@tangle-network/agent-eval 0.120.2 → 0.120.4

package/dist/chunk-VQMK5FMP.js

@@ -0,0 +1,247 @@
+// src/trace/emitter.ts
+var TraceEmitter = class {
+  store;
+  stack = [];
+  _runId;
+  now;
+  id;
+  hooks;
+  hookErrors;
+  constructor(store, options = {}) {
+    this.store = store;
+    this.now = options.now ?? (() => Date.now());
+    this.id = options.id ?? (() => cryptoRandomId());
+    this._runId = options.runId ?? this.id();
+    this.hooks = options.onRunComplete ?? [];
+    this.hookErrors = options.hookErrors ?? "swallow";
+  }
+  get runId() {
+    return this._runId;
+  }
+  get traceStore() {
+    return this.store;
+  }
+  /** Append a hook after construction (e.g. attach the trace analyst). */
+  addRunCompleteHook(hook) {
+    this.hooks.push(hook);
+  }
+  // ── Run lifecycle ──────────────────────────────────────────────────
+  /**
+   * Begin a Run.
+   *
+   * `scenarioId` is required on the persisted Run shape — every Run downstream
+   * gets a non-empty scenarioId so filters and aggregations stay simple — but
+   * the INPUT here accepts it as optional. When omitted, startRun substitutes
+   * a sensible default (`run.layer ?? run.tags?.['kind'] ?? 'runtime'`) so
+   * runtime / operator / meta-eval runs that have no curated-scenario corpus
+   * to anchor to don't have to invent placeholder strings at the call site.
+   */
+  async startRun(run) {
+    const scenarioId = run.scenarioId ?? run.layer ?? run.tags?.kind ?? "runtime";
+    const full = {
+      ...run,
+      scenarioId,
+      runId: this._runId,
+      startedAt: this.now(),
+      status: "running"
+    };
+    await this.store.appendRun(full);
+    return full;
+  }
+  async endRun(outcome) {
+    const status = outcome?.pass === false ? "failed" : "completed";
+    await this.store.updateRun(this._runId, { endedAt: this.now(), status, outcome });
+    await this.runHooks({ runId: this._runId, emitter: this, store: this.store, outcome, status });
+  }
+  async abortRun(reason) {
+    const outcome = { pass: false, notes: reason };
+    await this.store.updateRun(this._runId, {
+      endedAt: this.now(),
+      status: "aborted",
+      outcome
+    });
+    await this.runHooks({
+      runId: this._runId,
+      emitter: this,
+      store: this.store,
+      outcome,
+      status: "aborted"
+    });
+  }
+  async runHooks(ctx) {
+    for (const hook of this.hooks) {
+      try {
+        await hook(ctx);
+      } catch (err) {
+        if (this.hookErrors === "throw") throw err;
+        try {
+          await this.store.appendEvent({
+            eventId: this.id(),
+            runId: this._runId,
+            kind: "log",
+            timestamp: this.now(),
+            payload: {
+              source: "run_complete_hook",
+              error: err instanceof Error ? err.message : String(err)
+            }
+          });
+        } catch {
+        }
+      }
+    }
+  }
+  // ── Generic span ───────────────────────────────────────────────────
+  async span(init) {
+    const spanId = this.id();
+    const parent = init.parentSpanId ?? this.stack[this.stack.length - 1];
+    const span = {
+      spanId,
+      parentSpanId: parent,
+      runId: this._runId,
+      startedAt: this.now(),
+      ...init
+    };
+    await this.store.appendSpan(span);
+    this.stack.push(spanId);
+    return this.handle(span);
+  }
+  handle(span) {
+    return {
+      span,
+      end: async (patch) => {
+        const endedAt = this.now();
+        await this.store.updateSpan(span.spanId, {
+          endedAt,
+          status: "ok",
+          ...patch
+        });
+        this.pop(span.spanId);
+      },
+      fail: async (error, patch) => {
+        const endedAt = this.now();
+        const errStr = error instanceof Error ? error.message : error;
+        await this.store.updateSpan(span.spanId, {
+          endedAt,
+          status: "error",
+          error: errStr,
+          ...patch
+        });
+        this.pop(span.spanId);
+      }
+    };
+  }
+  pop(spanId) {
+    const idx = this.stack.lastIndexOf(spanId);
+    if (idx >= 0) this.stack.splice(idx, 1);
+  }
+  // ── Typed span conveniences ────────────────────────────────────────
+  llm(init) {
+    return this.span({ kind: "llm", ...init });
+  }
+  tool(init) {
+    return this.span({ kind: "tool", ...init });
+  }
+  retrieval(init) {
+    return this.span({ kind: "retrieval", ...init });
+  }
+  async recordJudge(verdict) {
+    const spanId = this.id();
+    const now = this.now();
+    const full = {
+      spanId,
+      runId: this._runId,
+      kind: "judge",
+      startedAt: now,
+      endedAt: now,
+      status: "ok",
+      ...verdict
+    };
+    await this.store.appendSpan(full);
+    return full;
+  }
+  sandbox(init) {
+    return this.span({ kind: "sandbox", ...init });
+  }
+  // ── Events ─────────────────────────────────────────────────────────
+  async emit(event) {
+    const full = {
+      eventId: this.id(),
+      runId: this._runId,
+      spanId: event.spanId ?? this.stack[this.stack.length - 1],
+      kind: event.kind,
+      timestamp: this.now(),
+      payload: event.payload ?? {}
+    };
+    await this.store.appendEvent(full);
+    return full;
+  }
+  // ── Budget ledger ──────────────────────────────────────────────────
+  async recordBudget(entry) {
+    const full = {
+      runId: this._runId,
+      timestamp: entry.timestamp ?? this.now(),
+      dimension: entry.dimension,
+      limit: entry.limit,
+      consumed: entry.consumed,
+      remaining: entry.remaining,
+      breached: entry.breached,
+      spanId: entry.spanId ?? this.stack[this.stack.length - 1]
+    };
+    await this.store.appendBudgetEntry(full);
+    if (full.breached) {
+      await this.emit({
+        kind: "budget_breach",
+        spanId: full.spanId,
+        payload: { dimension: full.dimension, limit: full.limit, consumed: full.consumed }
+      });
+    }
+    return full;
+  }
+  // ── Artifacts ──────────────────────────────────────────────────────
+  async recordArtifact(artifact) {
+    const full = { artifactId: this.id(), runId: this._runId, ...artifact };
+    await this.store.appendArtifact(full);
+    return full;
+  }
+  // ── Nested composition ─────────────────────────────────────────────
+  /**
+   * Runs `fn` inside a span; auto-ends on success, auto-fails on throw.
+   * Returns the fn's return value. Use this for the 95% case.
+   */
+  async within(init, fn) {
+    const handle = await this.span(init);
+    try {
+      const result = await fn(handle);
+      await handle.end();
+      return result;
+    } catch (err) {
+      await handle.fail(err instanceof Error ? err : String(err));
+      throw err;
+    }
+  }
+};
+function cryptoRandomId() {
+  if (typeof globalThis.crypto?.randomUUID === "function") return globalThis.crypto.randomUUID();
+  return `${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 10)}`;
+}
+function llmSpanFromProvider(args) {
+  return {
+    name: args.name ?? args.model,
+    model: args.model,
+    messages: args.messages,
+    output: args.output,
+    inputTokens: args.usage?.inputTokens,
+    outputTokens: args.usage?.outputTokens,
+    cachedTokens: args.usage?.cachedTokens,
+    cacheWriteTokens: args.usage?.cacheWriteTokens,
+    reasoningTokens: args.usage?.reasoningTokens,
+    costUsd: args.costUsd,
+    finishReason: args.finishReason
+  };
+}
+
+export {
+  TraceEmitter,
+  llmSpanFromProvider
+};
+//# sourceMappingURL=chunk-VQMK5FMP.js.map

package/dist/chunk-VQMK5FMP.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/trace/emitter.ts"],"sourcesContent":["/**\n * TraceEmitter — hierarchical span builder that auto-parents using an\n * internal stack. One emitter per Run; emitters do NOT share state.\n *\n * Convenience methods (`llm`, `tool`, `retrieval`, `judge`, `sandbox`)\n * return a `SpanHandle` with `.end()` / `.fail()` so callers don't\n * have to thread spanIds manually. For async workflows that can't use\n * the stack (e.g. fan-out parallel calls), pass `parentSpanId`\n * explicitly.\n */\n\nimport type {\n  Artifact,\n  BudgetLedgerEntry,\n  EventKind,\n  JudgeSpan,\n  LlmSpan,\n  Message,\n  RetrievalSpan,\n  Run,\n  RunOutcome,\n  SandboxSpan,\n  Span,\n  SpanKind,\n  ToolSpan,\n  TraceEvent,\n} from './schema'\nimport type { TraceStore } from './store'\n\nexport interface SpanHandle<S extends Span = Span> {\n  span: S\n  end(patch?: Partial<S>): Promise<void>\n  fail(error: string | Error, patch?: Partial<S>): Promise<void>\n}\n\nexport interface RunCompleteHookContext {\n  runId: string\n  emitter: TraceEmitter\n  store: TraceStore\n  /** Outcome the caller passed to `endRun` (undefined for `abortRun`). */\n  outcome?: RunOutcome\n  /** Final run status. */\n  status: 'completed' | 'failed' | 'aborted'\n}\n\nexport type RunCompleteHook = (ctx: RunCompleteHookContext) => Promise<void> | void\n\nexport interface TraceEmitterOptions {\n  runId?: string\n  /** Inject a clock for deterministic tests. */\n  now?: () => number\n  /** Inject an id generator for deterministic tests. */\n  id?: () => string\n  /**\n   * Hooks fired after `endRun` / `abortRun` writes the final run state.\n   * Designed for trace-analyst auto-execution, integrity assertions, and\n   * outbound notifications. Hooks run sequentially in the order supplied.\n   *\n   * By default a hook that throws is swallowed and logged as a `note` event\n   * on the run — auto-orchestration must not crash the underlying flow.\n   * Set `hookErrors: 'throw'` to propagate.\n   */\n  onRunComplete?: RunCompleteHook[]\n  /** `'swallow'` (default) | `'throw'`. */\n  hookErrors?: 'swallow' | 'throw'\n}\n\nexport class TraceEmitter {\n  private store: TraceStore\n  private stack: string[] = []\n  private _runId: string\n  private now: () => number\n  private id: () => string\n  private hooks: RunCompleteHook[]\n  private hookErrors: 'swallow' | 'throw'\n\n  constructor(store: TraceStore, options: TraceEmitterOptions = {}) {\n    this.store = store\n    this.now = options.now ?? (() => Date.now())\n    this.id = options.id ?? (() => cryptoRandomId())\n    this._runId = options.runId ?? this.id()\n    this.hooks = options.onRunComplete ?? []\n    this.hookErrors = options.hookErrors ?? 'swallow'\n  }\n\n  get runId(): string {\n    return this._runId\n  }\n\n  get traceStore(): TraceStore {\n    return this.store\n  }\n\n  /** Append a hook after construction (e.g. attach the trace analyst). */\n  addRunCompleteHook(hook: RunCompleteHook): void {\n    this.hooks.push(hook)\n  }\n\n  // ── Run lifecycle ──────────────────────────────────────────────────\n\n  /**\n   * Begin a Run.\n   *\n   * `scenarioId` is required on the persisted Run shape — every Run downstream\n   * gets a non-empty scenarioId so filters and aggregations stay simple — but\n   * the INPUT here accepts it as optional. When omitted, startRun substitutes\n   * a sensible default (`run.layer ?? run.tags?.['kind'] ?? 'runtime'`) so\n   * runtime / operator / meta-eval runs that have no curated-scenario corpus\n   * to anchor to don't have to invent placeholder strings at the call site.\n   */\n  async startRun(\n    run: Omit<Run, 'runId' | 'scenarioId' | 'startedAt' | 'status'> & { scenarioId?: string },\n  ): Promise<Run> {\n    const scenarioId = run.scenarioId ?? run.layer ?? run.tags?.kind ?? 'runtime'\n    const full: Run = {\n      ...run,\n      scenarioId,\n      runId: this._runId,\n      startedAt: this.now(),\n      status: 'running',\n    }\n    await this.store.appendRun(full)\n    return full\n  }\n\n  async endRun(outcome?: RunOutcome): Promise<void> {\n    const status: 'completed' | 'failed' = outcome?.pass === false ? 'failed' : 'completed'\n    await this.store.updateRun(this._runId, { endedAt: this.now(), status, outcome })\n    await this.runHooks({ runId: this._runId, emitter: this, store: this.store, outcome, status })\n  }\n\n  async abortRun(reason: string): Promise<void> {\n    const outcome = { pass: false, notes: reason }\n    await this.store.updateRun(this._runId, {\n      endedAt: this.now(),\n      status: 'aborted',\n      outcome,\n    })\n    await this.runHooks({\n      runId: this._runId,\n      emitter: this,\n      store: this.store,\n      outcome,\n      status: 'aborted',\n    })\n  }\n\n  private async runHooks(ctx: RunCompleteHookContext): Promise<void> {\n    for (const hook of this.hooks) {\n      try {\n        await hook(ctx)\n      } catch (err) {\n        if (this.hookErrors === 'throw') throw err\n        try {\n          await this.store.appendEvent({\n            eventId: this.id(),\n            runId: this._runId,\n            kind: 'log',\n            timestamp: this.now(),\n            payload: {\n              source: 'run_complete_hook',\n              error: err instanceof Error ? err.message : String(err),\n            },\n          })\n        } catch {\n          // best-effort\n        }\n      }\n    }\n  }\n\n  // ── Generic span ───────────────────────────────────────────────────\n\n  async span<S extends Span = Span>(\n    init: {\n      kind: SpanKind\n      name: string\n      parentSpanId?: string\n      attributes?: Record<string, unknown>\n    } & Partial<Omit<S, 'spanId' | 'runId' | 'startedAt' | 'kind' | 'name'>>,\n  ): Promise<SpanHandle<S>> {\n    const spanId = this.id()\n    const parent = init.parentSpanId ?? this.stack[this.stack.length - 1]\n    const span = {\n      spanId,\n      parentSpanId: parent,\n      runId: this._runId,\n      startedAt: this.now(),\n      ...init,\n    } as unknown as S\n    await this.store.appendSpan(span)\n    this.stack.push(spanId)\n    return this.handle<S>(span)\n  }\n\n  private handle<S extends Span>(span: S): SpanHandle<S> {\n    return {\n      span,\n      end: async (patch?: Partial<S>) => {\n        const endedAt = this.now()\n        await this.store.updateSpan(span.spanId, {\n          endedAt,\n          status: 'ok',\n          ...patch,\n        } as Partial<Span>)\n        this.pop(span.spanId)\n      },\n      fail: async (error: string | Error, patch?: Partial<S>) => {\n        const endedAt = this.now()\n        const errStr = error instanceof Error ? error.message : error\n        await this.store.updateSpan(span.spanId, {\n          endedAt,\n          status: 'error',\n          error: errStr,\n          ...patch,\n        } as Partial<Span>)\n        this.pop(span.spanId)\n      },\n    }\n  }\n\n  private pop(spanId: string): void {\n    const idx = this.stack.lastIndexOf(spanId)\n    if (idx >= 0) this.stack.splice(idx, 1)\n  }\n\n  // ── Typed span conveniences ────────────────────────────────────────\n\n  llm(\n    init: Omit<LlmSpan, 'spanId' | 'runId' | 'kind' | 'startedAt'>,\n  ): Promise<SpanHandle<LlmSpan>> {\n    return this.span<LlmSpan>({ kind: 'llm', ...init })\n  }\n\n  tool(\n    init: Omit<ToolSpan, 'spanId' | 'runId' | 'kind' | 'startedAt'>,\n  ): Promise<SpanHandle<ToolSpan>> {\n    return this.span<ToolSpan>({ kind: 'tool', ...init })\n  }\n\n  retrieval(\n    init: Omit<RetrievalSpan, 'spanId' | 'runId' | 'kind' | 'startedAt'>,\n  ): Promise<SpanHandle<RetrievalSpan>> {\n    return this.span<RetrievalSpan>({ kind: 'retrieval', ...init })\n  }\n\n  async recordJudge(\n    verdict: Omit<JudgeSpan, 'spanId' | 'runId' | 'kind' | 'startedAt' | 'endedAt'>,\n  ): Promise<JudgeSpan> {\n    const spanId = this.id()\n    const now = this.now()\n    const full: JudgeSpan = {\n      spanId,\n      runId: this._runId,\n      kind: 'judge',\n      startedAt: now,\n      endedAt: now,\n      status: 'ok',\n      ...verdict,\n    }\n    await this.store.appendSpan(full)\n    return full\n  }\n\n  sandbox(\n    init: Omit<SandboxSpan, 'spanId' | 'runId' | 'kind' | 'startedAt'>,\n  ): Promise<SpanHandle<SandboxSpan>> {\n    return this.span<SandboxSpan>({ kind: 'sandbox', ...init })\n  }\n\n  // ── Events ─────────────────────────────────────────────────────────\n\n  async emit(event: {\n    kind: EventKind\n    spanId?: string\n    payload?: Record<string, unknown>\n  }): Promise<TraceEvent> {\n    const full: TraceEvent = {\n      eventId: this.id(),\n      runId: this._runId,\n      spanId: event.spanId ?? this.stack[this.stack.length - 1],\n      kind: event.kind,\n      timestamp: this.now(),\n      payload: event.payload ?? {},\n    }\n    await this.store.appendEvent(full)\n    return full\n  }\n\n  // ── Budget ledger ──────────────────────────────────────────────────\n\n  async recordBudget(\n    entry: Omit<BudgetLedgerEntry, 'runId' | 'timestamp'> & { timestamp?: number },\n  ): Promise<BudgetLedgerEntry> {\n    const full: BudgetLedgerEntry = {\n      runId: this._runId,\n      timestamp: entry.timestamp ?? this.now(),\n      dimension: entry.dimension,\n      limit: entry.limit,\n      consumed: entry.consumed,\n      remaining: entry.remaining,\n      breached: entry.breached,\n      spanId: entry.spanId ?? this.stack[this.stack.length - 1],\n    }\n    await this.store.appendBudgetEntry(full)\n    if (full.breached) {\n      await this.emit({\n        kind: 'budget_breach',\n        spanId: full.spanId,\n        payload: { dimension: full.dimension, limit: full.limit, consumed: full.consumed },\n      })\n    }\n    return full\n  }\n\n  // ── Artifacts ──────────────────────────────────────────────────────\n\n  async recordArtifact(artifact: Omit<Artifact, 'artifactId' | 'runId'>): Promise<Artifact> {\n    const full: Artifact = { artifactId: this.id(), runId: this._runId, ...artifact }\n    await this.store.appendArtifact(full)\n    return full\n  }\n\n  // ── Nested composition ─────────────────────────────────────────────\n\n  /**\n   * Runs `fn` inside a span; auto-ends on success, auto-fails on throw.\n   * Returns the fn's return value. Use this for the 95% case.\n   */\n  async within<T>(\n    init: Parameters<TraceEmitter['span']>[0],\n    fn: (handle: SpanHandle) => Promise<T>,\n  ): Promise<T> {\n    const handle = await this.span(init)\n    try {\n      const result = await fn(handle)\n      await handle.end()\n      return result\n    } catch (err) {\n      await handle.fail(err instanceof Error ? err : String(err))\n      throw err\n    }\n  }\n}\n\n// Helpers -------------------------------------------------------------\n\nfunction cryptoRandomId(): string {\n  if (typeof globalThis.crypto?.randomUUID === 'function') return globalThis.crypto.randomUUID()\n  return `${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 10)}`\n}\n\n/** Helper to build an LLM span handle args object from a provider-shaped response. */\nexport function llmSpanFromProvider(args: {\n  name?: string\n  model: string\n  messages: Message[]\n  output: string\n  usage?: {\n    inputTokens?: number\n    outputTokens?: number\n    cachedTokens?: number\n    cacheWriteTokens?: number\n    reasoningTokens?: number\n  }\n  costUsd?: number\n  finishReason?: string\n}): Omit<LlmSpan, 'spanId' | 'runId' | 'kind' | 'startedAt'> {\n  return {\n    name: args.name ?? args.model,\n    model: args.model,\n    messages: args.messages,\n    output: args.output,\n    inputTokens: args.usage?.inputTokens,\n    outputTokens: args.usage?.outputTokens,\n    cachedTokens: args.usage?.cachedTokens,\n    cacheWriteTokens: args.usage?.cacheWriteTokens,\n    reasoningTokens: args.usage?.reasoningTokens,\n    costUsd: args.costUsd,\n    finishReason: args.finishReason,\n  }\n}\n"],"mappings":";AAmEO,IAAM,eAAN,MAAmB;AAAA,EAChB;AAAA,EACA,QAAkB,CAAC;AAAA,EACnB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAER,YAAY,OAAmB,UAA+B,CAAC,GAAG;AAChE,SAAK,QAAQ;AACb,SAAK,MAAM,QAAQ,QAAQ,MAAM,KAAK,IAAI;AAC1C,SAAK,KAAK,QAAQ,OAAO,MAAM,eAAe;AAC9C,SAAK,SAAS,QAAQ,SAAS,KAAK,GAAG;AACvC,SAAK,QAAQ,QAAQ,iBAAiB,CAAC;AACvC,SAAK,aAAa,QAAQ,cAAc;AAAA,EAC1C;AAAA,EAEA,IAAI,QAAgB;AAClB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,IAAI,aAAyB;AAC3B,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,mBAAmB,MAA6B;AAC9C,SAAK,MAAM,KAAK,IAAI;AAAA,EACtB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAcA,MAAM,SACJ,KACc;AACd,UAAM,aAAa,IAAI,cAAc,IAAI,SAAS,IAAI,MAAM,QAAQ;AACpE,UAAM,OAAY;AAAA,MAChB,GAAG;AAAA,MACH;AAAA,MACA,OAAO,KAAK;AAAA,MACZ,WAAW,KAAK,IAAI;AAAA,MACpB,QAAQ;AAAA,IACV;AACA,UAAM,KAAK,MAAM,UAAU,IAAI;AAC/B,WAAO;AAAA,EACT;AAAA,EAEA,MAAM,OAAO,SAAqC;AAChD,UAAM,SAAiC,SAAS,SAAS,QAAQ,WAAW;AAC5E,UAAM,KAAK,MAAM,UAAU,KAAK,QAAQ,EAAE,SAAS,KAAK,IAAI,GAAG,QAAQ,QAAQ,CAAC;AAChF,UAAM,KAAK,SAAS,EAAE,OAAO,KAAK,QAAQ,SAAS,MAAM,OAAO,KAAK,OAAO,SAAS,OAAO,CAAC;AAAA,EAC/F;AAAA,EAEA,MAAM,SAAS,QAA+B;AAC5C,UAAM,UAAU,EAAE,MAAM,OAAO,OAAO,OAAO;AAC7C,UAAM,KAAK,MAAM,UAAU,KAAK,QAAQ;AAAA,MACtC,SAAS,KAAK,IAAI;AAAA,MAClB,QAAQ;AAAA,MACR;AAAA,IACF,CAAC;AACD,UAAM,KAAK,SAAS;AAAA,MAClB,OAAO,KAAK;AAAA,MACZ,SAAS;AAAA,MACT,OAAO,KAAK;AAAA,MACZ;AAAA,MACA,QAAQ;AAAA,IACV,CAAC;AAAA,EACH;AAAA,EAEA,MAAc,SAAS,KAA4C;AACjE,eAAW,QAAQ,KAAK,OAAO;AAC7B,UAAI;AACF,cAAM,KAAK,GAAG;AAAA,MAChB,SAAS,KAAK;AACZ,YAAI,KAAK,eAAe,QAAS,OAAM;AACvC,YAAI;AACF,gBAAM,KAAK,MAAM,YAAY;AAAA,YAC3B,SAAS,KAAK,GAAG;AAAA,YACjB,OAAO,KAAK;AAAA,YACZ,MAAM;AAAA,YACN,WAAW,KAAK,IAAI;AAAA,YACpB,SAAS;AAAA,cACP,QAAQ;AAAA,cACR,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,YACxD;AAAA,UACF,CAAC;AAAA,QACH,QAAQ;AAAA,QAER;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAAA;AAAA,EAIA,MAAM,KACJ,MAMwB;AACxB,UAAM,SAAS,KAAK,GAAG;AACvB,UAAM,SAAS,KAAK,gBAAgB,KAAK,MAAM,KAAK,MAAM,SAAS,CAAC;AACpE,UAAM,OAAO;AAAA,MACX;AAAA,MACA,cAAc;AAAA,MACd,OAAO,KAAK;AAAA,MACZ,WAAW,KAAK,IAAI;AAAA,MACpB,GAAG;AAAA,IACL;AACA,UAAM,KAAK,MAAM,WAAW,IAAI;AAChC,SAAK,MAAM,KAAK,MAAM;AACtB,WAAO,KAAK,OAAU,IAAI;AAAA,EAC5B;AAAA,EAEQ,OAAuB,MAAwB;AACrD,WAAO;AAAA,MACL;AAAA,MACA,KAAK,OAAO,UAAuB;AACjC,cAAM,UAAU,KAAK,IAAI;AACzB,cAAM,KAAK,MAAM,WAAW,KAAK,QAAQ;AAAA,UACvC;AAAA,UACA,QAAQ;AAAA,UACR,GAAG;AAAA,QACL,CAAkB;AAClB,aAAK,IAAI,KAAK,MAAM;AAAA,MACtB;AAAA,MACA,MAAM,OAAO,OAAuB,UAAuB;AACzD,cAAM,UAAU,KAAK,IAAI;AACzB,cAAM,SAAS,iBAAiB,QAAQ,MAAM,UAAU;AACxD,cAAM,KAAK,MAAM,WAAW,KAAK,QAAQ;AAAA,UACvC;AAAA,UACA,QAAQ;AAAA,UACR,OAAO;AAAA,UACP,GAAG;AAAA,QACL,CAAkB;AAClB,aAAK,IAAI,KAAK,MAAM;AAAA,MACtB;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,IAAI,QAAsB;AAChC,UAAM,MAAM,KAAK,MAAM,YAAY,MAAM;AACzC,QAAI,OAAO,EAAG,MAAK,MAAM,OAAO,KAAK,CAAC;AAAA,EACxC;AAAA;AAAA,EAIA,IACE,MAC8B;AAC9B,WAAO,KAAK,KAAc,EAAE,MAAM,OAAO,GAAG,KAAK,CAAC;AAAA,EACpD;AAAA,EAEA,KACE,MAC+B;AAC/B,WAAO,KAAK,KAAe,EAAE,MAAM,QAAQ,GAAG,KAAK,CAAC;AAAA,EACtD;AAAA,EAEA,UACE,MACoC;AACpC,WAAO,KAAK,KAAoB,EAAE,MAAM,aAAa,GAAG,KAAK,CAAC;AAAA,EAChE;AAAA,EAEA,MAAM,YACJ,SACoB;AACpB,UAAM,SAAS,KAAK,GAAG;AACvB,UAAM,MAAM,KAAK,IAAI;AACrB,UAAM,OAAkB;AAAA,MACtB;AAAA,MACA,OAAO,KAAK;AAAA,MACZ,MAAM;AAAA,MACN,WAAW;AAAA,MACX,SAAS;AAAA,MACT,QAAQ;AAAA,MACR,GAAG;AAAA,IACL;AACA,UAAM,KAAK,MAAM,WAAW,IAAI;AAChC,WAAO;AAAA,EACT;AAAA,EAEA,QACE,MACkC;AAClC,WAAO,KAAK,KAAkB,EAAE,MAAM,WAAW,GAAG,KAAK,CAAC;AAAA,EAC5D;AAAA;AAAA,EAIA,MAAM,KAAK,OAIa;AACtB,UAAM,OAAmB;AAAA,MACvB,SAAS,KAAK,GAAG;AAAA,MACjB,OAAO,KAAK;AAAA,MACZ,QAAQ,MAAM,UAAU,KAAK,MAAM,KAAK,MAAM,SAAS,CAAC;AAAA,MACxD,MAAM,MAAM;AAAA,MACZ,WAAW,KAAK,IAAI;AAAA,MACpB,SAAS,MAAM,WAAW,CAAC;AAAA,IAC7B;AACA,UAAM,KAAK,MAAM,YAAY,IAAI;AACjC,WAAO;AAAA,EACT;AAAA;AAAA,EAIA,MAAM,aACJ,OAC4B;AAC5B,UAAM,OAA0B;AAAA,MAC9B,OAAO,KAAK;AAAA,MACZ,WAAW,MAAM,aAAa,KAAK,IAAI;AAAA,MACvC,WAAW,MAAM;AAAA,MACjB,OAAO,MAAM;AAAA,MACb,UAAU,MAAM;AAAA,MAChB,WAAW,MAAM;AAAA,MACjB,UAAU,MAAM;AAAA,MAChB,QAAQ,MAAM,UAAU,KAAK,MAAM,KAAK,MAAM,SAAS,CAAC;AAAA,IAC1D;AACA,UAAM,KAAK,MAAM,kBAAkB,IAAI;AACvC,QAAI,KAAK,UAAU;AACjB,YAAM,KAAK,KAAK;AAAA,QACd,MAAM;AAAA,QACN,QAAQ,KAAK;AAAA,QACb,SAAS,EAAE,WAAW,KAAK,WAAW,OAAO,KAAK,OAAO,UAAU,KAAK,SAAS;AAAA,MACnF,CAAC;AAAA,IACH;AACA,WAAO;AAAA,EACT;AAAA;AAAA,EAIA,MAAM,eAAe,UAAqE;AACxF,UAAM,OAAiB,EAAE,YAAY,KAAK,GAAG,GAAG,OAAO,KAAK,QAAQ,GAAG,SAAS;AAChF,UAAM,KAAK,MAAM,eAAe,IAAI;AACpC,WAAO;AAAA,EACT;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,OACJ,MACA,IACY;AACZ,UAAM,SAAS,MAAM,KAAK,KAAK,IAAI;AACnC,QAAI;AACF,YAAM,SAAS,MAAM,GAAG,MAAM;AAC9B,YAAM,OAAO,IAAI;AACjB,aAAO;AAAA,IACT,SAAS,KAAK;AACZ,YAAM,OAAO,KAAK,eAAe,QAAQ,MAAM,OAAO,GAAG,CAAC;AAC1D,YAAM;AAAA,IACR;AAAA,EACF;AACF;AAIA,SAAS,iBAAyB;AAChC,MAAI,OAAO,WAAW,QAAQ,eAAe,WAAY,QAAO,WAAW,OAAO,WAAW;AAC7F,SAAO,GAAG,KAAK,IAAI,EAAE,SAAS,EAAE,CAAC,IAAI,KAAK,OAAO,EAAE,SAAS,EAAE,EAAE,MAAM,GAAG,EAAE,CAAC;AAC9E;AAGO,SAAS,oBAAoB,MAcyB;AAC3D,SAAO;AAAA,IACL,MAAM,KAAK,QAAQ,KAAK;AAAA,IACxB,OAAO,KAAK;AAAA,IACZ,UAAU,KAAK;AAAA,IACf,QAAQ,KAAK;AAAA,IACb,aAAa,KAAK,OAAO;AAAA,IACzB,cAAc,KAAK,OAAO;AAAA,IAC1B,cAAc,KAAK,OAAO;AAAA,IAC1B,kBAAkB,KAAK,OAAO;AAAA,IAC9B,iBAAiB,KAAK,OAAO;AAAA,IAC7B,SAAS,KAAK;AAAA,IACd,cAAc,KAAK;AAAA,EACrB;AACF;","names":[]}

package/dist/chunk-VSMTAMNK.js

@@ -0,0 +1,53 @@
+// src/pre-registration.ts
+function canonicalize(v) {
+  if (v === null || typeof v !== "object") return v;
+  if (Array.isArray(v)) return v.map(canonicalize);
+  const keys = Object.keys(v).sort();
+  const out = {};
+  for (const k of keys) out[k] = canonicalize(v[k]);
+  return out;
+}
+async function hashJson(obj) {
+  const canonical = canonicalize(obj);
+  const bytes = new TextEncoder().encode(JSON.stringify(canonical));
+  const digest = await globalThis.crypto.subtle.digest("SHA-256", bytes);
+  return Array.from(new Uint8Array(digest)).map((b) => b.toString(16).padStart(2, "0")).join("");
+}
+async function signManifest(m) {
+  const hash = await hashJson(m);
+  return { ...m, contentHash: hash, algo: "sha256-content" };
+}
+async function verifyManifest(m) {
+  const { contentHash, algo: _algo, ...rest } = m;
+  void _algo;
+  const resigned = await signManifest(rest);
+  return resigned.contentHash === contentHash;
+}
+async function evaluateHypothesis(manifest, observed) {
+  if (!await verifyManifest(manifest)) {
+    throw new Error("evaluateHypothesis: manifest content hash mismatch (tampered)");
+  }
+  const reasons = [];
+  const directionOk = manifest.direction === "increase" ? observed.effect > 0 : observed.effect < 0;
+  if (!directionOk) reasons.push("wrong_direction");
+  if (Math.abs(observed.effect) < manifest.minEffect) reasons.push("effect_too_small");
+  if (observed.pValue >= manifest.alpha) reasons.push("not_significant");
+  if (observed.n < manifest.preRegisteredN) reasons.push("undersampled");
+  return {
+    manifest,
+    observedN: observed.n,
+    observedEffect: observed.effect,
+    observedPValue: observed.pValue,
+    confirmed: reasons.length === 0,
+    rejectionReasons: reasons
+  };
+}
+
+export {
+  canonicalize,
+  hashJson,
+  signManifest,
+  verifyManifest,
+  evaluateHypothesis
+};
+//# sourceMappingURL=chunk-VSMTAMNK.js.map

package/dist/chunk-VSMTAMNK.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/pre-registration.ts"],"sourcesContent":["/**\n * Pre-registered hypotheses — declare what you're testing BEFORE the\n * run, check it AFTER. Prevents p-hacking, optional stopping, and the\n * \"we ran until it looked good\" failure mode.\n *\n * Manifest is a plain JSON-friendly object. Sign it with a content hash\n * + timestamp; the registered record becomes immutable. Post-run,\n * evaluate the manifest against observed results — the library refuses\n * to let you re-interpret a different metric as the declared one.\n */\n\nexport interface HypothesisManifest {\n  id: string\n  /** Human prose — goes into the audit trail. */\n  hypothesis: string\n  /** Metric the hypothesis claims to move. */\n  metric: string\n  /** 'increase' = candidate should score higher than baseline; 'decrease' = lower. */\n  direction: 'increase' | 'decrease'\n  /** Minimum effect size to count (same units as the metric). */\n  minEffect: number\n  /** Alpha threshold. */\n  alpha: number\n  /** Target statistical power at which sample size was pre-computed. */\n  power: number\n  /** Declared N per arm before running. */\n  preRegisteredN: number\n  /** ISO8601 timestamp the manifest was registered. */\n  registeredAt: string\n  /** Optional identifiers to tie into the trace corpus. */\n  baselineLabel?: string\n  candidateLabel?: string\n}\n\n/**\n * Identifier for the hashing scheme used to produce `contentHash`.\n *\n * `'sha256-content'` — sha256 hex over the canonicalized manifest with\n * the `contentHash` and `algo` fields stripped. Held as a string union\n * so future schemes can be added without breaking parsers; SignedManifest\n * values without `algo` deserialize cleanly because the field is optional.\n */\nexport type SignedManifestAlgo = 'sha256-content'\n\nexport interface SignedManifest extends HypothesisManifest {\n  /** sha256 hex of canonicalized manifest (everything except contentHash and algo). */\n  contentHash: string\n  /**\n   * Algorithm string describing how `contentHash` was produced.\n   *\n   * Optional on the type so serialized manifests without it still parse,\n   * but ALWAYS populated by {@link signManifest}. Consumers that want to\n   * enforce a known algorithm should reject manifests where this field\n   * is missing or unrecognized.\n   */\n  algo?: SignedManifestAlgo\n}\n\nexport interface HypothesisResult {\n  manifest: SignedManifest\n  observedN: number\n  observedEffect: number\n  observedPValue: number\n  /** True iff the observed effect hits the pre-declared direction with\n   *  magnitude ≥ minEffect AND p < alpha. */\n  confirmed: boolean\n  /** Enumerated reasons the hypothesis was rejected (each a machine-tag). */\n  rejectionReasons: Array<\n    'wrong_direction' | 'effect_too_small' | 'not_significant' | 'undersampled'\n  >\n  notes?: string\n}\n\n/**\n * Deterministic JSON canonicalization — sort object keys recursively.\n *\n * Two semantically-equal objects produce byte-identical canonicalized output;\n * this is what makes a content-hash stable across encoders, key insertion\n * orders, and runtime versions. Exported for any consumer that needs the same\n * canonicalization guarantee outside the manifest-signing path (e.g., signing\n * an artifact bundle, hashing a dataset version, etc.).\n */\nexport function canonicalize(v: unknown): unknown {\n  if (v === null || typeof v !== 'object') return v\n  if (Array.isArray(v)) return v.map(canonicalize)\n  const keys = Object.keys(v as Record<string, unknown>).sort()\n  const out: Record<string, unknown> = {}\n  for (const k of keys) out[k] = canonicalize((v as Record<string, unknown>)[k])\n  return out\n}\n\n/**\n * SHA-256 hex (full 64 chars) over the canonicalized JSON encoding of `obj`.\n *\n * The same primitive `signManifest` and `verifyManifest` are built on, exposed\n * directly so consumers signing arbitrary structured content (artifact bundles,\n * production packets, dataset manifests, etc.) don't have to re-derive\n * canonicalize+sha256 from scratch.\n *\n * Stable across:\n *   - object key insertion order (canonicalization sorts keys recursively)\n *   - encoder choice (UTF-8 via TextEncoder, fixed)\n *   - runtime (uses the Web Crypto subtle digest, present in Node ≥18 and browsers)\n *\n * Named `hashJson` to disambiguate from `prompt-registry.ts`'s `hashContent`,\n * which takes a string input and returns a truncated 12-char prompt id.\n * Use `hashJson` when you mean \"canonicalize then hash.\"\n *\n * @example\n *   const hash = await hashJson({ id: '1', kind: 'spec' })\n *   // 'a3f1...' (64 hex chars)\n */\nexport async function hashJson<T>(obj: T): Promise<string> {\n  const canonical = canonicalize(obj)\n  const bytes = new TextEncoder().encode(JSON.stringify(canonical))\n  const digest = await globalThis.crypto.subtle.digest('SHA-256', bytes)\n  return Array.from(new Uint8Array(digest))\n    .map((b) => b.toString(16).padStart(2, '0'))\n    .join('')\n}\n\n/**\n * Sign a manifest with a SHA-256 content hash.\n *\n * The hash covers the canonicalized manifest with the `contentHash`\n * and `algo` fields stripped; this lets verifiers re-sign the rest and\n * compare. Returned manifest always carries `algo: 'sha256-content'`\n * so downstream consumers can identify the scheme; manifests without\n * `algo` still verify because it is stripped before hashing on both sides.\n */\nexport async function signManifest(m: HypothesisManifest): Promise<SignedManifest> {\n  const hash = await hashJson(m)\n  return { ...m, contentHash: hash, algo: 'sha256-content' }\n}\n\n/**\n * Verify that a signed manifest has not been tampered with.\n *\n * Strips `contentHash` and `algo` before re-signing so manifests without\n * `algo` verify identically to ones that carry it.\n */\nexport async function verifyManifest(m: SignedManifest): Promise<boolean> {\n  const { contentHash, algo: _algo, ...rest } = m\n  void _algo\n  const resigned = await signManifest(rest)\n  return resigned.contentHash === contentHash\n}\n\n/**\n * Evaluate a pre-registered hypothesis against observed results.\n * Mechanical — no re-interpretation permitted.\n */\nexport async function evaluateHypothesis(\n  manifest: SignedManifest,\n  observed: { n: number; effect: number; pValue: number },\n): Promise<HypothesisResult> {\n  if (!(await verifyManifest(manifest))) {\n    throw new Error('evaluateHypothesis: manifest content hash mismatch (tampered)')\n  }\n  const reasons: HypothesisResult['rejectionReasons'] = []\n  const directionOk = manifest.direction === 'increase' ? observed.effect > 0 : observed.effect < 0\n  if (!directionOk) reasons.push('wrong_direction')\n  if (Math.abs(observed.effect) < manifest.minEffect) reasons.push('effect_too_small')\n  if (observed.pValue >= manifest.alpha) reasons.push('not_significant')\n  if (observed.n < manifest.preRegisteredN) reasons.push('undersampled')\n  return {\n    manifest,\n    observedN: observed.n,\n    observedEffect: observed.effect,\n    observedPValue: observed.pValue,\n    confirmed: reasons.length === 0,\n    rejectionReasons: reasons,\n  }\n}\n"],"mappings":";AAkFO,SAAS,aAAa,GAAqB;AAChD,MAAI,MAAM,QAAQ,OAAO,MAAM,SAAU,QAAO;AAChD,MAAI,MAAM,QAAQ,CAAC,EAAG,QAAO,EAAE,IAAI,YAAY;AAC/C,QAAM,OAAO,OAAO,KAAK,CAA4B,EAAE,KAAK;AAC5D,QAAM,MAA+B,CAAC;AACtC,aAAW,KAAK,KAAM,KAAI,CAAC,IAAI,aAAc,EAA8B,CAAC,CAAC;AAC7E,SAAO;AACT;AAuBA,eAAsB,SAAY,KAAyB;AACzD,QAAM,YAAY,aAAa,GAAG;AAClC,QAAM,QAAQ,IAAI,YAAY,EAAE,OAAO,KAAK,UAAU,SAAS,CAAC;AAChE,QAAM,SAAS,MAAM,WAAW,OAAO,OAAO,OAAO,WAAW,KAAK;AACrE,SAAO,MAAM,KAAK,IAAI,WAAW,MAAM,CAAC,EACrC,IAAI,CAAC,MAAM,EAAE,SAAS,EAAE,EAAE,SAAS,GAAG,GAAG,CAAC,EAC1C,KAAK,EAAE;AACZ;AAWA,eAAsB,aAAa,GAAgD;AACjF,QAAM,OAAO,MAAM,SAAS,CAAC;AAC7B,SAAO,EAAE,GAAG,GAAG,aAAa,MAAM,MAAM,iBAAiB;AAC3D;AAQA,eAAsB,eAAe,GAAqC;AACxE,QAAM,EAAE,aAAa,MAAM,OAAO,GAAG,KAAK,IAAI;AAC9C,OAAK;AACL,QAAM,WAAW,MAAM,aAAa,IAAI;AACxC,SAAO,SAAS,gBAAgB;AAClC;AAMA,eAAsB,mBACpB,UACA,UAC2B;AAC3B,MAAI,CAAE,MAAM,eAAe,QAAQ,GAAI;AACrC,UAAM,IAAI,MAAM,+DAA+D;AAAA,EACjF;AACA,QAAM,UAAgD,CAAC;AACvD,QAAM,cAAc,SAAS,cAAc,aAAa,SAAS,SAAS,IAAI,SAAS,SAAS;AAChG,MAAI,CAAC,YAAa,SAAQ,KAAK,iBAAiB;AAChD,MAAI,KAAK,IAAI,SAAS,MAAM,IAAI,SAAS,UAAW,SAAQ,KAAK,kBAAkB;AACnF,MAAI,SAAS,UAAU,SAAS,MAAO,SAAQ,KAAK,iBAAiB;AACrE,MAAI,SAAS,IAAI,SAAS,eAAgB,SAAQ,KAAK,cAAc;AACrE,SAAO;AAAA,IACL;AAAA,IACA,WAAW,SAAS;AAAA,IACpB,gBAAgB,SAAS;AAAA,IACzB,gBAAgB,SAAS;AAAA,IACzB,WAAW,QAAQ,WAAW;AAAA,IAC9B,kBAAkB;AAAA,EACpB;AACF;","names":[]}

package/dist/chunk-VZSRQ272.js

@@ -0,0 +1,149 @@
+// src/rl/active-curriculum.ts
+function varianceBasedCurriculum(observations, candidateCells, opts) {
+  const variancePrior = opts.variancePrior ?? 0.05;
+  const floor = opts.floorPerCell ?? 1;
+  const budget = opts.budget;
+  const grouped = /* @__PURE__ */ new Map();
+  for (const o of observations) {
+    const k = `${o.variantId}::${o.scenarioId}`;
+    const arr = grouped.get(k) ?? [];
+    arr.push(o.score);
+    grouped.set(k, arr);
+  }
+  const cellStats = candidateCells.map((c) => {
+    const k = `${c.variantId}::${c.scenarioId}`;
+    const samples = grouped.get(k) ?? [];
+    const n = samples.length;
+    const mean = n === 0 ? 0.5 : samples.reduce((s, v) => s + v, 0) / n;
+    const variance = n < 2 ? variancePrior : samples.reduce((s, v) => s + (v - mean) ** 2, 0) / (n - 1) + variancePrior;
+    const weight = Math.sqrt(variance) + 1 / Math.sqrt(Math.max(1, n));
+    return { variantId: c.variantId, scenarioId: c.scenarioId, n, mean, variance, weight };
+  });
+  const floorTotal = floor * cellStats.length;
+  if (floorTotal >= budget) {
+    const each = Math.max(1, Math.floor(budget / Math.max(1, cellStats.length)));
+    return cellStats.map((c) => ({
+      variantId: c.variantId,
+      scenarioId: c.scenarioId,
+      count: each,
+      reason: `floor allocation (budget tight; n=${c.n})`
+    }));
+  }
+  const remaining = budget - floorTotal;
+  const totalWeight = cellStats.reduce((s, c) => s + c.weight, 0);
+  return cellStats.map((c) => {
+    const proportional = totalWeight === 0 ? 0 : Math.round(c.weight / totalWeight * remaining);
+    return {
+      variantId: c.variantId,
+      scenarioId: c.scenarioId,
+      count: floor + proportional,
+      reason: `variance ${c.variance.toFixed(3)} (n=${c.n}, mean=${c.mean.toFixed(3)})`
+    };
+  });
+}
+function thompsonCurriculum(observations, candidateCells, opts) {
+  const threshold = opts.decisionThreshold ?? 0.5;
+  const alpha0 = opts.priorAlpha ?? 1;
+  const beta0 = opts.priorBeta ?? 1;
+  const rng = makeRng(opts.seed);
+  const grouped = /* @__PURE__ */ new Map();
+  for (const o of observations) {
+    const k = `${o.variantId}::${o.scenarioId}`;
+    const cur = grouped.get(k) ?? { passes: 0, failures: 0 };
+    const pass = o.pass ?? o.score >= threshold;
+    if (pass) cur.passes += 1;
+    else cur.failures += 1;
+    grouped.set(k, cur);
+  }
+  const stats = candidateCells.map((c) => {
+    const k = `${c.variantId}::${c.scenarioId}`;
+    const cur = grouped.get(k) ?? { passes: 0, failures: 0 };
+    const a = alpha0 + cur.passes;
+    const b = beta0 + cur.failures;
+    const sampled = sampleBeta(a, b, rng);
+    const distance = Math.abs(sampled - threshold);
+    const variance = a * b / ((a + b) ** 2 * (a + b + 1));
+    const sigma = Math.max(0.05, Math.sqrt(variance));
+    const weight = Math.exp(-((distance / sigma) ** 2));
+    return {
+      variantId: c.variantId,
+      scenarioId: c.scenarioId,
+      n: cur.passes + cur.failures,
+      sampled,
+      sigma,
+      weight,
+      a,
+      b
+    };
+  });
+  const totalWeight = stats.reduce((s, c) => s + c.weight, 0);
+  return stats.map((c) => {
+    const proportional = totalWeight === 0 ? 0 : Math.round(c.weight / totalWeight * opts.budget);
+    return {
+      variantId: c.variantId,
+      scenarioId: c.scenarioId,
+      count: Math.max(0, proportional),
+      reason: `Beta(${c.a.toFixed(1)},${c.b.toFixed(1)}) sample=${c.sampled.toFixed(3)} (target ${threshold})`
+    };
+  });
+}
+function observationsFromRunRecords(runs, opts = {}) {
+  const threshold = opts.passThreshold ?? 0.5;
+  const useHoldout = opts.useHoldout ?? true;
+  const out = [];
+  for (const r of runs) {
+    if (!r.scenarioId) continue;
+    const score = useHoldout ? r.outcome.holdoutScore ?? r.outcome.searchScore : r.outcome.searchScore ?? r.outcome.holdoutScore;
+    if (typeof score !== "number" || !Number.isFinite(score)) continue;
+    out.push({
+      variantId: r.candidateId,
+      scenarioId: r.scenarioId,
+      score,
+      pass: score >= threshold
+    });
+  }
+  return out;
+}
+function makeRng(seed) {
+  if (seed === void 0) return Math.random;
+  let s = seed >>> 0;
+  return () => {
+    s = s + 1831565813 >>> 0;
+    let t = s;
+    t = Math.imul(t ^ t >>> 15, t | 1);
+    t ^= t + Math.imul(t ^ t >>> 7, t | 61);
+    return ((t ^ t >>> 14) >>> 0) / 4294967296;
+  };
+}
+function sampleBeta(alpha, beta, rng) {
+  const a = Math.max(1, alpha);
+  const b = Math.max(1, beta);
+  const x = sampleGamma(a, rng);
+  const y = sampleGamma(b, rng);
+  return x / (x + y);
+}
+function sampleGamma(shape, rng) {
+  const d = shape - 1 / 3;
+  const c = 1 / Math.sqrt(9 * d);
+  while (true) {
+    let x;
+    let v;
+    do {
+      const u1 = rng() || 1e-12;
+      const u2 = rng() || 1e-12;
+      x = Math.sqrt(-2 * Math.log(u1)) * Math.cos(2 * Math.PI * u2);
+      v = 1 + c * x;
+    } while (v <= 0);
+    v = v * v * v;
+    const u = rng();
+    if (u < 1 - 0.0331 * x ** 4) return d * v;
+    if (Math.log(u) < 0.5 * x * x + d * (1 - v + Math.log(v))) return d * v;
+  }
+}
+
+export {
+  varianceBasedCurriculum,
+  thompsonCurriculum,
+  observationsFromRunRecords
+};
+//# sourceMappingURL=chunk-VZSRQ272.js.map

package/dist/chunk-VZSRQ272.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/rl/active-curriculum.ts"],"sourcesContent":["/**\n * Adaptive curriculum / active scenario selection.\n *\n * Fixed scenario sets waste sample budget on cells the policy already\n * passes (no information left) and cells the policy never passes (no\n * gradient available either). Active learning over scenarios fixes this\n * by allocating the next sample budget to cells where the policy's\n * outcome is *uncertain* — those carry the most decision-relevant signal.\n *\n * This module ships two complementary strategies:\n *\n *   1. **Variance-based** — score each (variant, scenario) cell by the\n *      empirical variance of past observations. Allocate next-round budget\n *      proportional to variance. Standard active-learning-by-uncertainty\n *      heuristic; works well when the policy is non-deterministic and\n *      cells differ in observation noise.\n *\n *   2. **Bandit-based (Thompson sampling)** — model each (variant,\n *      scenario) cell as a Beta-Bernoulli arm; sample a posterior; pick\n *      cells whose posterior mean is closest to the per-scenario decision\n *      threshold. The right primitive when scenarios are\n *      \"pass/fail\" rather than continuous, and when promotion gates fire\n *      at a known threshold (e.g., 0.5).\n *\n * The output is a *next-round budget allocation* — a list of (variant,\n * scenario, count) triples. The consumer's matrix runner consumes the\n * allocation, runs those cells, feeds the new observations back. Loop.\n *\n * Out of scope (deliberate): scenario *generation* — that's the\n * adversarial primitive's job. This module allocates over an existing\n * scenario pool.\n */\n\nimport type { RunRecord } from '../run-record'\n\nexport interface CellObservation {\n  variantId: string\n  scenarioId: string\n  /** Observed score in [0, 1]. */\n  score: number\n  /** For Bernoulli arms — derive from the score with a threshold if needed. */\n  pass?: boolean\n}\n\nexport interface CurriculumAllocation {\n  variantId: string\n  scenarioId: string\n  /** How many additional reps to run on this cell. */\n  count: number\n  /** Strategy-specific reason for the allocation. */\n  reason: string\n}\n\nexport interface VarianceCurriculumOptions {\n  /** Total reps to allocate across all cells. */\n  budget: number\n  /**\n   * Smoothing prior on variance — keeps the allocator from concentrating\n   * on a cell with one observation just because its 1-sample variance is\n   * 0. Default 0.05.\n   */\n  variancePrior?: number\n  /**\n   * Minimum reps per cell — even when the variance estimate is low, give\n   * every cell at least this many. Default 1.\n   */\n  floorPerCell?: number\n}\n\n/**\n * Variance-proportional allocation. For each cell, estimate variance from\n * past observations + a prior, then allocate the budget proportional to\n * (sqrt(variance) + 1/sqrt(n)) — a classical optimal-allocation rule\n * (Neyman 1934) that balances \"explore noisy cells\" with \"explore\n * under-sampled cells.\"\n */\nexport function varianceBasedCurriculum(\n  observations: CellObservation[],\n  candidateCells: Array<{ variantId: string; scenarioId: string }>,\n  opts: VarianceCurriculumOptions,\n): CurriculumAllocation[] {\n  const variancePrior = opts.variancePrior ?? 0.05\n  const floor = opts.floorPerCell ?? 1\n  const budget = opts.budget\n\n  const grouped = new Map<string, number[]>()\n  for (const o of observations) {\n    const k = `${o.variantId}::${o.scenarioId}`\n    const arr = grouped.get(k) ?? []\n    arr.push(o.score)\n    grouped.set(k, arr)\n  }\n\n  const cellStats = candidateCells.map((c) => {\n    const k = `${c.variantId}::${c.scenarioId}`\n    const samples = grouped.get(k) ?? []\n    const n = samples.length\n    const mean = n === 0 ? 0.5 : samples.reduce((s, v) => s + v, 0) / n\n    const variance =\n      n < 2\n        ? variancePrior\n        : samples.reduce((s, v) => s + (v - mean) ** 2, 0) / (n - 1) + variancePrior\n    // Neyman optimal allocation: weight ∝ √variance; add √(1/n) to break\n    // ties toward under-sampled cells.\n    const weight = Math.sqrt(variance) + 1 / Math.sqrt(Math.max(1, n))\n    return { variantId: c.variantId, scenarioId: c.scenarioId, n, mean, variance, weight }\n  })\n\n  // Reserve floor*N for the floor; allocate the rest proportional to weight.\n  const floorTotal = floor * cellStats.length\n  if (floorTotal >= budget) {\n    const each = Math.max(1, Math.floor(budget / Math.max(1, cellStats.length)))\n    return cellStats.map((c) => ({\n      variantId: c.variantId,\n      scenarioId: c.scenarioId,\n      count: each,\n      reason: `floor allocation (budget tight; n=${c.n})`,\n    }))\n  }\n  const remaining = budget - floorTotal\n  const totalWeight = cellStats.reduce((s, c) => s + c.weight, 0)\n  return cellStats.map((c) => {\n    const proportional = totalWeight === 0 ? 0 : Math.round((c.weight / totalWeight) * remaining)\n    return {\n      variantId: c.variantId,\n      scenarioId: c.scenarioId,\n      count: floor + proportional,\n      reason: `variance ${c.variance.toFixed(3)} (n=${c.n}, mean=${c.mean.toFixed(3)})`,\n    }\n  })\n}\n\nexport interface ThompsonCurriculumOptions {\n  budget: number\n  /**\n   * The per-scenario decision threshold. Cells whose posterior mean is\n   * closest to this get the most budget — that's where the next observation\n   * has the highest information value for the gate decision. Default 0.5.\n   */\n  decisionThreshold?: number\n  /** Beta prior parameters. Default α=β=1 (uniform). */\n  priorAlpha?: number\n  priorBeta?: number\n  /** Seed the Thompson sampler. Default unset (Math.random). */\n  seed?: number\n}\n\n/**\n * Thompson-sampling-style allocation for pass/fail cells. For each cell:\n *\n *   - Maintain Beta(α + passes, β + failures) posterior on pass-rate\n *   - Allocation weight ∝ exp(-((sampledMean - threshold) / σ)^2):\n *     cells whose sampled posterior straddles the decision boundary get\n *     the most weight; cells already clearly above or below get less.\n *\n * This is the right primitive when promotion gates fire at a known\n * threshold and you want to sharpen the posterior near the boundary.\n */\nexport function thompsonCurriculum(\n  observations: CellObservation[],\n  candidateCells: Array<{ variantId: string; scenarioId: string }>,\n  opts: ThompsonCurriculumOptions,\n): CurriculumAllocation[] {\n  const threshold = opts.decisionThreshold ?? 0.5\n  const alpha0 = opts.priorAlpha ?? 1\n  const beta0 = opts.priorBeta ?? 1\n  const rng = makeRng(opts.seed)\n\n  const grouped = new Map<string, { passes: number; failures: number }>()\n  for (const o of observations) {\n    const k = `${o.variantId}::${o.scenarioId}`\n    const cur = grouped.get(k) ?? { passes: 0, failures: 0 }\n    const pass = o.pass ?? o.score >= threshold\n    if (pass) cur.passes += 1\n    else cur.failures += 1\n    grouped.set(k, cur)\n  }\n\n  const stats = candidateCells.map((c) => {\n    const k = `${c.variantId}::${c.scenarioId}`\n    const cur = grouped.get(k) ?? { passes: 0, failures: 0 }\n    const a = alpha0 + cur.passes\n    const b = beta0 + cur.failures\n    // Sample a single Beta draw — the Thompson signal.\n    const sampled = sampleBeta(a, b, rng)\n    const distance = Math.abs(sampled - threshold)\n    // Information-near-threshold weight: closer = higher.\n    // Use Gaussian-shaped kernel with σ tuned to posterior std.\n    const variance = (a * b) / ((a + b) ** 2 * (a + b + 1))\n    const sigma = Math.max(0.05, Math.sqrt(variance))\n    const weight = Math.exp(-((distance / sigma) ** 2))\n    return {\n      variantId: c.variantId,\n      scenarioId: c.scenarioId,\n      n: cur.passes + cur.failures,\n      sampled,\n      sigma,\n      weight,\n      a,\n      b,\n    }\n  })\n\n  const totalWeight = stats.reduce((s, c) => s + c.weight, 0)\n  return stats.map((c) => {\n    const proportional = totalWeight === 0 ? 0 : Math.round((c.weight / totalWeight) * opts.budget)\n    return {\n      variantId: c.variantId,\n      scenarioId: c.scenarioId,\n      count: Math.max(0, proportional),\n      reason: `Beta(${c.a.toFixed(1)},${c.b.toFixed(1)}) sample=${c.sampled.toFixed(3)} (target ${threshold})`,\n    }\n  })\n}\n\n/** Convenience: extract `CellObservation[]` directly from `RunRecord[]`. */\nexport function observationsFromRunRecords(\n  runs: RunRecord[],\n  opts: { passThreshold?: number; useHoldout?: boolean } = {},\n): CellObservation[] {\n  const threshold = opts.passThreshold ?? 0.5\n  const useHoldout = opts.useHoldout ?? true\n  const out: CellObservation[] = []\n  for (const r of runs) {\n    if (!r.scenarioId) continue\n    const score = useHoldout\n      ? (r.outcome.holdoutScore ?? r.outcome.searchScore)\n      : (r.outcome.searchScore ?? r.outcome.holdoutScore)\n    if (typeof score !== 'number' || !Number.isFinite(score)) continue\n    out.push({\n      variantId: r.candidateId,\n      scenarioId: r.scenarioId,\n      score,\n      pass: score >= threshold,\n    })\n  }\n  return out\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction makeRng(seed?: number): () => number {\n  if (seed === undefined) return Math.random\n  let s = seed >>> 0\n  return () => {\n    s = (s + 0x6d2b79f5) >>> 0\n    let t = s\n    t = Math.imul(t ^ (t >>> 15), t | 1)\n    t ^= t + Math.imul(t ^ (t >>> 7), t | 61)\n    return ((t ^ (t >>> 14)) >>> 0) / 4294967296\n  }\n}\n\n/**\n * Sample from Beta(α, β) via the Marsaglia–Tsang method using two Gamma\n * variates. Accuracy is good for α, β > 1; we floor the parameters at 1\n * to avoid degenerate cases.\n */\nfunction sampleBeta(alpha: number, beta: number, rng: () => number): number {\n  const a = Math.max(1, alpha)\n  const b = Math.max(1, beta)\n  const x = sampleGamma(a, rng)\n  const y = sampleGamma(b, rng)\n  return x / (x + y)\n}\n\nfunction sampleGamma(shape: number, rng: () => number): number {\n  // Marsaglia–Tsang for shape ≥ 1.\n  const d = shape - 1 / 3\n  const c = 1 / Math.sqrt(9 * d)\n  while (true) {\n    let x: number\n    let v: number\n    do {\n      const u1 = rng() || 1e-12\n      const u2 = rng() || 1e-12\n      // Box-Muller for a normal sample.\n      x = Math.sqrt(-2 * Math.log(u1)) * Math.cos(2 * Math.PI * u2)\n      v = 1 + c * x\n    } while (v <= 0)\n    v = v * v * v\n    const u = rng()\n    if (u < 1 - 0.0331 * x ** 4) return d * v\n    if (Math.log(u) < 0.5 * x * x + d * (1 - v + Math.log(v))) return d * v\n  }\n}\n"],"mappings":";AA4EO,SAAS,wBACd,cACA,gBACA,MACwB;AACxB,QAAM,gBAAgB,KAAK,iBAAiB;AAC5C,QAAM,QAAQ,KAAK,gBAAgB;AACnC,QAAM,SAAS,KAAK;AAEpB,QAAM,UAAU,oBAAI,IAAsB;AAC1C,aAAW,KAAK,cAAc;AAC5B,UAAM,IAAI,GAAG,EAAE,SAAS,KAAK,EAAE,UAAU;AACzC,UAAM,MAAM,QAAQ,IAAI,CAAC,KAAK,CAAC;AAC/B,QAAI,KAAK,EAAE,KAAK;AAChB,YAAQ,IAAI,GAAG,GAAG;AAAA,EACpB;AAEA,QAAM,YAAY,eAAe,IAAI,CAAC,MAAM;AAC1C,UAAM,IAAI,GAAG,EAAE,SAAS,KAAK,EAAE,UAAU;AACzC,UAAM,UAAU,QAAQ,IAAI,CAAC,KAAK,CAAC;AACnC,UAAM,IAAI,QAAQ;AAClB,UAAM,OAAO,MAAM,IAAI,MAAM,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAClE,UAAM,WACJ,IAAI,IACA,gBACA,QAAQ,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,SAAS,GAAG,CAAC,KAAK,IAAI,KAAK;AAGnE,UAAM,SAAS,KAAK,KAAK,QAAQ,IAAI,IAAI,KAAK,KAAK,KAAK,IAAI,GAAG,CAAC,CAAC;AACjE,WAAO,EAAE,WAAW,EAAE,WAAW,YAAY,EAAE,YAAY,GAAG,MAAM,UAAU,OAAO;AAAA,EACvF,CAAC;AAGD,QAAM,aAAa,QAAQ,UAAU;AACrC,MAAI,cAAc,QAAQ;AACxB,UAAM,OAAO,KAAK,IAAI,GAAG,KAAK,MAAM,SAAS,KAAK,IAAI,GAAG,UAAU,MAAM,CAAC,CAAC;AAC3E,WAAO,UAAU,IAAI,CAAC,OAAO;AAAA,MAC3B,WAAW,EAAE;AAAA,MACb,YAAY,EAAE;AAAA,MACd,OAAO;AAAA,MACP,QAAQ,qCAAqC,EAAE,CAAC;AAAA,IAClD,EAAE;AAAA,EACJ;AACA,QAAM,YAAY,SAAS;AAC3B,QAAM,cAAc,UAAU,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,QAAQ,CAAC;AAC9D,SAAO,UAAU,IAAI,CAAC,MAAM;AAC1B,UAAM,eAAe,gBAAgB,IAAI,IAAI,KAAK,MAAO,EAAE,SAAS,cAAe,SAAS;AAC5F,WAAO;AAAA,MACL,WAAW,EAAE;AAAA,MACb,YAAY,EAAE;AAAA,MACd,OAAO,QAAQ;AAAA,MACf,QAAQ,YAAY,EAAE,SAAS,QAAQ,CAAC,CAAC,OAAO,EAAE,CAAC,UAAU,EAAE,KAAK,QAAQ,CAAC,CAAC;AAAA,IAChF;AAAA,EACF,CAAC;AACH;AA4BO,SAAS,mBACd,cACA,gBACA,MACwB;AACxB,QAAM,YAAY,KAAK,qBAAqB;AAC5C,QAAM,SAAS,KAAK,cAAc;AAClC,QAAM,QAAQ,KAAK,aAAa;AAChC,QAAM,MAAM,QAAQ,KAAK,IAAI;AAE7B,QAAM,UAAU,oBAAI,IAAkD;AACtE,aAAW,KAAK,cAAc;AAC5B,UAAM,IAAI,GAAG,EAAE,SAAS,KAAK,EAAE,UAAU;AACzC,UAAM,MAAM,QAAQ,IAAI,CAAC,KAAK,EAAE,QAAQ,GAAG,UAAU,EAAE;AACvD,UAAM,OAAO,EAAE,QAAQ,EAAE,SAAS;AAClC,QAAI,KAAM,KAAI,UAAU;AAAA,QACnB,KAAI,YAAY;AACrB,YAAQ,IAAI,GAAG,GAAG;AAAA,EACpB;AAEA,QAAM,QAAQ,eAAe,IAAI,CAAC,MAAM;AACtC,UAAM,IAAI,GAAG,EAAE,SAAS,KAAK,EAAE,UAAU;AACzC,UAAM,MAAM,QAAQ,IAAI,CAAC,KAAK,EAAE,QAAQ,GAAG,UAAU,EAAE;AACvD,UAAM,IAAI,SAAS,IAAI;AACvB,UAAM,IAAI,QAAQ,IAAI;AAEtB,UAAM,UAAU,WAAW,GAAG,GAAG,GAAG;AACpC,UAAM,WAAW,KAAK,IAAI,UAAU,SAAS;AAG7C,UAAM,WAAY,IAAI,MAAO,IAAI,MAAM,KAAK,IAAI,IAAI;AACpD,UAAM,QAAQ,KAAK,IAAI,MAAM,KAAK,KAAK,QAAQ,CAAC;AAChD,UAAM,SAAS,KAAK,IAAI,GAAG,WAAW,UAAU,EAAE;AAClD,WAAO;AAAA,MACL,WAAW,EAAE;AAAA,MACb,YAAY,EAAE;AAAA,MACd,GAAG,IAAI,SAAS,IAAI;AAAA,MACpB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF,CAAC;AAED,QAAM,cAAc,MAAM,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,QAAQ,CAAC;AAC1D,SAAO,MAAM,IAAI,CAAC,MAAM;AACtB,UAAM,eAAe,gBAAgB,IAAI,IAAI,KAAK,MAAO,EAAE,SAAS,cAAe,KAAK,MAAM;AAC9F,WAAO;AAAA,MACL,WAAW,EAAE;AAAA,MACb,YAAY,EAAE;AAAA,MACd,OAAO,KAAK,IAAI,GAAG,YAAY;AAAA,MAC/B,QAAQ,QAAQ,EAAE,EAAE,QAAQ,CAAC,CAAC,IAAI,EAAE,EAAE,QAAQ,CAAC,CAAC,YAAY,EAAE,QAAQ,QAAQ,CAAC,CAAC,YAAY,SAAS;AAAA,IACvG;AAAA,EACF,CAAC;AACH;AAGO,SAAS,2BACd,MACA,OAAyD,CAAC,GACvC;AACnB,QAAM,YAAY,KAAK,iBAAiB;AACxC,QAAM,aAAa,KAAK,cAAc;AACtC,QAAM,MAAyB,CAAC;AAChC,aAAW,KAAK,MAAM;AACpB,QAAI,CAAC,EAAE,WAAY;AACnB,UAAM,QAAQ,aACT,EAAE,QAAQ,gBAAgB,EAAE,QAAQ,cACpC,EAAE,QAAQ,eAAe,EAAE,QAAQ;AACxC,QAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,EAAG;AAC1D,QAAI,KAAK;AAAA,MACP,WAAW,EAAE;AAAA,MACb,YAAY,EAAE;AAAA,MACd;AAAA,MACA,MAAM,SAAS;AAAA,IACjB,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAIA,SAAS,QAAQ,MAA6B;AAC5C,MAAI,SAAS,OAAW,QAAO,KAAK;AACpC,MAAI,IAAI,SAAS;AACjB,SAAO,MAAM;AACX,QAAK,IAAI,eAAgB;AACzB,QAAI,IAAI;AACR,QAAI,KAAK,KAAK,IAAK,MAAM,IAAK,IAAI,CAAC;AACnC,SAAK,IAAI,KAAK,KAAK,IAAK,MAAM,GAAI,IAAI,EAAE;AACxC,aAAS,IAAK,MAAM,QAAS,KAAK;AAAA,EACpC;AACF;AAOA,SAAS,WAAW,OAAe,MAAc,KAA2B;AAC1E,QAAM,IAAI,KAAK,IAAI,GAAG,KAAK;AAC3B,QAAM,IAAI,KAAK,IAAI,GAAG,IAAI;AAC1B,QAAM,IAAI,YAAY,GAAG,GAAG;AAC5B,QAAM,IAAI,YAAY,GAAG,GAAG;AAC5B,SAAO,KAAK,IAAI;AAClB;AAEA,SAAS,YAAY,OAAe,KAA2B;AAE7D,QAAM,IAAI,QAAQ,IAAI;AACtB,QAAM,IAAI,IAAI,KAAK,KAAK,IAAI,CAAC;AAC7B,SAAO,MAAM;AACX,QAAI;AACJ,QAAI;AACJ,OAAG;AACD,YAAM,KAAK,IAAI,KAAK;AACpB,YAAM,KAAK,IAAI,KAAK;AAEpB,UAAI,KAAK,KAAK,KAAK,KAAK,IAAI,EAAE,CAAC,IAAI,KAAK,IAAI,IAAI,KAAK,KAAK,EAAE;AAC5D,UAAI,IAAI,IAAI;AAAA,IACd,SAAS,KAAK;AACd,QAAI,IAAI,IAAI;AACZ,UAAM,IAAI,IAAI;AACd,QAAI,IAAI,IAAI,SAAS,KAAK,EAAG,QAAO,IAAI;AACxC,QAAI,KAAK,IAAI,CAAC,IAAI,MAAM,IAAI,IAAI,KAAK,IAAI,IAAI,KAAK,IAAI,CAAC,GAAI,QAAO,IAAI;AAAA,EACxE;AACF;","names":[]}