@nexus-cortex/core 4.26.0

package/dist/training/ResearchBacklog.js

@@ -0,0 +1,149 @@
+/**
+ * ResearchBacklog — the recursive auto-research task lifecycle.
+ *
+ * Every harness deficiency found during benchmarking is auto-added here as a
+ * tracked, triaged task. The harness identifies its own weaknesses, prioritizes
+ * them (severity × impact × confidence ÷ effort), and walks each through a
+ * lifecycle: open → triaged → in_progress → fixed → verified → closed.
+ *
+ * OVERFITTING GUARD (first-class — see operator note "be wary of overfitting as
+ * recursion progresses"): a deficiency is NOT `verified` just because the fix
+ * passes the task that SURFACED it — that is only `fixed`. It becomes `verified`
+ * only after the fix is confirmed on HELD-OUT tasks it was never tuned against.
+ * The `fixed`→`verified` gap is the guard against the harness gaming its own eval.
+ * (Pair with ModelRouterMatrix's `split:'holdout'` records — keep/discard uses
+ * `train`, verification uses `holdout`.)
+ *
+ * Store: append-only JSONL at `.cortex/research-backlog.jsonl`. Concurrent-safe
+ * for two agents sharing the tree — latest snapshot per `id` wins; never
+ * rewritten in place.
+ */
+import * as fs from 'fs';
+import * as path from 'path';
+import * as crypto from 'crypto';
+const SEVERITY_WEIGHT = { low: 1, medium: 2, high: 3, critical: 5 };
+/**
+ * Triage priority: severity-weighted value per unit effort, scaled by confidence.
+ * `(severityWeight × impact × confidence) / effort`. Higher = fix first.
+ * Confidence scaling is the overfitting/ noise guard at the triage layer —
+ * low-confidence "deficiencies" (could be model noise) sink in priority.
+ */
+export function computePriority(severity, impact, effort, confidence) {
+    const sw = SEVERITY_WEIGHT[severity] ?? 1;
+    const e = Math.max(1, effort);
+    const c = Math.min(1, Math.max(0, confidence));
+    return Math.round(((sw * impact * c) / e) * 100) / 100;
+}
+const STORE_RELATIVE_PATH = path.join('.cortex', 'research-backlog.jsonl');
+export class ResearchBacklog {
+    storePath;
+    constructor(projectRoot) {
+        this.storePath = path.join(projectRoot, STORE_RELATIVE_PATH);
+    }
+    /** Stable id derived from the title, so re-finding the same deficiency updates
+     *  rather than duplicating. */
+    static idFor(title) {
+        return 'def-' + crypto.createHash('sha256').update(title.trim().toLowerCase(), 'utf8').digest('hex').slice(0, 10);
+    }
+    append(rec) {
+        fs.mkdirSync(path.dirname(this.storePath), { recursive: true });
+        fs.appendFileSync(this.storePath, JSON.stringify(rec) + '\n', 'utf8');
+    }
+    /** Current state = latest snapshot per id (append-only; last write wins). */
+    list(opts = {}) {
+        let raw = '';
+        try {
+            raw = fs.readFileSync(this.storePath, 'utf8');
+        }
+        catch {
+            return [];
+        }
+        const byId = new Map();
+        for (const line of raw.split('\n')) {
+            if (!line.trim())
+                continue;
+            try {
+                const r = JSON.parse(line);
+                byId.set(r.id, r);
+            }
+            catch { /* skip corrupt line */ }
+        }
+        let out = [...byId.values()];
+        if (opts.status) {
+            const want = Array.isArray(opts.status) ? opts.status : [opts.status];
+            out = out.filter(r => want.includes(r.status));
+        }
+        if (opts.sortByPriority !== false)
+            out.sort((a, b) => b.priorityScore - a.priorityScore);
+        return out;
+    }
+    get(id) {
+        return this.list({ sortByPriority: false }).find(r => r.id === id);
+    }
+    /** Auto-add a found deficiency. Idempotent by title; triage runs automatically
+     *  so `priorityScore` is set on add. */
+    add(d) {
+        const id = ResearchBacklog.idFor(d.title);
+        const existing = this.get(id);
+        const now = new Date().toISOString();
+        const severity = d.severity ?? existing?.severity ?? 'medium';
+        const impact = d.impact ?? existing?.impact ?? 3;
+        const effort = d.effort ?? existing?.effort ?? 3;
+        const confidence = d.confidence ?? existing?.confidence ?? 0.6;
+        const triaged = d.severity !== undefined || d.impact !== undefined || d.effort !== undefined;
+        const rec = {
+            id,
+            title: d.title,
+            description: d.description,
+            bugClass: d.bugClass ?? existing?.bugClass ?? 'Other',
+            // never downgrade an already-progressing item back to 'open' on re-find
+            status: existing && existing.status !== 'open' ? existing.status : (triaged ? 'triaged' : 'open'),
+            severity, impact, effort, confidence,
+            priorityScore: computePriority(severity, impact, effort, confidence),
+            discoveredRound: d.discoveredRound ?? existing?.discoveredRound,
+            discoveredRef: d.discoveredRef ?? existing?.discoveredRef,
+            affectedModels: d.affectedModels ?? existing?.affectedModels,
+            affectedTaskFingerprints: d.affectedTaskFingerprints ?? existing?.affectedTaskFingerprints,
+            experimentTag: existing?.experimentTag,
+            fixedRef: existing?.fixedRef,
+            verifiedRound: existing?.verifiedRound,
+            notes: d.notes ?? existing?.notes,
+            createdAt: existing?.createdAt ?? now,
+            updatedAt: now,
+        };
+        this.append(rec);
+        return rec;
+    }
+    /** Patch fields + re-snapshot. Recomputes priority when triage fields change. */
+    update(id, patch) {
+        const cur = this.get(id);
+        if (!cur)
+            return undefined;
+        const merged = { ...cur, ...patch, id, updatedAt: new Date().toISOString() };
+        merged.priorityScore = computePriority(merged.severity, merged.impact, merged.effort, merged.confidence);
+        this.append(merged);
+        return merged;
+    }
+    triage(id, t) {
+        return this.update(id, { ...t, status: 'triaged' });
+    }
+    /** Highest-priority actionable item (open/triaged) — what the recursion fixes next. */
+    next() {
+        return this.list({ status: ['open', 'triaged'], sortByPriority: true })[0];
+    }
+    markInProgress(id, experimentTag) {
+        return this.update(id, { status: 'in_progress', experimentTag });
+    }
+    /** Fix passes the DISCOVERY task — but NOT yet generalized. Do not close here. */
+    markFixed(id, fixedRef) {
+        return this.update(id, { status: 'fixed', fixedRef });
+    }
+    /** OVERFITTING GUARD: only call after the fix holds on HELD-OUT tasks it was not tuned against. */
+    markVerified(id, verifiedRound) {
+        return this.update(id, { status: 'verified', verifiedRound });
+    }
+    close(id) { return this.update(id, { status: 'closed' }); }
+    wontFix(id, reason) { return this.update(id, { status: 'wont_fix', notes: reason }); }
+    reopenRegressed(id) { return this.update(id, { status: 'regressed' }); }
+}
+//# sourceMappingURL=ResearchBacklog.js.map

package/dist/training/ResearchBacklog.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"ResearchBacklog.js","sourceRoot":"","sources":["../../src/training/ResearchBacklog.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAEH,OAAO,KAAK,EAAE,MAAM,IAAI,CAAC;AACzB,OAAO,KAAK,IAAI,MAAM,MAAM,CAAC;AAC7B,OAAO,KAAK,MAAM,MAAM,QAAQ,CAAC;AAwCjC,MAAM,eAAe,GAA6B,EAAE,GAAG,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,IAAI,EAAE,CAAC,EAAE,QAAQ,EAAE,CAAC,EAAE,CAAC;AAE9F;;;;;GAKG;AACH,MAAM,UAAU,eAAe,CAAC,QAAkB,EAAE,MAAc,EAAE,MAAc,EAAE,UAAkB;IACpG,MAAM,EAAE,GAAG,eAAe,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC;IAC1C,MAAM,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC;IAC9B,MAAM,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,UAAU,CAAC,CAAC,CAAC;IAC/C,OAAO,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,GAAG,MAAM,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC,GAAG,GAAG,CAAC;AACzD,CAAC;AAiBD,MAAM,mBAAmB,GAAG,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,wBAAwB,CAAC,CAAC;AAE3E,MAAM,OAAO,eAAe;IACT,SAAS,CAAS;IAEnC,YAAY,WAAmB;QAC7B,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,IAAI,CAAC,WAAW,EAAE,mBAAmB,CAAC,CAAC;IAC/D,CAAC;IAED;mCAC+B;IAC/B,MAAM,CAAC,KAAK,CAAC,KAAa;QACxB,OAAO,MAAM,GAAG,MAAM,CAAC,UAAU,CAAC,QAAQ,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,EAAE,MAAM,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;IACpH,CAAC;IAEO,MAAM,CAAC,GAAqB;QAClC,EAAE,CAAC,SAAS,CAAC,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,SAAS,CAAC,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;QAChE,EAAE,CAAC,cAAc,CAAC,IAAI,CAAC,SAAS,EAAE,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,GAAG,IAAI,EAAE,MAAM,CAAC,CAAC;IACxE,CAAC;IAED,6EAA6E;IAC7E,IAAI,CAAC,OAAqF,EAAE;QAC1F,IAAI,GAAG,GAAG,EAAE,CAAC;QACb,IAAI,CAAC;YAAC,GAAG,GAAG,EAAE,CAAC,YAAY,CAAC,IAAI,CAAC,SAAS,EAAE,MAAM,CAAC,CAAC;QAAC,CAAC;QAAC,MAAM,CAAC;YAAC,OAAO,EAAE,CAAC;QAAC,CAAC;QAC3E,MAAM,IAAI,GAAG,IAAI,GAAG,EAA4B,CAAC;QACjD,KAAK,MAAM,IAAI,IAAI,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC;YACnC,IAAI,CAAC,IAAI,CAAC,IAAI,EAAE;gBAAE,SAAS;YAC3B,IAAI,CAAC;gBAAC,MAAM,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAqB,CAAC;gBAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC;YAAC,CAAC;YAAC,MAAM,CAAC,CAAC,uBAAuB,CAAC,CAAC;QAC9G,CAAC;QACD,IAAI,GAAG,GAAG,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC;QAC7B,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC;YAChB,MAAM,IAAI,GAAG,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;YACtE,GAAG,GAAG,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC;QACjD,CAAC;QACD,IAAI,IAAI,CAAC,cAAc,KAAK,KAAK;YAAE,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,aAAa,GAAG,CAAC,CAAC,aAAa,CAAC,CAAC;QACzF,OAAO,GAAG,CAAC;IACb,CAAC;IAED,GAAG,CAAC,EAAU;QACZ,OAAO,IAAI,CAAC,IAAI,CAAC,EAAE,cAAc,EAAE,KAAK,EAAE,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC;IACrE,CAAC;IAED;4CACwC;IACxC,GAAG,CAAC,CAAgB;QAClB,MAAM,EAAE,GAAG,eAAe,CAAC,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC;QAC1C,MAAM,QAAQ,GAAG,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;QAC9B,MAAM,GAAG,GAAG,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;QACrC,MAAM,QAAQ,GAAG,CAAC,CAAC,QAAQ,IAAI,QAAQ,EAAE,QAAQ,IAAI,QAAQ,CAAC;QAC9D,MAAM,MAAM,GAAG,CAAC,CAAC,MAAM,IAAI,QAAQ,EAAE,MAAM,IAAI,CAAC,CAAC;QACjD,MAAM,MAAM,GAAG,CAAC,CAAC,MAAM,IAAI,QAAQ,EAAE,MAAM,IAAI,CAAC,CAAC;QACjD,MAAM,UAAU,GAAG,CAAC,CAAC,UAAU,IAAI,QAAQ,EAAE,UAAU,IAAI,GAAG,CAAC;QAC/D,MAAM,OAAO,GAAG,CAAC,CAAC,QAAQ,KAAK,SAAS,IAAI,CAAC,CAAC,MAAM,KAAK,SAAS,IAAI,CAAC,CAAC,MAAM,KAAK,SAAS,CAAC;QAC7F,MAAM,GAAG,GAAqB;YAC5B,EAAE;YACF,KAAK,EAAE,CAAC,CAAC,KAAK;YACd,WAAW,EAAE,CAAC,CAAC,WAAW;YAC1B,QAAQ,EAAE,CAAC,CAAC,QAAQ,IAAI,QAAQ,EAAE,QAAQ,IAAI,OAAO;YACrD,wEAAwE;YACxE,MAAM,EAAE,QAAQ,IAAI,QAAQ,CAAC,MAAM,KAAK,MAAM,CAAC,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,MAAM,CAAC;YACjG,QAAQ,EAAE,MAAM,EAAE,MAAM,EAAE,UAAU;YACpC,aAAa,EAAE,eAAe,CAAC,QAAQ,EAAE,MAAM,EAAE,MAAM,EAAE,UAAU,CAAC;YACpE,eAAe,EAAE,CAAC,CAAC,eAAe,IAAI,QAAQ,EAAE,eAAe;YAC/D,aAAa,EAAE,CAAC,CAAC,aAAa,IAAI,QAAQ,EAAE,aAAa;YACzD,cAAc,EAAE,CAAC,CAAC,cAAc,IAAI,QAAQ,EAAE,cAAc;YAC5D,wBAAwB,EAAE,CAAC,CAAC,wBAAwB,IAAI,QAAQ,EAAE,wBAAwB;YAC1F,aAAa,EAAE,QAAQ,EAAE,aAAa;YACtC,QAAQ,EAAE,QAAQ,EAAE,QAAQ;YAC5B,aAAa,EAAE,QAAQ,EAAE,aAAa;YACtC,KAAK,EAAE,CAAC,CAAC,KAAK,IAAI,QAAQ,EAAE,KAAK;YACjC,SAAS,EAAE,QAAQ,EAAE,SAAS,IAAI,GAAG;YACrC,SAAS,EAAE,GAAG;SACf,CAAC;QACF,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;QACjB,OAAO,GAAG,CAAC;IACb,CAAC;IAED,iFAAiF;IACjF,MAAM,CAAC,EAAU,EAAE,KAAgC;QACjD,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;QACzB,IAAI,CAAC,GAAG;YAAE,OAAO,SAAS,CAAC;QAC3B,MAAM,MAAM,GAAqB,EAAE,GAAG,GAAG,EAAE,GAAG,KAAK,EAAE,EAAE,EAAE,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,EAAE,CAAC;QAC/F,MAAM,CAAC,aAAa,GAAG,eAAe,CAAC,MAAM,CAAC,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,UAAU,CAAC,CAAC;QACzG,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;QACpB,OAAO,MAAM,CAAC;IAChB,CAAC;IAED,MAAM,CAAC,EAAU,EAAE,CAAiF;QAClG,OAAO,IAAI,CAAC,MAAM,CAAC,EAAE,EAAE,EAAE,GAAG,CAAC,EAAE,MAAM,EAAE,SAAS,EAAE,CAAC,CAAC;IACtD,CAAC;IAED,uFAAuF;IACvF,IAAI;QACF,OAAO,IAAI,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,CAAC,MAAM,EAAE,SAAS,CAAC,EAAE,cAAc,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;IAC7E,CAAC;IAED,cAAc,CAAC,EAAU,EAAE,aAAqB;QAC9C,OAAO,IAAI,CAAC,MAAM,CAAC,EAAE,EAAE,EAAE,MAAM,EAAE,aAAa,EAAE,aAAa,EAAE,CAAC,CAAC;IACnE,CAAC;IAED,kFAAkF;IAClF,SAAS,CAAC,EAAU,EAAE,QAAgB;QACpC,OAAO,IAAI,CAAC,MAAM,CAAC,EAAE,EAAE,EAAE,MAAM,EAAE,OAAO,EAAE,QAAQ,EAAE,CAAC,CAAC;IACxD,CAAC;IAED,mGAAmG;IACnG,YAAY,CAAC,EAAU,EAAE,aAAqB;QAC5C,OAAO,IAAI,CAAC,MAAM,CAAC,EAAE,EAAE,EAAE,MAAM,EAAE,UAAU,EAAE,aAAa,EAAE,CAAC,CAAC;IAChE,CAAC;IAED,KAAK,CAAC,EAAU,IAAkC,OAAO,IAAI,CAAC,MAAM,CAAC,EAAE,EAAE,EAAE,MAAM,EAAE,QAAQ,EAAE,CAAC,CAAC,CAAC,CAAC;IACjG,OAAO,CAAC,EAAU,EAAE,MAAc,IAAkC,OAAO,IAAI,CAAC,MAAM,CAAC,EAAE,EAAE,EAAE,MAAM,EAAE,UAAU,EAAE,KAAK,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,CAAC;IACpI,eAAe,CAAC,EAAU,IAAkC,OAAO,IAAI,CAAC,MAAM,CAAC,EAAE,EAAE,EAAE,MAAM,EAAE,WAAW,EAAE,CAAC,CAAC,CAAC,CAAC;CAC/G"}

package/dist/training/TaskClassifier.d.ts

@@ -0,0 +1,16 @@
+/**
+ * TaskClassifier — heuristic task-type detector for model routing.
+ *
+ * Classifies a user prompt into one of the benchmark task types (T1-T5)
+ * so the ModelRouterMatrix can recommend the best model. Classification
+ * is keyword-based and intentionally simple — a 90% heuristic that can
+ * be replaced with an LLM classifier later without changing the surface.
+ */
+export type TaskType = 'T1' | 'T2' | 'T3' | 'T4' | 'T5' | 'GENERAL';
+interface ClassificationResult {
+    taskType: TaskType;
+    confidence: number;
+}
+export declare function classifyTask(prompt: string): ClassificationResult;
+export {};
+//# sourceMappingURL=TaskClassifier.d.ts.map

package/dist/training/TaskClassifier.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"TaskClassifier.d.ts","sourceRoot":"","sources":["../../src/training/TaskClassifier.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,MAAM,MAAM,QAAQ,GAAG,IAAI,GAAG,IAAI,GAAG,IAAI,GAAG,IAAI,GAAG,IAAI,GAAG,SAAS,CAAC;AAEpE,UAAU,oBAAoB;IAC5B,QAAQ,EAAE,QAAQ,CAAC;IACnB,UAAU,EAAE,MAAM,CAAC;CACpB;AAyCD,wBAAgB,YAAY,CAAC,MAAM,EAAE,MAAM,GAAG,oBAAoB,CAuBjE"}

package/dist/training/TaskClassifier.js

@@ -0,0 +1,70 @@
+/**
+ * TaskClassifier — heuristic task-type detector for model routing.
+ *
+ * Classifies a user prompt into one of the benchmark task types (T1-T5)
+ * so the ModelRouterMatrix can recommend the best model. Classification
+ * is keyword-based and intentionally simple — a 90% heuristic that can
+ * be replaced with an LLM classifier later without changing the surface.
+ */
+const PATTERNS = [
+    {
+        taskType: 'T1',
+        keywords: [
+            /\b(grep|search|find|locate|where is|which file|defined in|look for)\b/i,
+            /\b(read|show me|what does .+ look like|open|cat)\b/i,
+            /\b(codebase|source code|implementation|function|class|method)\b/i,
+        ],
+        weight: 1.0,
+    },
+    {
+        taskType: 'T3',
+        keywords: [
+            /\b(list|discover|available|what .+ exist|show .+ tools|capabilities)\b/i,
+            /\b(session|history|previous|past conversation)\b/i,
+            /\b(what can you|how many|enumerate)\b/i,
+        ],
+        weight: 1.0,
+    },
+    {
+        taskType: 'T4',
+        keywords: [
+            /\b(audit|analyze|review|assess|evaluate|architecture)\b/i,
+            /\b(security|middleware|pipeline|flow|dependency|coupling)\b/i,
+            /\b(how does .+ work|explain .+ system|trace .+ through)\b/i,
+        ],
+        weight: 1.1,
+    },
+    {
+        taskType: 'T5',
+        keywords: [
+            /\b(create|build|implement|write|generate|design|make)\b/i,
+            /\b(new (tool|function|component|module|page|feature))\b/i,
+            /\b(landing page|dashboard|CLI command|UI)\b/i,
+        ],
+        weight: 1.2,
+    },
+];
+export function classifyTask(prompt) {
+    const text = typeof prompt === 'string' ? prompt : JSON.stringify(prompt);
+    let bestType = 'GENERAL';
+    let bestScore = 0;
+    for (const pattern of PATTERNS) {
+        let hits = 0;
+        for (const kw of pattern.keywords) {
+            if (kw.test(text))
+                hits++;
+        }
+        const score = (hits / pattern.keywords.length) * pattern.weight;
+        if (score > bestScore) {
+            bestScore = score;
+            bestType = pattern.taskType;
+        }
+    }
+    // T2 (multi-turn) is detected by conversation length, not prompt content.
+    // The orchestrator upgrades to T2 when messageHistory.length > 6.
+    const confidence = Math.min(1.0, bestScore);
+    if (confidence < 0.2)
+        return { taskType: 'GENERAL', confidence: 0 };
+    return { taskType: bestType, confidence };
+}
+//# sourceMappingURL=TaskClassifier.js.map

package/dist/training/TaskClassifier.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"TaskClassifier.js","sourceRoot":"","sources":["../../src/training/TaskClassifier.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AASH,MAAM,QAAQ,GAAiE;IAC7E;QACE,QAAQ,EAAE,IAAI;QACd,QAAQ,EAAE;YACR,wEAAwE;YACxE,qDAAqD;YACrD,kEAAkE;SACnE;QACD,MAAM,EAAE,GAAG;KACZ;IACD;QACE,QAAQ,EAAE,IAAI;QACd,QAAQ,EAAE;YACR,yEAAyE;YACzE,mDAAmD;YACnD,wCAAwC;SACzC;QACD,MAAM,EAAE,GAAG;KACZ;IACD;QACE,QAAQ,EAAE,IAAI;QACd,QAAQ,EAAE;YACR,0DAA0D;YAC1D,8DAA8D;YAC9D,4DAA4D;SAC7D;QACD,MAAM,EAAE,GAAG;KACZ;IACD;QACE,QAAQ,EAAE,IAAI;QACd,QAAQ,EAAE;YACR,0DAA0D;YAC1D,0DAA0D;YAC1D,8CAA8C;SAC/C;QACD,MAAM,EAAE,GAAG;KACZ;CACF,CAAC;AAEF,MAAM,UAAU,YAAY,CAAC,MAAc;IACzC,MAAM,IAAI,GAAG,OAAO,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,CAAC;IAC1E,IAAI,QAAQ,GAAa,SAAS,CAAC;IACnC,IAAI,SAAS,GAAG,CAAC,CAAC;IAElB,KAAK,MAAM,OAAO,IAAI,QAAQ,EAAE,CAAC;QAC/B,IAAI,IAAI,GAAG,CAAC,CAAC;QACb,KAAK,MAAM,EAAE,IAAI,OAAO,CAAC,QAAQ,EAAE,CAAC;YAClC,IAAI,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC;gBAAE,IAAI,EAAE,CAAC;QAC5B,CAAC;QACD,MAAM,KAAK,GAAG,CAAC,IAAI,GAAG,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAC,GAAG,OAAO,CAAC,MAAM,CAAC;QAChE,IAAI,KAAK,GAAG,SAAS,EAAE,CAAC;YACtB,SAAS,GAAG,KAAK,CAAC;YAClB,QAAQ,GAAG,OAAO,CAAC,QAAQ,CAAC;QAC9B,CAAC;IACH,CAAC;IAED,0EAA0E;IAC1E,kEAAkE;IAElE,MAAM,UAAU,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,SAAS,CAAC,CAAC;IAC5C,IAAI,UAAU,GAAG,GAAG;QAAE,OAAO,EAAE,QAAQ,EAAE,SAAS,EAAE,UAAU,EAAE,CAAC,EAAE,CAAC;IACpE,OAAO,EAAE,QAAQ,EAAE,QAAQ,EAAE,UAAU,EAAE,CAAC;AAC5C,CAAC"}

package/dist/training/ThompsonRouter.d.ts

@@ -0,0 +1,53 @@
+/**
+ * ThompsonRouter — explore/exploit model selection by posterior sampling.
+ *
+ * `recommend()`/`recommendTrusted()` are greedy: they always return the current
+ * argmax. That locks the router onto early winners and never re-checks a model
+ * that looked weak on thin data — a form of overfitting in the routing policy
+ * itself. Thompson sampling fixes this: instead of taking the top mean, it draws
+ * one sample from each model's score POSTERIOR and routes to the sampled-argmax.
+ *
+ * A model with a high mean and many samples (tight posterior) usually wins
+ * (exploit). A model with few samples (wide posterior) occasionally samples high
+ * and gets a chance (explore) — which generates the very data that sharpens its
+ * posterior. Over many dispatches the matrix stops being self-confirming.
+ *
+ * POSTERIOR: a Gaussian approximation on the 0-100 composite score. mean =
+ * compositeScore; sd = explorationSpread / sqrt(sampleCount + 1). Cold-start
+ * (sampleCount 0) → widest sd → maximal exploration; as observations accumulate,
+ * sd shrinks toward exploitation. Simple, continuous, no Gamma sampler needed.
+ *
+ * OPT-IN + SAFE: this is only consulted when MODEL_ROUTER_EXPLORATION is on
+ * (default off — greedy routing is unchanged). `exclude` drops models that must
+ * never be auto-selected (e.g. the operator's cost-constrained providers); the
+ * caller passes MODEL_ROUTER_EXCLUDE so exploration can never route to a banned
+ * model.
+ */
+import type { ModelScore } from './ModelRouterMatrix.js';
+import type { RNG } from './AutoResearchStats.js';
+export interface ThompsonOptions {
+    /** RNG in [0,1). Defaults to Math.random (fresh draw per dispatch). Inject a
+     *  seeded RNG (mulberry32) for deterministic tests. */
+    rng?: RNG;
+    /** posterior sd at zero samples, on the 0-100 scale (default 20). Bigger =
+     *  more exploration of thinly-sampled models. */
+    explorationSpread?: number;
+    /** floor on posterior sd so a heavily-sampled model still has a little jitter
+     *  (default 1). Prevents total lock-in. */
+    minSpread?: number;
+    /** model IDs that must never be selected (cost/policy bans). An entry ending
+     *  in '*' is a prefix wildcard — e.g. 'grok*' excludes every xAI model,
+     *  present and future, in one rule. */
+    exclude?: string[];
+}
+/** Exact match, or prefix match when the exclude entry ends in '*'. Each entry of
+ *  the (comma-split) exclude list is matched independently, so multiple models or
+ *  providers can be banned at once — e.g. ['grok*', 'gpt-4o']. */
+export declare function isExcluded(modelId: string, exclude: string[]): boolean;
+/**
+ * Draw one posterior sample per model and return the sampled-argmax model id.
+ * Returns null if there are no eligible models (all excluded / empty input) —
+ * the caller then falls back to its safe default.
+ */
+export declare function thompsonSelect(scores: ModelScore[], opts?: ThompsonOptions): string | null;
+//# sourceMappingURL=ThompsonRouter.d.ts.map

package/dist/training/ThompsonRouter.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"ThompsonRouter.d.ts","sourceRoot":"","sources":["../../src/training/ThompsonRouter.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAEH,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,wBAAwB,CAAC;AACzD,OAAO,KAAK,EAAE,GAAG,EAAE,MAAM,wBAAwB,CAAC;AAElD,MAAM,WAAW,eAAe;IAC9B;2DACuD;IACvD,GAAG,CAAC,EAAE,GAAG,CAAC;IACV;qDACiD;IACjD,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B;+CAC2C;IAC3C,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB;;2CAEuC;IACvC,OAAO,CAAC,EAAE,MAAM,EAAE,CAAC;CACpB;AAED;;kEAEkE;AAClE,wBAAgB,UAAU,CAAC,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,EAAE,GAAG,OAAO,CAMtE;AAUD;;;;GAIG;AACH,wBAAgB,cAAc,CAAC,MAAM,EAAE,UAAU,EAAE,EAAE,IAAI,GAAE,eAAoB,GAAG,MAAM,GAAG,IAAI,CAe9F"}

package/dist/training/ThompsonRouter.js

@@ -0,0 +1,73 @@
+/**
+ * ThompsonRouter — explore/exploit model selection by posterior sampling.
+ *
+ * `recommend()`/`recommendTrusted()` are greedy: they always return the current
+ * argmax. That locks the router onto early winners and never re-checks a model
+ * that looked weak on thin data — a form of overfitting in the routing policy
+ * itself. Thompson sampling fixes this: instead of taking the top mean, it draws
+ * one sample from each model's score POSTERIOR and routes to the sampled-argmax.
+ *
+ * A model with a high mean and many samples (tight posterior) usually wins
+ * (exploit). A model with few samples (wide posterior) occasionally samples high
+ * and gets a chance (explore) — which generates the very data that sharpens its
+ * posterior. Over many dispatches the matrix stops being self-confirming.
+ *
+ * POSTERIOR: a Gaussian approximation on the 0-100 composite score. mean =
+ * compositeScore; sd = explorationSpread / sqrt(sampleCount + 1). Cold-start
+ * (sampleCount 0) → widest sd → maximal exploration; as observations accumulate,
+ * sd shrinks toward exploitation. Simple, continuous, no Gamma sampler needed.
+ *
+ * OPT-IN + SAFE: this is only consulted when MODEL_ROUTER_EXPLORATION is on
+ * (default off — greedy routing is unchanged). `exclude` drops models that must
+ * never be auto-selected (e.g. the operator's cost-constrained providers); the
+ * caller passes MODEL_ROUTER_EXCLUDE so exploration can never route to a banned
+ * model.
+ */
+/** Exact match, or prefix match when the exclude entry ends in '*'. Each entry of
+ *  the (comma-split) exclude list is matched independently, so multiple models or
+ *  providers can be banned at once — e.g. ['grok*', 'gpt-4o']. */
+export function isExcluded(modelId, exclude) {
+    for (const e of exclude) {
+        if (e.endsWith('*')) {
+            if (modelId.startsWith(e.slice(0, -1)))
+                return true;
+        }
+        else if (modelId === e)
+            return true;
+    }
+    return false;
+}
+/** Standard normal via Box-Muller, driven by the injected RNG. */
+function sampleNormal(rng) {
+    let u = 0, v = 0;
+    while (u === 0)
+        u = rng(); // avoid log(0)
+    while (v === 0)
+        v = rng();
+    return Math.sqrt(-2 * Math.log(u)) * Math.cos(2 * Math.PI * v);
+}
+/**
+ * Draw one posterior sample per model and return the sampled-argmax model id.
+ * Returns null if there are no eligible models (all excluded / empty input) —
+ * the caller then falls back to its safe default.
+ */
+export function thompsonSelect(scores, opts = {}) {
+    const rng = opts.rng ?? Math.random;
+    const spread = opts.explorationSpread ?? 20;
+    const minSpread = opts.minSpread ?? 1;
+    const exclude = opts.exclude ?? [];
+    let bestId = null;
+    let bestDraw = -Infinity;
+    for (const s of scores) {
+        if (isExcluded(s.modelId, exclude))
+            continue;
+        const sd = Math.max(minSpread, spread / Math.sqrt(s.sampleCount + 1));
+        const draw = s.compositeScore + sd * sampleNormal(rng);
+        if (draw > bestDraw) {
+            bestDraw = draw;
+            bestId = s.modelId;
+        }
+    }
+    return bestId;
+}
+//# sourceMappingURL=ThompsonRouter.js.map

package/dist/training/ThompsonRouter.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"ThompsonRouter.js","sourceRoot":"","sources":["../../src/training/ThompsonRouter.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAqBH;;kEAEkE;AAClE,MAAM,UAAU,UAAU,CAAC,OAAe,EAAE,OAAiB;IAC3D,KAAK,MAAM,CAAC,IAAI,OAAO,EAAE,CAAC;QACxB,IAAI,CAAC,CAAC,QAAQ,CAAC,GAAG,CAAC,EAAE,CAAC;YAAC,IAAI,OAAO,CAAC,UAAU,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;gBAAE,OAAO,IAAI,CAAC;QAAC,CAAC;aACxE,IAAI,OAAO,KAAK,CAAC;YAAE,OAAO,IAAI,CAAC;IACtC,CAAC;IACD,OAAO,KAAK,CAAC;AACf,CAAC;AAED,kEAAkE;AAClE,SAAS,YAAY,CAAC,GAAQ;IAC5B,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC;IACjB,OAAO,CAAC,KAAK,CAAC;QAAE,CAAC,GAAG,GAAG,EAAE,CAAC,CAAC,eAAe;IAC1C,OAAO,CAAC,KAAK,CAAC;QAAE,CAAC,GAAG,GAAG,EAAE,CAAC;IAC1B,OAAO,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,EAAE,GAAG,CAAC,CAAC,CAAC;AACjE,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,cAAc,CAAC,MAAoB,EAAE,OAAwB,EAAE;IAC7E,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,IAAI,IAAI,CAAC,MAAM,CAAC;IACpC,MAAM,MAAM,GAAG,IAAI,CAAC,iBAAiB,IAAI,EAAE,CAAC;IAC5C,MAAM,SAAS,GAAG,IAAI,CAAC,SAAS,IAAI,CAAC,CAAC;IACtC,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,IAAI,EAAE,CAAC;IAEnC,IAAI,MAAM,GAAkB,IAAI,CAAC;IACjC,IAAI,QAAQ,GAAG,CAAC,QAAQ,CAAC;IACzB,KAAK,MAAM,CAAC,IAAI,MAAM,EAAE,CAAC;QACvB,IAAI,UAAU,CAAC,CAAC,CAAC,OAAO,EAAE,OAAO,CAAC;YAAE,SAAS;QAC7C,MAAM,EAAE,GAAG,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,WAAW,GAAG,CAAC,CAAC,CAAC,CAAC;QACtE,MAAM,IAAI,GAAG,CAAC,CAAC,cAAc,GAAG,EAAE,GAAG,YAAY,CAAC,GAAG,CAAC,CAAC;QACvD,IAAI,IAAI,GAAG,QAAQ,EAAE,CAAC;YAAC,QAAQ,GAAG,IAAI,CAAC;YAAC,MAAM,GAAG,CAAC,CAAC,OAAO,CAAC;QAAC,CAAC;IAC/D,CAAC;IACD,OAAO,MAAM,CAAC;AAChB,CAAC"}

package/dist/training/VersionComparison.d.ts

@@ -0,0 +1,90 @@
+/**
+ * VersionComparison — task-by-task base-vs-candidate harness comparison.
+ *
+ * The auto-research loop asks one question per experiment: did harness version
+ * `candidateRef` actually do better than `baseRef` on the fixed-eval tasks, or is
+ * the apparent difference noise? This module produces the per-task evidence
+ * (raw per-run score arrays + means + deltas) that the Monte-Carlo gate
+ * (AutoResearchStats) turns into a keep/discard decision, and that the
+ * ExperimentLedger records.
+ *
+ * OVERFITTING GUARD (enforced here): comparisons read `split:'train'` records by
+ * default. Holdout records are NEVER used for keep/discard — only to confirm a
+ * kept candidate generalizes. The caller verifying generalization passes
+ * split:'holdout' explicitly.
+ *
+ * SCORE: the per-run scalar is `qualitativeScore` (0-100). Graders already fold
+ * correctness into it (and for pass/fail benchmarks like swebench the grader
+ * emits 100/0). The bootstrap/permutation gate resamples these per-run values,
+ * so this module returns the raw arrays, not just the means.
+ */
+import type { ModelRouterMatrix, BenchmarkRecord } from './ModelRouterMatrix.js';
+/** Per-task comparison evidence for one (taskFingerprint, base, candidate) cell. */
+export interface TaskComparison {
+    taskFingerprint: string;
+    baseRef: string;
+    candidateRef: string;
+    /** raw per-run scores on the base version (for bootstrap/permutation). */
+    baseScores: number[];
+    /** raw per-run scores on the candidate version. */
+    candScores: number[];
+    baseMean: number;
+    candMean: number;
+    /** candMean - baseMean. Positive = candidate improved the task. */
+    delta: number;
+    baseN: number;
+    candN: number;
+    /** true when BOTH arms have >= minN runs — the minimum for any trustworthy
+     *  delta. Cells failing this should not feed a keep decision. */
+    sufficient: boolean;
+}
+export interface CompareOptions {
+    split?: 'train' | 'holdout';
+    benchmarkSource?: string;
+    modelId?: string;
+    /** minimum runs per arm for `sufficient` (default 2 — n>=2 significance rule). */
+    minN?: number;
+    /** per-run scalar extractor; defaults to qualitativeScore. */
+    scoreOf?: (r: BenchmarkRecord) => number;
+}
+export interface RegressionReport {
+    candidateRef: string;
+    baseRef: string;
+    /** candMean < baseMean - epsilon, with sufficient runs. The blockers. */
+    regressions: TaskComparison[];
+    /** candMean > baseMean + epsilon, with sufficient runs. */
+    improvements: TaskComparison[];
+    /** within epsilon, or insufficient runs to judge. */
+    neutral: TaskComparison[];
+    /** tasks the candidate ran but the base never did — can't compare. */
+    uncompared: string[];
+    /** mean of all sufficient-cell deltas. Quick directional summary, NOT a
+     *  significance verdict (that's AutoResearchStats). */
+    meanDelta: number;
+}
+/**
+ * Compare a single task across two harness versions. Returns the raw per-run
+ * score arrays plus summary stats. Empty arrays when a version never ran the
+ * task (sufficient=false).
+ */
+export declare function compareVersions(matrix: ModelRouterMatrix, taskFingerprint: string, baseRef: string, candidateRef: string, opts?: CompareOptions): TaskComparison;
+/**
+ * Compare EVERY task the candidate version ran against the base version. The
+ * broad regression net: a candidate that improves its target task but quietly
+ * breaks three others must not be kept. `epsilon` is the dead-band that ignores
+ * trivial score wobble (default 0.5 on the 0-100 scale).
+ */
+export declare function regressionScan(matrix: ModelRouterMatrix, candidateRef: string, baseRef: string, opts?: CompareOptions & {
+    epsilon?: number;
+}): RegressionReport;
+/** Adapt a TaskComparison to the ExperimentLedger child-row shape. */
+export declare function toTaskResult(experimentTag: string, cmp: TaskComparison): {
+    experimentTag: string;
+    taskFingerprint: string;
+    baseScore: number;
+    candScore: number;
+    delta: number;
+    baseN: number;
+    candN: number;
+};
+//# sourceMappingURL=VersionComparison.d.ts.map

package/dist/training/VersionComparison.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"VersionComparison.d.ts","sourceRoot":"","sources":["../../src/training/VersionComparison.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAEH,OAAO,KAAK,EAAE,iBAAiB,EAAE,eAAe,EAAE,MAAM,wBAAwB,CAAC;AAEjF,oFAAoF;AACpF,MAAM,WAAW,cAAc;IAC7B,eAAe,EAAE,MAAM,CAAC;IACxB,OAAO,EAAE,MAAM,CAAC;IAChB,YAAY,EAAE,MAAM,CAAC;IACrB,0EAA0E;IAC1E,UAAU,EAAE,MAAM,EAAE,CAAC;IACrB,mDAAmD;IACnD,UAAU,EAAE,MAAM,EAAE,CAAC;IACrB,QAAQ,EAAE,MAAM,CAAC;IACjB,QAAQ,EAAE,MAAM,CAAC;IACjB,mEAAmE;IACnE,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;IACd;qEACiE;IACjE,UAAU,EAAE,OAAO,CAAC;CACrB;AAED,MAAM,WAAW,cAAc;IAC7B,KAAK,CAAC,EAAE,OAAO,GAAG,SAAS,CAAC;IAC5B,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,kFAAkF;IAClF,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,8DAA8D;IAC9D,OAAO,CAAC,EAAE,CAAC,CAAC,EAAE,eAAe,KAAK,MAAM,CAAC;CAC1C;AAED,MAAM,WAAW,gBAAgB;IAC/B,YAAY,EAAE,MAAM,CAAC;IACrB,OAAO,EAAE,MAAM,CAAC;IAChB,yEAAyE;IACzE,WAAW,EAAE,cAAc,EAAE,CAAC;IAC9B,2DAA2D;IAC3D,YAAY,EAAE,cAAc,EAAE,CAAC;IAC/B,qDAAqD;IACrD,OAAO,EAAE,cAAc,EAAE,CAAC;IAC1B,sEAAsE;IACtE,UAAU,EAAE,MAAM,EAAE,CAAC;IACrB;2DACuD;IACvD,SAAS,EAAE,MAAM,CAAC;CACnB;AAWD;;;;GAIG;AACH,wBAAgB,eAAe,CAC7B,MAAM,EAAE,iBAAiB,EACzB,eAAe,EAAE,MAAM,EACvB,OAAO,EAAE,MAAM,EACf,YAAY,EAAE,MAAM,EACpB,IAAI,GAAE,cAAmB,GACxB,cAAc,CA6BhB;AAED;;;;;GAKG;AACH,wBAAgB,cAAc,CAC5B,MAAM,EAAE,iBAAiB,EACzB,YAAY,EAAE,MAAM,EACpB,OAAO,EAAE,MAAM,EACf,IAAI,GAAE,cAAc,GAAG;IAAE,OAAO,CAAC,EAAE,MAAM,CAAA;CAAO,GAC/C,gBAAgB,CAkClB;AAED,sEAAsE;AACtE,wBAAgB,YAAY,CAAC,aAAa,EAAE,MAAM,EAAE,GAAG,EAAE,cAAc;;;;;;;;EAUtE"}

package/dist/training/VersionComparison.js

@@ -0,0 +1,117 @@
+/**
+ * VersionComparison — task-by-task base-vs-candidate harness comparison.
+ *
+ * The auto-research loop asks one question per experiment: did harness version
+ * `candidateRef` actually do better than `baseRef` on the fixed-eval tasks, or is
+ * the apparent difference noise? This module produces the per-task evidence
+ * (raw per-run score arrays + means + deltas) that the Monte-Carlo gate
+ * (AutoResearchStats) turns into a keep/discard decision, and that the
+ * ExperimentLedger records.
+ *
+ * OVERFITTING GUARD (enforced here): comparisons read `split:'train'` records by
+ * default. Holdout records are NEVER used for keep/discard — only to confirm a
+ * kept candidate generalizes. The caller verifying generalization passes
+ * split:'holdout' explicitly.
+ *
+ * SCORE: the per-run scalar is `qualitativeScore` (0-100). Graders already fold
+ * correctness into it (and for pass/fail benchmarks like swebench the grader
+ * emits 100/0). The bootstrap/permutation gate resamples these per-run values,
+ * so this module returns the raw arrays, not just the means.
+ */
+function mean(xs) {
+    if (xs.length === 0)
+        return 0;
+    return xs.reduce((s, x) => s + x, 0) / xs.length;
+}
+function round3(n) { return Math.round(n * 1000) / 1000; }
+const defaultScoreOf = (r) => r.qualitativeScore;
+/**
+ * Compare a single task across two harness versions. Returns the raw per-run
+ * score arrays plus summary stats. Empty arrays when a version never ran the
+ * task (sufficient=false).
+ */
+export function compareVersions(matrix, taskFingerprint, baseRef, candidateRef, opts = {}) {
+    const split = opts.split ?? 'train';
+    const scoreOf = opts.scoreOf ?? defaultScoreOf;
+    const minN = opts.minN ?? 2;
+    const pull = (harnessRef) => matrix
+        .getRecords({ taskFingerprint, harnessRef, split, benchmarkSource: opts.benchmarkSource, modelId: opts.modelId })
+        .map(scoreOf)
+        .filter((n) => typeof n === 'number' && !Number.isNaN(n));
+    const baseScores = pull(baseRef);
+    const candScores = pull(candidateRef);
+    const baseMean = round3(mean(baseScores));
+    const candMean = round3(mean(candScores));
+    return {
+        taskFingerprint,
+        baseRef,
+        candidateRef,
+        baseScores,
+        candScores,
+        baseMean,
+        candMean,
+        delta: round3(candMean - baseMean),
+        baseN: baseScores.length,
+        candN: candScores.length,
+        sufficient: baseScores.length >= minN && candScores.length >= minN,
+    };
+}
+/**
+ * Compare EVERY task the candidate version ran against the base version. The
+ * broad regression net: a candidate that improves its target task but quietly
+ * breaks three others must not be kept. `epsilon` is the dead-band that ignores
+ * trivial score wobble (default 0.5 on the 0-100 scale).
+ */
+export function regressionScan(matrix, candidateRef, baseRef, opts = {}) {
+    const split = opts.split ?? 'train';
+    const epsilon = opts.epsilon ?? 0.5;
+    const fingerprints = matrix.taskFingerprintsAt(candidateRef, split);
+    const regressions = [];
+    const improvements = [];
+    const neutral = [];
+    const uncompared = [];
+    const sufficientDeltas = [];
+    for (const fp of fingerprints) {
+        const cmp = compareVersions(matrix, fp, baseRef, candidateRef, opts);
+        if (cmp.baseN === 0) {
+            uncompared.push(fp);
+            continue;
+        }
+        if (!cmp.sufficient) {
+            neutral.push(cmp);
+            continue;
+        }
+        sufficientDeltas.push(cmp.delta);
+        if (cmp.delta < -epsilon)
+            regressions.push(cmp);
+        else if (cmp.delta > epsilon)
+            improvements.push(cmp);
+        else
+            neutral.push(cmp);
+    }
+    // Worst regressions first; biggest improvements first.
+    regressions.sort((a, b) => a.delta - b.delta);
+    improvements.sort((a, b) => b.delta - a.delta);
+    return {
+        candidateRef,
+        baseRef,
+        regressions,
+        improvements,
+        neutral,
+        uncompared,
+        meanDelta: round3(mean(sufficientDeltas)),
+    };
+}
+/** Adapt a TaskComparison to the ExperimentLedger child-row shape. */
+export function toTaskResult(experimentTag, cmp) {
+    return {
+        experimentTag,
+        taskFingerprint: cmp.taskFingerprint,
+        baseScore: cmp.baseMean,
+        candScore: cmp.candMean,
+        delta: cmp.delta,
+        baseN: cmp.baseN,
+        candN: cmp.candN,
+    };
+}
+//# sourceMappingURL=VersionComparison.js.map

package/dist/training/VersionComparison.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"VersionComparison.js","sourceRoot":"","sources":["../../src/training/VersionComparison.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAkDH,SAAS,IAAI,CAAC,EAAY;IACxB,IAAI,EAAE,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,CAAC,CAAC;IAC9B,OAAO,EAAE,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,GAAG,EAAE,CAAC,MAAM,CAAC;AACnD,CAAC;AAED,SAAS,MAAM,CAAC,CAAS,IAAY,OAAO,IAAI,CAAC,KAAK,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC;AAE1E,MAAM,cAAc,GAAG,CAAC,CAAkB,EAAU,EAAE,CAAC,CAAC,CAAC,gBAAgB,CAAC;AAE1E;;;;GAIG;AACH,MAAM,UAAU,eAAe,CAC7B,MAAyB,EACzB,eAAuB,EACvB,OAAe,EACf,YAAoB,EACpB,OAAuB,EAAE;IAEzB,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,IAAI,OAAO,CAAC;IACpC,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,IAAI,cAAc,CAAC;IAC/C,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,IAAI,CAAC,CAAC;IAE5B,MAAM,IAAI,GAAG,CAAC,UAAkB,EAAY,EAAE,CAC5C,MAAM;SACH,UAAU,CAAC,EAAE,eAAe,EAAE,UAAU,EAAE,KAAK,EAAE,eAAe,EAAE,IAAI,CAAC,eAAe,EAAE,OAAO,EAAE,IAAI,CAAC,OAAO,EAAE,CAAC;SAChH,GAAG,CAAC,OAAO,CAAC;SACZ,MAAM,CAAC,CAAC,CAAC,EAAe,EAAE,CAAC,OAAO,CAAC,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC;IAE3E,MAAM,UAAU,GAAG,IAAI,CAAC,OAAO,CAAC,CAAC;IACjC,MAAM,UAAU,GAAG,IAAI,CAAC,YAAY,CAAC,CAAC;IACtC,MAAM,QAAQ,GAAG,MAAM,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC;IAC1C,MAAM,QAAQ,GAAG,MAAM,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC;IAE1C,OAAO;QACL,eAAe;QACf,OAAO;QACP,YAAY;QACZ,UAAU;QACV,UAAU;QACV,QAAQ;QACR,QAAQ;QACR,KAAK,EAAE,MAAM,CAAC,QAAQ,GAAG,QAAQ,CAAC;QAClC,KAAK,EAAE,UAAU,CAAC,MAAM;QACxB,KAAK,EAAE,UAAU,CAAC,MAAM;QACxB,UAAU,EAAE,UAAU,CAAC,MAAM,IAAI,IAAI,IAAI,UAAU,CAAC,MAAM,IAAI,IAAI;KACnE,CAAC;AACJ,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,cAAc,CAC5B,MAAyB,EACzB,YAAoB,EACpB,OAAe,EACf,OAA8C,EAAE;IAEhD,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,IAAI,OAAO,CAAC;IACpC,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,IAAI,GAAG,CAAC;IACpC,MAAM,YAAY,GAAG,MAAM,CAAC,kBAAkB,CAAC,YAAY,EAAE,KAAK,CAAC,CAAC;IAEpE,MAAM,WAAW,GAAqB,EAAE,CAAC;IACzC,MAAM,YAAY,GAAqB,EAAE,CAAC;IAC1C,MAAM,OAAO,GAAqB,EAAE,CAAC;IACrC,MAAM,UAAU,GAAa,EAAE,CAAC;IAChC,MAAM,gBAAgB,GAAa,EAAE,CAAC;IAEtC,KAAK,MAAM,EAAE,IAAI,YAAY,EAAE,CAAC;QAC9B,MAAM,GAAG,GAAG,eAAe,CAAC,MAAM,EAAE,EAAE,EAAE,OAAO,EAAE,YAAY,EAAE,IAAI,CAAC,CAAC;QACrE,IAAI,GAAG,CAAC,KAAK,KAAK,CAAC,EAAE,CAAC;YAAC,UAAU,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YAAC,SAAS;QAAC,CAAC;QACvD,IAAI,CAAC,GAAG,CAAC,UAAU,EAAE,CAAC;YAAC,OAAO,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;YAAC,SAAS;QAAC,CAAC;QACrD,gBAAgB,CAAC,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC;QACjC,IAAI,GAAG,CAAC,KAAK,GAAG,CAAC,OAAO;YAAE,WAAW,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;aAC3C,IAAI,GAAG,CAAC,KAAK,GAAG,OAAO;YAAE,YAAY,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;;YAChD,OAAO,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;IACzB,CAAC;IAED,uDAAuD;IACvD,WAAW,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC;IAC9C,YAAY,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC;IAE/C,OAAO;QACL,YAAY;QACZ,OAAO;QACP,WAAW;QACX,YAAY;QACZ,OAAO;QACP,UAAU;QACV,SAAS,EAAE,MAAM,CAAC,IAAI,CAAC,gBAAgB,CAAC,CAAC;KAC1C,CAAC;AACJ,CAAC;AAED,sEAAsE;AACtE,MAAM,UAAU,YAAY,CAAC,aAAqB,EAAE,GAAmB;IACrE,OAAO;QACL,aAAa;QACb,eAAe,EAAE,GAAG,CAAC,eAAe;QACpC,SAAS,EAAE,GAAG,CAAC,QAAQ;QACvB,SAAS,EAAE,GAAG,CAAC,QAAQ;QACvB,KAAK,EAAE,GAAG,CAAC,KAAK;QAChB,KAAK,EAAE,GAAG,CAAC,KAAK;QAChB,KAAK,EAAE,GAAG,CAAC,KAAK;KACjB,CAAC;AACJ,CAAC"}

package/dist/ui/index.d.ts

@@ -0,0 +1,13 @@
+/**
+ * UI Types and Utilities
+ *
+ * Shared types for model pickers, session pickers, and other UI components
+ * that can be used across different CLI implementations.
+ *
+ * Includes:
+ * - Model/Session/Theme picker result types
+ * - Interactive menu system types (for shared config menus)
+ */
+export * from './types.js';
+export * from './menu-types.js';
+//# sourceMappingURL=index.d.ts.map

package/dist/ui/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/ui/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAGH,cAAc,YAAY,CAAC;AAG3B,cAAc,iBAAiB,CAAC"}

package/dist/ui/index.js

@@ -0,0 +1,15 @@
+/**
+ * UI Types and Utilities
+ *
+ * Shared types for model pickers, session pickers, and other UI components
+ * that can be used across different CLI implementations.
+ *
+ * Includes:
+ * - Model/Session/Theme picker result types
+ * - Interactive menu system types (for shared config menus)
+ */
+// Picker result types
+export * from './types.js';
+// Interactive menu system
+export * from './menu-types.js';
+//# sourceMappingURL=index.js.map

package/dist/ui/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/ui/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAEH,sBAAsB;AACtB,cAAc,YAAY,CAAC;AAE3B,0BAA0B;AAC1B,cAAc,iBAAiB,CAAC"}