eAAe,EAAE,CAAC;IAClB,aAAa,EAAE,CAAC;IAChB,kBAAkB,EAAE,CAAC;IACrB,SAAS,EAAE,CAAC;IACZ,kBAAkB,EAAE,CAAC;IACrB,WAAW,EAAE,CAAC;IACd,SAAS,EAAE,CAAC;IACZ,kBAAkB,EAAE,CAAC;IACrB,WAAW,EAAE,CAAC;IACd,oBAAoB,EAAE,CAAC;IACvB,sBAAsB,EAAE,CAAC;IACzB,sBAAsB,EAAE,CAAC;IACzB,aAAa,EAAE;QACd,OAAO,EAAE,CAAC;QACV,aAAa,EAAE,CAAC;QAChB,UAAU,EAAE,CAAC;QACb,KAAK,EAAE,CAAC;QACR,QAAQ,EAAE,CAAC;QACX,MAAM,EAAE,CAAC;QACT,IAAI,EAAE,CAAC;KACP;IACD,uBAAuB,EAAE,CAAC;IAC1B,cAAc,EAAE,CAAC;IACjB,UAAU,EAAE,CAAC;IACb,iFAAiF;IACjF,+EAA6E;IAC7E,uEAAuE;IACvE,gFAAgF;IAChF,8EAA8E;IAC9E,8DAA8D;IAC9D,cAAc,EAAE,CAAC;IACjB,cAAc,EAAE,CAAC;IACjB,SAAS,EAAE,CAAC;IACZ,mBAAmB,EAAE,EAAE;IACvB,sBAAsB,EAAE,IAAI;IAC5B,uBAAuB,EAAE,GAAG;CAC5B,CAAC","sourcesContent":["/**\n * Reasoning-router v4 — frozen types and calibratable weight schema.\n *\n * Pure, side-effect-free, erasable TypeScript. No runtime imports. This file\n * owns the closed task/lane unions, named scorer weights, and calibrated default\n * preset used by `/think auto`.\n *\n * `keywordFamily` is the class -> weight table for whole-prompt keyword-family\n * patterns (see reasoning-router-v4.ts). `trivial` carries weight 0 because it\n * is reached only via the zero-score fallback cascade.\n *\n * `secondClauseIntent` applies a bounded bonus when a second, distinct leading\n * verb is found after a short-range conjunction split. `negationWindowChars`\n * bounds negation gating so a nearby \"don't/skip/without/no need\" cue can\n * suppress a whole-prompt signal without crossing a clause boundary.\n *\n * `lowConfidenceThreshold` / `highConfidenceThreshold` bound the confidence\n * bands (`ClassifierVerdictV4.confidenceBand`) computed from score separation\n * plus absolute evidence strength in reasoning-router-v4.ts. They are\n * metadata-only: they never change `taskClass`, only how\n * `resolveThinkingLevelV4WithUncertainty` treats an already-decided class.\n */\n\n/** Closed set of v4 task classes. */\nexport type TaskClassV4 = \"trivial\" | \"simple-edit\" | \"code-gen\" | \"debug\" | \"refactor\" | \"review\" | \"plan\";\n\n/** Subagent lane types recognized by the v4 classifier and resolvers. */\nexport type ReasoningLaneTypeV4 = \"planner\" | \"security\" | \"explorer\" | \"coder\" | \"reviewer\" | \"tester\";\n\n/** Fixed iteration order over TaskClassV4 (matches the type-declaration order; also v4's argmax tie-break precedence, highest first). */\nexport const TASK_CLASSES_V4: readonly TaskClassV4[] = [\n\t\"debug\",\n\t\"refactor\",\n\t\"review\",\n\t\"plan\",\n\t\"simple-edit\",\n\t\"code-gen\",\n\t\"trivial\",\n];\n\n/**\n * Named weight schema for the v4 scorer (reasoning-router-v4.ts). Every field\n * is a plain, deterministic integer/ratio; there is no clock, randomness, or\n * I/O anywhere in this file or its consumer.\n */\nexport interface RouterWeightsV4 {\n\t/** Bump to \"code-gen\" when a code fence or diff hunk is present. */\n\treadonly codeFenceOrDiff: number;\n\t/** Bump to the primary leading-clause intent class, if any. */\n\treadonly leadingIntent: number;\n\t/** Bump to a distinct second-clause intent class in a detected compound prompt (v4-new; default half of `leadingIntent`). */\n\treadonly secondClauseIntent: number;\n\t/** Bump to \"simple-edit\" when the composite local-edit feature fires. */\n\treadonly localEdit: number;\n\t/** Bump to \"debug\" when the composite diagnostic-evidence feature fires. */\n\treadonly diagnosticEvidence: number;\n\t/** Bump to \"review\" when the composite review-scope feature fires. */\n\treadonly reviewScope: number;\n\t/** Bump to \"plan\" when the composite plan-brief feature fires. */\n\treadonly planBrief: number;\n\t/** Bump to \"plan\" for release-bound operational runbooks (commit/push/tag/publish/changelog/CI bundles). */\n\treadonly operationalRunbook: number;\n\t/** Bump to \"refactor\" when the refactor-cue pattern matches anywhere in the prompt. */\n\treadonly refactorCue: number;\n\t/** Bump to \"code-gen\" when an implementation-object noun is present. */\n\treadonly implementationObject: number;\n\t/** Bump to \"code-gen\" when a first-person request names a code artifact (for example, \"I need a script\"). */\n\treadonly codeGenArtifactRequest: number;\n\t/** Bump to \"review\" when a leading review request names an evaluative object such as holes/regressions/issues. */\n\treadonly evaluativeReviewObject: number;\n\t/** Per-class whole-prompt keyword-family bump; trivial has no family, 0. */\n\treadonly keywordFamily: Readonly<Record<TaskClassV4, number>>;\n\t/** Bump to a class when normalized seed->cluster anchors match (medium generalized evidence). */\n\treadonly normalizedIntentCluster: number;\n\t/** Bump to a class when the bounded intent-skeleton extractor matches (medium generalized evidence). */\n\treadonly intentSkeleton: number;\n\t/** Bump to \"code-gen\" for a bare \"add\" keyword, gated by !localEdit. */\n\treadonly addKeyword: number;\n\t/** Bump to the class matching `history[0]`, if supplied (default 2; gated on existing prompt evidence). */\n\treadonly multiTurnPrior: number;\n\t/** Linear per-bucket bump applied to debug/review/plan under context pressure (default 1). */\n\treadonly pressureBucket: number;\n\t/** Bump to an externally supplied judge vote, if any (default 2; gated on existing prompt evidence). */\n\treadonly judgeVote: number;\n\t/** Bounded look-back window (characters) for negation-cue gating, never crossing a .,;!? boundary. */\n\treadonly negationWindowChars: number;\n\t/** confidence <= this value bands as \"low\" (confidence combines score separation and evidence strength). */\n\treadonly lowConfidenceThreshold: number;\n\t/** confidence >= this value bands as \"high\". */\n\treadonly highConfidenceThreshold: number;\n}\n\n/**\n * Calibrated default preset. The named weights below are the production v4\n * scorer configuration covered by\n * test/suite/regressions/013-reasoning-router-v4-accuracy.test.ts. The bounded\n * extension fields (`multiTurnPrior`, `pressureBucket`, `judgeVote`) are active\n * only after the base scorer finds prompt evidence; zero-score fallbacks ignore them.\n */\nexport const DEFAULT_WEIGHTS_V4: RouterWeightsV4 = {\n\tcodeFenceOrDiff: 4,\n\tleadingIntent: 8,\n\tsecondClauseIntent: 4,\n\tlocalEdit: 7,\n\tdiagnosticEvidence: 7,\n\treviewScope: 3,\n\tplanBrief: 8,\n\toperationalRunbook: 8,\n\trefactorCue: 6,\n\timplementationObject: 3,\n\tcodeGenArtifactRequest: 5,\n\tevaluativeReviewObject: 3,\n\tkeywordFamily: {\n\t\ttrivial: 0,\n\t\t\"simple-edit\": 4,\n\t\t\"code-gen\": 4,\n\t\tdebug: 4,\n\t\trefactor: 4,\n\t\treview: 4,\n\t\tplan: 4,\n\t},\n\tnormalizedIntentCluster: 4,\n\tintentSkeleton: 5,\n\taddKeyword: 1,\n\t// Goal 016/L2 calibration (scripts/reasoning-router/calibrate-inert-weights.ts):\n\t// adversarial bounds on the frozen gold set after the topScore>0 vote gate —\n\t// multiTurnPrior/judgeVote flip at 4 (gold-0066), pressureBucket at 2.\n\t// multiTurnPrior=2 / judgeVote=2 sit one step below the bound; pressureBucket=1\n\t// is provably safe because the runtime bucket is hard-clamped to <= 3, so the\n\t// max possible bump (+3) stays under the flip threshold (+6).\n\tmultiTurnPrior: 2,\n\tpressureBucket: 1,\n\tjudgeVote: 2,\n\tnegationWindowChars: 24,\n\tlowConfidenceThreshold: 0.35,\n\thighConfidenceThreshold: 0.7,\n};\n"]}