@tangle-network/agent-eval 0.82.0 → 0.84.0

package/dist/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "@tangle-network/agent-eval — wire protocol",
-    "version": "0.82.0",
+    "version": "0.84.0",
     "description": "HTTP and stdio RPC interface to agent-eval. The TypeScript runtime is the source of truth; this spec is the contract that cross-language clients (Python, Rust, Go) generate from.\n\nWire-protocol version: 1.0.0. Bumps on breaking changes to request/response schemas.",
     "contact": {
       "name": "Tangle Network",

package/dist/runtime-trajectory-BLRiaifm.d.ts

@@ -0,0 +1,49 @@
+import { a as RunSplitTag } from './run-record-De9VarXR.js';
+
+interface RuntimeTrajectoryHookEvent {
+    id: string;
+    runId: string;
+    scenarioId?: string;
+    target: string;
+    phase: string;
+    timestamp: number;
+    stepIndex?: number;
+    parentId?: string;
+    payload?: unknown;
+    metadata?: Record<string, unknown>;
+}
+interface RuntimeTrajectoryRecord {
+    id?: string;
+    scenarioId?: string;
+    splitTag?: RunSplitTag;
+    runtimeEvents?: unknown;
+    [key: string]: unknown;
+}
+interface RuntimeTrajectoryRunRecord {
+    runId: string;
+    scenarioId?: string;
+    splitTag: RunSplitTag;
+}
+interface RuntimeTrajectoryEvidenceSummary {
+    recordCount: number;
+    recordWithRuntimeEventsCount: number;
+    runtimeRunCount: number;
+    lifecycleEventCount: number;
+    defaultedSplitCount: number;
+}
+interface RuntimeTrajectoryEvidenceProjection {
+    runs: RuntimeTrajectoryRunRecord[];
+    events: RuntimeTrajectoryHookEvent[];
+    summary: RuntimeTrajectoryEvidenceSummary;
+    diagnostics: string[];
+}
+interface ProjectRuntimeTrajectoryEvidenceOptions<TRecord extends RuntimeTrajectoryRecord = RuntimeTrajectoryRecord> {
+    records: TRecord[];
+    defaultSplitTag?: RunSplitTag;
+    recordIdOf?: (record: TRecord, index: number) => string | undefined;
+    scenarioIdOf?: (record: TRecord, index: number) => string | undefined;
+}
+declare function projectRuntimeTrajectoryEvidence<TRecord extends RuntimeTrajectoryRecord>(options: ProjectRuntimeTrajectoryEvidenceOptions<TRecord>): RuntimeTrajectoryEvidenceProjection;
+declare function parseRuntimeTrajectoryHookEvent(input: unknown): RuntimeTrajectoryHookEvent | null;
+
+export { type ProjectRuntimeTrajectoryEvidenceOptions as P, type RuntimeTrajectoryRecord as R, type RuntimeTrajectoryEvidenceProjection as a, type RuntimeTrajectoryEvidenceSummary as b, type RuntimeTrajectoryHookEvent as c, type RuntimeTrajectoryRunRecord as d, projectRuntimeTrajectoryEvidence as e, parseRuntimeTrajectoryHookEvent as p };

package/docs/research/belief-state-agent-eval-roadmap.md

@@ -238,6 +238,8 @@ Follow-up local smoke on 2026-06-05 over 50 recent Codex JSONL sessions under 20
 
 Runtime hook bridge on 2026-06-05: `src/belief-state/runtime-hooks.ts` now converts `agent-runtime` decision hooks into outcome-blind shadow-probe inputs, attaches matching lifecycle hook events as probe evidence, and only converts them into full `BeliefDecisionPoint` rows when the observed action is supplied. This keeps `agent-eval` trace/analysis-only while letting the runtime emit producer-backed decision boundaries and context for the next experiment.
 
+Phase 0 measurement helper on 2026-06-06: `src/belief-state/phase0-measurement.ts` now joins runtime producer decisions, lifecycle evidence, observed action/outcome labels, and `RunRecord` split metadata into completed `BeliefDecisionPoint` rows plus a `BeliefDecisionResearchEvidencePacket`. This closes the local join/packet path; it does not clear the empirical gate until the corpus has >= 200 labeled producer-backed decisions with integrity checks and a recorded baseline.
+
 Taxonomy tightening on 2026-06-05: `src/belief-state/types.ts` now exports stable decision kinds, evidence sources, evidence quality labels, evaluation criteria, and reason codes. The intent is to make future dashboards and paper artifacts aggregate by stable IDs (`calibration`, `ope-support`, `memory-health`, `surface-attribution`, `promotion`, etc.) instead of parsing prose diagnostics.
 
 ### Q4 2026 - Phase 1: Selective Prediction and Abstention
@@ -567,6 +569,7 @@ Promotion:
 | `src/belief-state/report.test.ts` | full report status: `ship`, `hold`, `need_more_data`; recommendation cannot ship on OPE alone. |
 | `src/belief-state/code-agent-corpus.test.ts` | extracts code-agent decision corpora across Codex, Claude Code, OpenCode, Kimi Code, and Pi/PiGraph-shaped traces; inventories targets; picks failure recovery first; holds when OPE propensities are absent. |
 | `src/belief-state/runtime-hooks.test.ts` | converts runtime decision hooks to outcome-blind shadow probes; attaches matching lifecycle hook events as evidence; requires observed action for full belief rows; collector stays structurally compatible with runtime hooks. |
+| `tests/belief-state/phase0-measurement.test.ts` | joins runtime producer decisions, lifecycle events, observed labels, and run split metadata into completed rows and a research evidence packet without fabricating missing joins or propensities. |
 | `src/belief-state/types.test.ts` | guards the stable decision kinds, evidence sources/qualities, evaluation criteria, and reason-code taxonomy. |
 | `src/meta-eval/calibration.test.ts` | existing `calibrationCurve()` still works after extracting pure helper. |
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-eval",
-  "version": "0.82.0",
+  "version": "0.84.0",
   "description": "Evaluate and improve AI agents from runs, traces, judges, and feedback. Compare candidates, cluster failures, measure lift, and gate releases.",
   "homepage": "https://github.com/tangle-network/agent-eval#readme",
   "repository": {