@hmharness/evolution 0.14.13 → 0.14.15

package/dist/judge.d.ts

@@ -43,6 +43,10 @@ export declare function runJudgeBatch(opts: {
     provider: ProviderConfig;
     modelTag?: string;
     log?: (line: string) => void;
+    /** re-judge sessions that carry a HUMAN label too - the 2026-09-21
+     *  direction: those stars were blind guesses, not ground truth; the
+     *  judge re-grades them and the judge label takes precedence. */
+    includeHumanLabeled?: boolean;
 }): Promise<{
     labeled: number;
     skipped: number;

package/dist/judge.js

@@ -107,7 +107,7 @@ export async function runJudgeBatch(opts) {
     const modelTag = opts.modelTag ?? provider.model;
     const human = await readLabels(home);
     const judged = await readJudgeLabels(home);
-    const done = new Set([...human.map((l) => l.session), ...judged.map((l) => l.session)]);
+    const done = new Set([...(opts.includeHumanLabeled ? [] : human.map((l) => l.session)), ...judged.map((l) => l.session)]);
     let insights = [];
     try {
         const text = await readFile(join(home, 'insights', 'insights.jsonl'), 'utf8');

package/dist/reward-model.d.ts

@@ -64,11 +64,19 @@ export declare function exportDpoPairs(labels: HumanLabel[], insights: InsightLi
     crossBucketPrefix?: number;
     sources?: Map<string, 'human' | 'judge'>;
 }): DpoPair[];
-/** Deterministic ~80/20 train/eval split by pair identity. Pure. */
+/** Deterministic ~80/20 split BY PROMPT (group split): bench templates
+ *  mint many pairs per prompt, and a pair-level split leaked ~4.9k eval
+ *  prompts into train (audit 2026-09-21) - every pair of one prompt must
+ *  land on the same side or offline metrics are inflated. Pure. */
 export declare function splitDpoPairs(pairs: DpoPair[], salt?: string): {
     train: DpoPair[];
     eval: DpoPair[];
 };
+/** Exact-duplicate removal (same prompt+chosen+rejected): the 5-star army
+ *  against the same few degraded sessions minted 11k+ permutation copies
+ *  (audit 2026-09-21) - training would just over-weight those templates.
+ *  Keeps the first occurrence (deterministic). Pure. */
+export declare function dedupeDpoPairs(pairs: DpoPair[]): DpoPair[];
 /** Leakage/quality audit before training consumes the pairs: empty sides,
  *  exact duplicates, and eval pairs whose prompt ALSO appears in train
  *  (prompt leakage inflates offline metrics). Pure. */

package/dist/reward-model.js

@@ -202,15 +202,34 @@ function fnv1a(str) {
     }
     return h >>> 0;
 }
-/** Deterministic ~80/20 train/eval split by pair identity. Pure. */
+/** Deterministic ~80/20 split BY PROMPT (group split): bench templates
+ *  mint many pairs per prompt, and a pair-level split leaked ~4.9k eval
+ *  prompts into train (audit 2026-09-21) - every pair of one prompt must
+ *  land on the same side or offline metrics are inflated. Pure. */
 export function splitDpoPairs(pairs, salt = 'dpo') {
     const train = [];
     const evalP = [];
     for (const p of pairs) {
-        (fnv1a(salt + '|' + p.prompt + '|' + p.chosenSession + '|' + p.rejectedSession) % 5 === 0 ? evalP : train).push(p);
+        (fnv1a(salt + '|' + p.prompt) % 5 === 0 ? evalP : train).push(p);
     }
     return { train, eval: evalP };
 }
+/** Exact-duplicate removal (same prompt+chosen+rejected): the 5-star army
+ *  against the same few degraded sessions minted 11k+ permutation copies
+ *  (audit 2026-09-21) - training would just over-weight those templates.
+ *  Keeps the first occurrence (deterministic). Pure. */
+export function dedupeDpoPairs(pairs) {
+    const seen = new Set();
+    const out = [];
+    for (const p of pairs) {
+        const key = p.prompt + '\u0000' + p.chosen + '\u0000' + p.rejected;
+        if (seen.has(key))
+            continue;
+        seen.add(key);
+        out.push(p);
+    }
+    return out;
+}
 /** Leakage/quality audit before training consumes the pairs: empty sides,
  *  exact duplicates, and eval pairs whose prompt ALSO appears in train
  *  (prompt leakage inflates offline metrics). Pure. */

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hmharness/evolution",
-  "version": "0.14.13",
+  "version": "0.14.15",
   "description": "hmharness evolution subsystem: persistent memory, insight capture, skill library, and the bench that gives evolution its fitness signal. First-class kernel citizen, not a plugin.",
   "type": "module",
   "main": "dist/index.js",