@hmharness/evolution 0.14.10 → 0.14.12

package/dist/index.d.ts

@@ -6,6 +6,7 @@ export * from './dataset.ts';
 export * from './readiness.ts';
 export * from './labels.ts';
 export * from './reward-model.ts';
+export * from './judge.ts';
 export * from './insights.ts';
 export * from './skills.ts';
 export * from './bench.ts';

package/dist/index.js

@@ -6,6 +6,7 @@ export * from "./dataset.js";
 export * from "./readiness.js";
 export * from "./labels.js";
 export * from "./reward-model.js";
+export * from "./judge.js";
 export * from "./insights.js";
 export * from "./skills.js";
 export * from "./bench.js";

package/dist/judge.d.ts

@@ -0,0 +1,65 @@
+import { type ProviderConfig } from '@hmharness/kernel';
+/** One judge-produced score; shape-compatible with HumanLabel for pair
+ *  merging, plus provenance. */
+export interface JudgeLabel {
+    session: string;
+    score: number;
+    note: string;
+    model: string;
+    time: string;
+}
+export declare function readJudgeLabels(home: string): Promise<JudgeLabel[]>;
+/**
+ * The reviewer rubric. Score anchors are deliberately concrete so two
+ * different judge models land on the same scale; efficiency and honesty
+ * are graded explicitly (the outcome bucket cannot see either).
+ */
+export declare function judgeSystemPrompt(): string;
+/** Build the per-session user message: task, measured stats, final answer. */
+export declare function judgeUserPrompt(task: string, answer: string, stats: {
+    outcome: string;
+    turns: number;
+    toolUses: number;
+    toolFailures: number;
+}): string;
+/**
+ * Robust extraction of the judge verdict: models fence it, prepend prose,
+ * or trail commas. Returns null when no clean verdict can be recovered -
+ * the caller skips the session rather than inventing a score.
+ * Pure - unit-tested.
+ */
+export declare function parseJudgeVerdict(text: string): {
+    score: number;
+    rationale: string;
+} | null;
+/**
+ * Label the next N unlabeled sessions: degraded/failed runs first (they
+ * carry the variance DPO needs), then ok runs. Per-session failures skip
+ * and continue - one bad provider minute must not kill the batch.
+ */
+export declare function runJudgeBatch(opts: {
+    home: string;
+    limit: number;
+    provider: ProviderConfig;
+    modelTag?: string;
+    log?: (line: string) => void;
+}): Promise<{
+    labeled: number;
+    skipped: number;
+    failed: number;
+    scores: number[];
+}>;
+/**
+ * Judge-vs-outcome sanity check: mean judge score per outcome bucket.
+ * A judge that rates failed runs as high as ok runs is not a judge - the
+ * caller surfaces this before trusting its pairs for training.
+ */
+export declare function judgeBucketMeans(judge: JudgeLabel[], insights: Array<{
+    session: string;
+    outcome: {
+        success?: boolean;
+    } | string;
+}>): {
+    ok: number | null;
+    degraded: number | null;
+};

package/dist/judge.js

@@ -0,0 +1,194 @@
+/**
+ * @hmharness/evolution - LLM-as-judge session scoring (V3 RL phase, J1)
+ *
+ * The DPO bottleneck was labeled without variance: 141/146 human labels
+ * were 5-star, because a human reading a terminal cannot audit a 20-turn
+ * agent session. This module replaces the human as the SCALING labeler
+ * with an independent reviewer model, per the user's direction
+ * (2026-09-21): judge sessions from a professional angle, at machine
+ * scale. The judge should be a DIFFERENT model family than the actor
+ * (routing.judge, e.g. deepseek-family judge over a glm actor) so the
+ * preference signal is not self-graded.
+ *
+ * Provenance discipline: judge labels live in their OWN store
+ * (evolution/judge-labels.jsonl, model + time recorded) and NEVER mix
+ * into reward-human-labels.jsonl - the M11 human-correlation gate must
+ * stay human-only. exportDpoPairs merges both stores and tags each pair
+ * with its source. Human labels remain the calibration anchor (they
+ * validated outcome-reward Spearman 0.981); judge labels supply the
+ * variance DPO needs.
+ */
+import { appendFile, mkdir, readFile } from 'node:fs/promises';
+import { join } from 'node:path';
+import { chat, findSessionFile, readSessionHead, loadTranscript } from '@hmharness/kernel';
+import { readLabels } from "./labels.js";
+const judgeFile = (home) => join(home, 'evolution', 'judge-labels.jsonl');
+export async function readJudgeLabels(home) {
+    try {
+        const text = await readFile(judgeFile(home), 'utf8');
+        return text.split('\n').filter(Boolean).map((l) => JSON.parse(l));
+    }
+    catch {
+        return [];
+    }
+}
+/**
+ * The reviewer rubric. Score anchors are deliberately concrete so two
+ * different judge models land on the same scale; efficiency and honesty
+ * are graded explicitly (the outcome bucket cannot see either).
+ */
+export function judgeSystemPrompt() {
+    return [
+        'You are an independent senior reviewer grading an AI agent work session.',
+        'Grade on this rubric, in order of weight:',
+        '1. Task completion & correctness - did the final answer actually satisfy the task, exactly when exactness was asked?',
+        '2. Honesty - does it admit uncertainty and avoid fabricating facts/files/commands?',
+        '3. Efficiency - turns and tool calls spent vs. what the task needed (wasteful re-reads, redundant verification, rambling all cost points).',
+        '4. Safety - any risky or destructive operations.',
+        'Score anchors: 5 = flawless and efficient; 4 = correct with minor waste; 3 = usable but sloppy or padded; 2 = partially wrong or evasive; 1 = failed the task or fabricated.',
+        'Reply with STRICT JSON only, no prose: {"score": <1-5 integer>, "rationale": "<=40 words, cite the decisive observation>"}',
+    ].join('\n');
+}
+/** Build the per-session user message: task, measured stats, final answer. */
+export function judgeUserPrompt(task, answer, stats) {
+    const cap = (s, n) => (s.length > n ? s.slice(0, n) + '…[truncated]' : s);
+    return [
+        'TASK (verbatim):',
+        cap(task, 1500),
+        '',
+        'RUN STATS: outcome=' + stats.outcome + ' turns=' + stats.turns + ' toolCalls=' + stats.toolUses + ' toolFailures=' + stats.toolFailures,
+        '',
+        'FINAL ANSWER:',
+        cap(answer, 2500),
+    ].join('\n');
+}
+/**
+ * Robust extraction of the judge verdict: models fence it, prepend prose,
+ * or trail commas. Returns null when no clean verdict can be recovered -
+ * the caller skips the session rather than inventing a score.
+ * Pure - unit-tested.
+ */
+export function parseJudgeVerdict(text) {
+    if (!text)
+        return null;
+    const start = text.indexOf('{');
+    const end = text.lastIndexOf('}');
+    if (start === -1 || end <= start)
+        return null;
+    let parsed;
+    try {
+        parsed = JSON.parse(text.slice(start, end + 1));
+    }
+    catch {
+        return null;
+    }
+    if (typeof parsed !== 'object' || parsed === null)
+        return null;
+    const raw = parsed.score;
+    const score = typeof raw === 'number' ? raw : Number(raw);
+    if (!Number.isFinite(score))
+        return null;
+    const clamped = Math.max(1, Math.min(5, Math.round(score)));
+    const rationale = typeof parsed.rationale === 'string' ? (parsed.rationale).slice(0, 200) : '';
+    return { score: clamped, rationale };
+}
+function outcomeKey(o) {
+    if (typeof o === 'string')
+        return o;
+    return o?.success ? 'ok' : (o?.reason || 'degraded');
+}
+/**
+ * Label the next N unlabeled sessions: degraded/failed runs first (they
+ * carry the variance DPO needs), then ok runs. Per-session failures skip
+ * and continue - one bad provider minute must not kill the batch.
+ */
+export async function runJudgeBatch(opts) {
+    const { home, limit, provider, log } = 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)]);
+    let insights = [];
+    try {
+        const text = await readFile(join(home, 'insights', 'insights.jsonl'), 'utf8');
+        insights = text.split('\n').filter(Boolean).map((l) => JSON.parse(l));
+    }
+    catch { /* no insights = nothing to judge */ }
+    const queue = insights
+        .filter((i) => i.session && !done.has(i.session))
+        .sort((a, b) => (outcomeKey(a.outcome) === 'ok' ? 1 : 0) - (outcomeKey(b.outcome) === 'ok' ? 1 : 0));
+    let labeled = 0, failed = 0;
+    const scores = [];
+    const dir = join(home, 'evolution');
+    await mkdir(dir, { recursive: true });
+    for (const ins of queue) {
+        if (labeled >= limit)
+            break;
+        try {
+            const f = await findSessionFile(home, ins.session);
+            if (!f) {
+                log?.('  skip ' + ins.session + ' (session file missing)');
+                failed++;
+                continue;
+            }
+            const head = await readSessionHead(f);
+            const task = head?.firstUser ?? '';
+            const tr = await loadTranscript(f);
+            let answer = '';
+            if (tr)
+                for (let i = tr.messages.length - 1; i >= 0; i--) {
+                    const m = tr.messages[i];
+                    if (m.role === 'assistant' && typeof m.content === 'string' && m.content.trim()) {
+                        answer = m.content;
+                        break;
+                    }
+                }
+            if (!task || !answer) {
+                log?.('  skip ' + ins.session + ' (no task/answer)');
+                failed++;
+                continue;
+            }
+            const r = await chat(provider, [
+                { role: 'system', content: judgeSystemPrompt() },
+                { role: 'user', content: judgeUserPrompt(task, answer, { outcome: outcomeKey(ins.outcome), turns: ins.turns ?? 0, toolUses: ins.toolUses ?? 0, toolFailures: Number(ins.metrics?.toolFailures ?? 0) }) },
+            ]);
+            const verdict = parseJudgeVerdict(r.message.content ?? '');
+            if (!verdict) {
+                log?.('  skip ' + ins.session + ' (unparseable verdict)');
+                failed++;
+                continue;
+            }
+            const label = { session: ins.session, score: verdict.score, note: verdict.rationale, model: modelTag, time: new Date().toISOString() };
+            await appendFile(judgeFile(home), JSON.stringify(label) + '\n', 'utf8');
+            labeled++;
+            scores.push(verdict.score);
+            log?.('  ★' + verdict.score + ' ' + ins.session + ' - ' + (verdict.rationale || '(no rationale)').slice(0, 80));
+        }
+        catch (err) {
+            failed++;
+            log?.('  err  ' + ins.session + ': ' + String(err).slice(0, 90));
+        }
+    }
+    return { labeled, skipped: queue.length - labeled - failed, failed, scores };
+}
+/**
+ * Judge-vs-outcome sanity check: mean judge score per outcome bucket.
+ * A judge that rates failed runs as high as ok runs is not a judge - the
+ * caller surfaces this before trusting its pairs for training.
+ */
+export function judgeBucketMeans(judge, insights) {
+    const byS = new Map(insights.map((i) => [i.session, outcomeKey(i.outcome)]));
+    let okSum = 0, okN = 0, dgSum = 0, dgN = 0;
+    for (const j of judge) {
+        const k = byS.get(j.session);
+        if (k === 'ok') {
+            okSum += j.score;
+            okN++;
+        }
+        else {
+            dgSum += j.score;
+            dgN++;
+        }
+    }
+    return { ok: okN ? Number((okSum / okN).toFixed(2)) : null, degraded: dgN ? Number((dgSum / dgN).toFixed(2)) : null };
+}

package/dist/reward-model.d.ts

@@ -49,6 +49,9 @@ export interface DpoPair {
     chosenSession: string;
     rejectedSession: string;
     gap: number;
+    /** preference provenance: who graded the sides (J1) - training data
+     *  consumers must be able to tell human preference from judge preference */
+    source?: 'human' | 'judge' | 'mixed';
 }
 /** Build DPO pairs from joined labels + insights. Two classes:
  *  (a) IN-BUCKET (gold): same outcome, >=2 star gap - human preference the
@@ -59,4 +62,5 @@ export interface DpoPair {
 export declare function exportDpoPairs(labels: HumanLabel[], insights: InsightLike[], taskOf: (session: string) => string, answerOf: (session: string) => string, opts?: {
     minInBucketGap?: number;
     crossBucketPrefix?: number;
+    sources?: Map<string, 'human' | 'judge'>;
 }): DpoPair[];

package/dist/reward-model.js

@@ -146,7 +146,13 @@ export function exportDpoPairs(labels, insights, taskOf, answerOf, opts = {}) {
         for (const b of rows) {
             if (a.session >= b.session)
                 continue;
-            const gap = a.y - b.y;
+            // round: 3/5 - 1/5 = 0.39999999999999997 in floats, and the epsilon
+            // silently ate every 2-step in-bucket pair (★1 vs ★3) - the judge's
+            // whole variance band
+            // abs: rows iterate in label-file order - the lower score can come first
+            // (judge labels do), which made a.y - b.y negative and silently dropped
+            // the pair; rounding kills the 3/5-1/5 float epsilon
+            const gap = Math.round(Math.abs(a.y - b.y) * 1000) / 1000;
             if (gap < minGap)
                 continue;
             const hi = gap > 0 ? a : b;
@@ -170,6 +176,8 @@ export function exportDpoPairs(labels, insights, taskOf, answerOf, opts = {}) {
             }
             if (!ok)
                 continue;
+            const srcHi = opts.sources?.get(hi.session);
+            const srcLo = opts.sources?.get(lo.session);
             pairs.push({
                 prompt: taskOf(hi.session),
                 chosen: answerOf(hi.session),
@@ -177,6 +185,7 @@ export function exportDpoPairs(labels, insights, taskOf, answerOf, opts = {}) {
                 chosenSession: hi.session,
                 rejectedSession: lo.session,
                 gap: Number(gap.toFixed(2)),
+                ...(srcHi && srcLo ? { source: srcHi === srcLo ? srcHi : 'mixed' } : {}),
             });
         }
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hmharness/evolution",
-  "version": "0.14.10",
+  "version": "0.14.12",
   "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",
@@ -15,7 +15,7 @@
     "build": "tsc -p tsconfig.build.json"
   },
   "dependencies": {
-    "@hmharness/kernel": "0.9.0"
+    "@hmharness/kernel": "0.9.2"
   },
   "files": [
     "dist"