@tangle-network/agent-eval 0.27.0 → 0.28.0

package/dist/rl.d.ts

@@ -1,16 +1,16 @@
 import { R as RunRecord, a as RunSplitTag } from './run-record-CqzahIbx.js';
-import { V as VerificationReport } from './multi-layer-verifier-LkP3LVKj.js';
-import { t as TrialResult, V as VariantAggregate, q as PromptEvolutionResult, e as MultiShotOptimizationResult } from './summary-report-DZVXOCK_.js';
+import { V as VerificationReport } from './multi-layer-verifier-BNi4-8lR.js';
+import { t as TrialResult, V as VariantAggregate, q as PromptEvolutionResult, e as MultiShotOptimizationResult } from './summary-report-jrSGb2xZ.js';
 import { O as OutcomeStore } from './outcome-store-D6KWmYvj.js';
 import { b as RubricPredictiveValidityReport } from './rubric-predictive-validity-C0uDYwG6.js';
-import { I as InterimReleaseConfidence } from './sequential-Dgz1n51-.js';
-import { S as Span, T as TraceStore } from './store-Db2Bv8Cf.js';
-import { R as Researcher, l as FailureMode, S as SteeringChange, j as ExperimentPlan, k as ExperimentResult, i as EvalCampaignResult, E as EvalCampaignOptions } from './researcher-bGkI7vCl.js';
-export { r as runEvalCampaign } from './researcher-bGkI7vCl.js';
+import { I as InterimReleaseConfidence } from './sequential-5iSVfzl2.js';
+import { S as Span, T as TraceStore } from './store-BP5be6s7.js';
+import { R as Researcher, l as FailureMode, S as SteeringChange, j as ExperimentPlan, k as ExperimentResult, i as EvalCampaignResult, E as EvalCampaignOptions } from './researcher-ClDX3KZx.js';
+export { r as runEvalCampaign } from './researcher-ClDX3KZx.js';
 import './errors-BZ9sTdz7.js';
-import './failure-cluster-C2EGSDiT.js';
-import './integrity-DK2EBVZC.js';
-import './emitter-DP_cSSiw.js';
+import './failure-cluster-D1NZKqYu.js';
+import './integrity-BAxLGJ9I.js';
+import './emitter-BqjeOvJh.js';
 
 /**
  * Test-time compute scaling curves.
@@ -529,17 +529,12 @@ declare function toAnthropicFormat(triples: PreferenceTriple[]): Array<{
 }>;
 
 /**
- * Adapters: convert legacy optimization outputs into the canonical
- * `RunRecord[]` artifact that 0.22+ primitives consume.
+ * Adapters: convert `TrialResult[]` (from `runMultiShotOptimization`,
+ * `runPromptEvolution`) into the canonical `RunRecord[]` artifact that
+ * `replayCache`, `pairedEvalueSequence`, and `rubricPredictiveValidity`
+ * consume.
  *
- * The 0.22 release standardized the campaign artifact: every cell of an
- * eval matrix produces one `RunRecord`. The pre-0.22 optimization
- * primitives (`runMultiShotOptimization`, `runPromptEvolution`) produce
- * `TrialResult[]` with a different shape. This file bridges the two so
- * the new primitives (`replayCache`, `pairedEvalueSequence`,
- * `rubricPredictiveValidity`) compose cleanly with the existing RL stack.
- *
- * The adapters are thin and explicit — every mandatory `RunRecord` field
+ * Adapters are thin and explicit — every mandatory `RunRecord` field
  * comes from a caller-supplied context (`commitSha`, `model`,
  * `promptHash`, `configHash`) plus the trial's runtime data. Defaults
  * exist for fields the trial doesn't carry (`tokenUsage`, `costUsd`),
@@ -1505,18 +1500,16 @@ interface DetectRewardHackingInput {
 declare function detectRewardHacking(input: DetectRewardHackingInput): RewardHackingReport;
 
 /**
- * `analyzeOptimizationResult` — unifies the pre-0.22 auto-research stack
+ * `analyzeOptimizationResult` — unifies the auto-research stack
  * (`runPromptEvolution`, `runMultiShotOptimization`, reflective-mutation,
- * Ax/AxRLM trace analyst) with the 0.23 RL bridge in a single call.
+ * Ax/AxRLM trace analyst) with the RL bridge in a single call.
  *
- * What this fixes: until 0.23 the optimization stack and the RL bridge
- * lived in parallel namespaces. The optimization primitives produced
- * `TrialResult[]`; the RL bridge consumed `RunRecord[]`. Trace-analyst
- * was decoupled from both. `analyzeOptimizationResult` does the wiring
- * once so consumers don't have to:
+ * The optimization primitives produce `TrialResult[]`; the RL bridge
+ * consumes `RunRecord[]`. Trace-analyst is independent of both. This
+ * function does the wiring once so consumers don't have to:
  *
- *    Optimization (existing primitives)           RL bridge (0.23)
- *    ──────────────────────────────────           ────────────────
+ *    Optimization (existing primitives)           RL bridge
+ *    ──────────────────────────────────           ────────
  *    runPromptEvolution → TrialResult[]    →
  *    runMultiShotOptimization → MSTrial[]  → analyzeOptimizationResult →
  *    reflective-mutation → mutations.jsonl →                             ↓
@@ -1527,10 +1520,10 @@ declare function detectRewardHacking(input: DetectRewardHackingInput): RewardHac
  *    ↓                                                                   │
  *    TraceAnalyst.analyze(progressLog)         ←─────────────────────────┘
  *
- * The output of this function is the canonical RL artifact set:
- * `RunRecord[]` (so every other 0.22+ primitive composes), preference
- * triples, verifiable reward signals, reward-hacking diagnosis,
- * sequential interim verdict, and (when wired) trace-analyst summary.
+ * The output is the canonical RL artifact set: `RunRecord[]` (so every
+ * other RL primitive composes), preference triples, verifiable reward
+ * signals, reward-hacking diagnosis, sequential interim verdict, and
+ * (when wired) trace-analyst summary.
  *
  * What this primitive does NOT do: it does not modify the optimization
  * primitives' internals. They keep producing `TrialResult` and emitting
@@ -1609,11 +1602,7 @@ declare function analyzeOptimizationResult(opts: AnalyzeOptimizationResultOption
  * `PredictiveValidityResearcher` — concrete `Researcher` implementation
  * that drives selection from outcome-anchored predictive validity.
  *
- * `Researcher` was a placeholder interface plus `NoopResearcher` until
- * 0.23. The 0.23 panel critique called this out: shipping the interface
- * without a default implementation that drives the loop is incomplete.
- *
- * This researcher answers each method:
+ * Each method:
  *
  *   - `inspectFailures(runs)` — synthesizes failure modes from the
  *     bottom-quartile of `RunRecord`s on the configured proxy reward.
@@ -1676,14 +1665,10 @@ declare class PredictiveValidityResearcher implements Researcher {
 }
 
 /**
- * `runRLCampaign` — the missing top-level orchestrator.
+ * `runRLCampaign` — top-level orchestrator that runs the matrix and
+ * produces every RL-ready artifact in one call.
  *
- * `runEvalCampaign` runs the matrix and produces `RunRecord[]`. The 0.23
- * RL primitives consume that artifact in different ways. Until 0.24 they
- * had to be wired together by hand at every consumer; that defeats the
- * cohesion the package is supposed to provide.
- *
- * `runRLCampaign` wires:
+ * Wires:
  *   1. `runEvalCampaign` for the matrix run (capture, integrity, hooks)
  *   2. `extractVerifiableReward` over each run, separating deterministic
  *      from probabilistic reward sources for the trainer
@@ -1697,9 +1682,6 @@ declare class PredictiveValidityResearcher implements Researcher {
  * stage's output is in there. The consumer's downstream fits in a single
  * line: pass `result.preferences` to their DPO trainer, `result.grpoRows`
  * to GRPO, `result.runs` plus `result.rewardSignals` to a custom RL loop.
- *
- * This is what the 0.23 panel critique called the "missing top-level
- * primitive." Now shipped.
  */
 
 interface RunRLCampaignOptions<V> extends EvalCampaignOptions<V> {

package/dist/rl.js

@@ -1,23 +1,23 @@
 import {
   runEvalCampaign
-} from "./chunk-SESZDQPX.js";
+} from "./chunk-RUI6SIHY.js";
 import "./chunk-4S4BM3QQ.js";
 import {
   rubricPredictiveValidity
 } from "./chunk-YRZ4M5GS.js";
 import {
   evaluateInterimReleaseConfidence
-} from "./chunk-NU65VQ7M.js";
+} from "./chunk-MAZ26DC7.js";
 import {
   benjaminiHochberg
-} from "./chunk-2A5XJB43.js";
+} from "./chunk-5AKPEK5L.js";
 import {
   wilcoxonSignedRank
-} from "./chunk-I4MBDTY5.js";
+} from "./chunk-R5UQJNKC.js";
 import "./chunk-KTGTIOFD.js";
 import "./chunk-PC4UYEBM.js";
 import "./chunk-TVVP3ZZQ.js";
-import "./chunk-4F5DQN55.js";
+import "./chunk-VSMTAMNK.js";
 import {
   ValidationError
 } from "./chunk-NG236HPC.js";