@jinn-network/client 0.1.7-canary.d4e4e183 → 0.1.7-canary.e68c8f43

package/dist/learner/revert-decision.js

@@ -0,0 +1,53 @@
+/**
+ * Revert-decision logic for the learner's Memory-consolidation phase (#764).
+ *
+ * Given per-codeDigest pass-rate aggregates for a candidate Improve commit
+ * (`withCommit`) and its parent (`atParent`), decide whether the commit
+ * significantly regressed the frozen-eval pass rate and should be reverted.
+ *
+ * Thresholds are explicit and documented here (AC4) — not magic constants:
+ *   - minSamplesPerArm: minimum indexed attempts required in EACH arm before a
+ *     statistical test is meaningful. Below it, plateau is expected Level-1
+ *     behaviour, so we abstain (reason 'insufficient_samples').
+ *   - alpha: significance threshold (revert only when p < alpha, two-sided).
+ *   - recentAttemptsWindow: how many most-recent attempts per codeDigest the
+ *     aggregate is computed over. Overridable via implStateDir/policy.json
+ *     (`policy.revert.recentAttemptsWindow`), mirroring policy.maxNotesBytes.
+ */
+import { twoProportionZTest } from './revert-stats.js';
+export const DEFAULT_REVERT_POLICY = {
+    minSamplesPerArm: 30,
+    alpha: 0.05,
+    recentAttemptsWindow: 200,
+};
+export function decideRevert(input, policy = DEFAULT_REVERT_POLICY) {
+    const { withCommit, atParent } = input;
+    const base = {
+        withCommit: { codeDigest: withCommit.codeDigest, n: withCommit.attempts, passRate: withCommit.passRate },
+        atParent: { codeDigest: atParent.codeDigest, n: atParent.attempts, passRate: atParent.passRate },
+    };
+    // Sample floor first — a zero-attempt codeDigest is "insufficient_samples",
+    // NOT pass-rate zero (a fresh promotion that has not run yet is not a regression).
+    if (withCommit.attempts < policy.minSamplesPerArm || atParent.attempts < policy.minSamplesPerArm) {
+        return { ...base, delta: withCommit.passRate - atParent.passRate, pValue: 1, significant: false, recommendRevert: false, reason: 'insufficient_samples' };
+    }
+    const stats = twoProportionZTest({
+        passesA: withCommit.passes,
+        totalA: withCommit.attempts,
+        passesB: atParent.passes,
+        totalB: atParent.attempts,
+    });
+    const significant = stats.pValue < policy.alpha;
+    if (stats.delta >= 0) {
+        return { ...base, delta: stats.delta, pValue: stats.pValue, significant, recommendRevert: false, reason: 'no_regression' };
+    }
+    if (!significant) {
+        return { ...base, delta: stats.delta, pValue: stats.pValue, significant, recommendRevert: false, reason: 'not_significant' };
+    }
+    return { ...base, delta: stats.delta, pValue: stats.pValue, significant: true, recommendRevert: true, reason: 'significant_regression' };
+}
+/** Merge a partial policy (e.g. from implStateDir/policy.json) over the defaults. */
+export function resolveRevertPolicy(override) {
+    return { ...DEFAULT_REVERT_POLICY, ...(override ?? {}) };
+}
+//# sourceMappingURL=revert-decision.js.map

package/dist/learner/revert-decision.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"revert-decision.js","sourceRoot":"","sources":["../../src/learner/revert-decision.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAEH,OAAO,EAAE,kBAAkB,EAAE,MAAM,mBAAmB,CAAC;AAWvD,MAAM,CAAC,MAAM,qBAAqB,GAAiB;IACjD,gBAAgB,EAAE,EAAE;IACpB,KAAK,EAAE,IAAI;IACX,oBAAoB,EAAE,GAAG;CAC1B,CAAC;AAiCF,MAAM,UAAU,YAAY,CAC1B,KAA0B,EAC1B,SAAuB,qBAAqB;IAE5C,MAAM,EAAE,UAAU,EAAE,QAAQ,EAAE,GAAG,KAAK,CAAC;IACvC,MAAM,IAAI,GAAG;QACX,UAAU,EAAE,EAAE,UAAU,EAAE,UAAU,CAAC,UAAU,EAAE,CAAC,EAAE,UAAU,CAAC,QAAQ,EAAE,QAAQ,EAAE,UAAU,CAAC,QAAQ,EAAE;QACxG,QAAQ,EAAE,EAAE,UAAU,EAAE,QAAQ,CAAC,UAAU,EAAE,CAAC,EAAE,QAAQ,CAAC,QAAQ,EAAE,QAAQ,EAAE,QAAQ,CAAC,QAAQ,EAAE;KACjG,CAAC;IAEF,4EAA4E;IAC5E,mFAAmF;IACnF,IAAI,UAAU,CAAC,QAAQ,GAAG,MAAM,CAAC,gBAAgB,IAAI,QAAQ,CAAC,QAAQ,GAAG,MAAM,CAAC,gBAAgB,EAAE,CAAC;QACjG,OAAO,EAAE,GAAG,IAAI,EAAE,KAAK,EAAE,UAAU,CAAC,QAAQ,GAAG,QAAQ,CAAC,QAAQ,EAAE,MAAM,EAAE,CAAC,EAAE,WAAW,EAAE,KAAK,EAAE,eAAe,EAAE,KAAK,EAAE,MAAM,EAAE,sBAAsB,EAAE,CAAC;IAC5J,CAAC;IAED,MAAM,KAAK,GAAG,kBAAkB,CAAC;QAC/B,OAAO,EAAE,UAAU,CAAC,MAAM;QAC1B,MAAM,EAAE,UAAU,CAAC,QAAQ;QAC3B,OAAO,EAAE,QAAQ,CAAC,MAAM;QACxB,MAAM,EAAE,QAAQ,CAAC,QAAQ;KAC1B,CAAC,CAAC;IACH,MAAM,WAAW,GAAG,KAAK,CAAC,MAAM,GAAG,MAAM,CAAC,KAAK,CAAC;IAEhD,IAAI,KAAK,CAAC,KAAK,IAAI,CAAC,EAAE,CAAC;QACrB,OAAO,EAAE,GAAG,IAAI,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,EAAE,MAAM,EAAE,KAAK,CAAC,MAAM,EAAE,WAAW,EAAE,eAAe,EAAE,KAAK,EAAE,MAAM,EAAE,eAAe,EAAE,CAAC;IAC7H,CAAC;IACD,IAAI,CAAC,WAAW,EAAE,CAAC;QACjB,OAAO,EAAE,GAAG,IAAI,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,EAAE,MAAM,EAAE,KAAK,CAAC,MAAM,EAAE,WAAW,EAAE,eAAe,EAAE,KAAK,EAAE,MAAM,EAAE,iBAAiB,EAAE,CAAC;IAC/H,CAAC;IACD,OAAO,EAAE,GAAG,IAAI,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,EAAE,MAAM,EAAE,KAAK,CAAC,MAAM,EAAE,WAAW,EAAE,IAAI,EAAE,eAAe,EAAE,IAAI,EAAE,MAAM,EAAE,wBAAwB,EAAE,CAAC;AAC3I,CAAC;AAED,qFAAqF;AACrF,MAAM,UAAU,mBAAmB,CAAC,QAAgC;IAClE,OAAO,EAAE,GAAG,qBAAqB,EAAE,GAAG,CAAC,QAAQ,IAAI,EAAE,CAAC,EAAE,CAAC;AAC3D,CAAC"}

package/dist/learner/revert-stats.d.ts

@@ -0,0 +1,24 @@
+/**
+ * Pure statistics for the per-codeDigest revert decision (issue #764).
+ *
+ * Two-proportion z-test on pass/total for "arm A" (codeDigest WITH a candidate
+ * Improve commit) vs "arm B" (codeDigest AT the commit's parent). No I/O; unit-
+ * tested directly with hand-computed z-values. `delta = pA - pB` (negative means
+ * the commit made the pass rate worse).
+ */
+export interface TwoProportionInput {
+    passesA: number;
+    totalA: number;
+    passesB: number;
+    totalB: number;
+}
+export interface TwoProportionResult {
+    /** pA - pB (negative => arm A is worse). */
+    delta: number;
+    /** Test statistic; sign matches `delta`. 0 when either arm has no samples. */
+    z: number;
+    /** Two-sided p-value in [0, 1]. 1 when there is no signal. */
+    pValue: number;
+}
+/** Two-proportion z-test. Returns no-signal (z=0, p=1) if either total is 0. */
+export declare function twoProportionZTest(input: TwoProportionInput): TwoProportionResult;

package/dist/learner/revert-stats.js

@@ -0,0 +1,44 @@
+/**
+ * Pure statistics for the per-codeDigest revert decision (issue #764).
+ *
+ * Two-proportion z-test on pass/total for "arm A" (codeDigest WITH a candidate
+ * Improve commit) vs "arm B" (codeDigest AT the commit's parent). No I/O; unit-
+ * tested directly with hand-computed z-values. `delta = pA - pB` (negative means
+ * the commit made the pass rate worse).
+ */
+/** Two-proportion z-test. Returns no-signal (z=0, p=1) if either total is 0. */
+export function twoProportionZTest(input) {
+    const { passesA, totalA, passesB, totalB } = input;
+    if (totalA <= 0 || totalB <= 0) {
+        const pA = totalA > 0 ? passesA / totalA : 0;
+        const pB = totalB > 0 ? passesB / totalB : 0;
+        return { delta: pA - pB, z: 0, pValue: 1 };
+    }
+    const pA = passesA / totalA;
+    const pB = passesB / totalB;
+    const delta = pA - pB;
+    const pooled = (passesA + passesB) / (totalA + totalB);
+    const se = Math.sqrt(pooled * (1 - pooled) * (1 / totalA + 1 / totalB));
+    if (se === 0) {
+        // Both arms 0% or both 100% — no measurable difference.
+        return { delta, z: 0, pValue: 1 };
+    }
+    const z = delta / se;
+    const pValue = 2 * (1 - standardNormalCdf(Math.abs(z)));
+    return { delta, z, pValue };
+}
+/** Standard normal CDF via the Abramowitz & Stegun 7.1.26 erf approximation. */
+function standardNormalCdf(x) {
+    return 0.5 * (1 + erf(x / Math.SQRT2));
+}
+function erf(x) {
+    // Abramowitz & Stegun 7.1.26, max abs error ~1.5e-7.
+    const t = 1 / (1 + 0.3275911 * Math.abs(x));
+    const y = 1 -
+        ((((1.061405429 * t - 1.453152027) * t + 1.421413741) * t - 0.284496736) * t +
+            0.254829592) *
+            t *
+            Math.exp(-x * x);
+    return x >= 0 ? y : -y;
+}
+//# sourceMappingURL=revert-stats.js.map

package/dist/learner/revert-stats.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"revert-stats.js","sourceRoot":"","sources":["../../src/learner/revert-stats.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAkBH,gFAAgF;AAChF,MAAM,UAAU,kBAAkB,CAAC,KAAyB;IAC1D,MAAM,EAAE,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,EAAE,GAAG,KAAK,CAAC;IACnD,IAAI,MAAM,IAAI,CAAC,IAAI,MAAM,IAAI,CAAC,EAAE,CAAC;QAC/B,MAAM,EAAE,GAAG,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,OAAO,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC;QAC7C,MAAM,EAAE,GAAG,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,OAAO,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC;QAC7C,OAAO,EAAE,KAAK,EAAE,EAAE,GAAG,EAAE,EAAE,CAAC,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,CAAC;IAC7C,CAAC;IACD,MAAM,EAAE,GAAG,OAAO,GAAG,MAAM,CAAC;IAC5B,MAAM,EAAE,GAAG,OAAO,GAAG,MAAM,CAAC;IAC5B,MAAM,KAAK,GAAG,EAAE,GAAG,EAAE,CAAC;IACtB,MAAM,MAAM,GAAG,CAAC,OAAO,GAAG,OAAO,CAAC,GAAG,CAAC,MAAM,GAAG,MAAM,CAAC,CAAC;IACvD,MAAM,EAAE,GAAG,IAAI,CAAC,IAAI,CAAC,MAAM,GAAG,CAAC,CAAC,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC,GAAG,MAAM,GAAG,CAAC,GAAG,MAAM,CAAC,CAAC,CAAC;IACxE,IAAI,EAAE,KAAK,CAAC,EAAE,CAAC;QACb,wDAAwD;QACxD,OAAO,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,CAAC;IACpC,CAAC;IACD,MAAM,CAAC,GAAG,KAAK,GAAG,EAAE,CAAC;IACrB,MAAM,MAAM,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,iBAAiB,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IACxD,OAAO,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC;AAC9B,CAAC;AAED,gFAAgF;AAChF,SAAS,iBAAiB,CAAC,CAAS;IAClC,OAAO,GAAG,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC;AACzC,CAAC;AAED,SAAS,GAAG,CAAC,CAAS;IACpB,qDAAqD;IACrD,MAAM,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,SAAS,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;IAC5C,MAAM,CAAC,GACL,CAAC;QACD,CAAC,CAAC,CAAC,CAAC,WAAW,GAAG,CAAC,GAAG,WAAW,CAAC,GAAG,CAAC,GAAG,WAAW,CAAC,GAAG,CAAC,GAAG,WAAW,CAAC,GAAG,CAAC;YAC1E,WAAW,CAAC;YACZ,CAAC;YACD,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACrB,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;AACzB,CAAC"}