@tangle-network/agent-eval 0.126.1 → 0.126.2

package/dist/{chunk-NGUYT5CI.js → chunk-Y5CLI4PY.js}

@@ -7,7 +7,7 @@ import {
   pairHoldout,
   recoverTruncatedJson,
   surfaceContentHash
-} from "./chunk-VMUENW6F.js";
+} from "./chunk-7AN2E7BU.js";
 import {
   SearchLedgerConflictError,
   SearchLedgerError,
@@ -4634,4 +4634,4 @@ export {
   verifyCodeSurface,
   resolveWorktreePath
 };
-//# sourceMappingURL=chunk-NGUYT5CI.js.map
+//# sourceMappingURL=chunk-Y5CLI4PY.js.map

package/dist/contract/index.d.ts

@@ -1477,6 +1477,24 @@ interface RunOptimizationBaseOptions<TScenario extends Scenario$1, TArtifact> ex
         costLedger?: CostLedgerHandle;
         costPhase?: string;
     }) => Promise<unknown[]>;
+    /**
+     * Optional override for how the WINNER is selected among coverage-complete
+     * candidates (and how the incumbent bar is set). Returns a lexicographic rank
+     * key — each element higher-is-better; candidates are ranked by descending key
+     * (`compareRankKeys`) and the top must STRICTLY beat the incumbent's key to
+     * promote. Defaults to `[campaignMeanComposite(campaign)]`, i.e. the historical
+     * scalar-mean ranking (single-element key ⇒ identical behavior).
+     *
+     * A binary-with-replicates consumer (e.g. swe-arena, whose ship-gate counts an
+     * instance resolved only when EVERY replicate resolved) passes a fail-closed
+     * key built from the SAME reduction its gate uses, so winner-selection and the
+     * ship-gate rank on the identical metric and can never invert — the selector
+     * cannot promote a flaky per-cell-mean candidate the gate would reject over a
+     * fail-closed candidate the gate would accept. Only the winner CHOICE changes;
+     * the descriptive `composite` (mean) on every record and the Pareto objective
+     * vectors are untouched, so proposer diversity and reporting are unaffected.
+     */
+    selectionRankKey?: (campaign: CampaignResult<TArtifact, TScenario>) => number[];
 }
 type RunOptimizationOptions<TScenario extends Scenario$1, TArtifact> = RunOptimizationBaseOptions<TScenario, TArtifact>;
 interface RunOptimizationResult<TArtifact, TScenario extends Scenario$1> {
@@ -3726,6 +3744,13 @@ interface SelfImproveOptions<TScenario extends Scenario$1, TArtifact> {
     /** Static findings forwarded to the proposer's `propose()` as `ctx.findings`
      *  (a findings-grounded proposer consumes them). Default: none. */
     findings?: unknown[];
+    /** Override how the WINNER is selected among coverage-complete candidates.
+     *  Defaults to the scalar mean composite (historical behavior). A binary-with-
+     *  replicates consumer whose ship-gate counts an instance resolved only when
+     *  every replicate resolved passes a fail-closed lexicographic key here so that
+     *  winner-selection and the ship-gate rank on the identical metric and cannot
+     *  invert. See `RunOptimizationOptions.selectionRankKey`. */
+    selectionRankKey?: RunOptimizationOptions<TScenario, TArtifact>['selectionRankKey'];
 }
 interface SelfImproveResult<TScenario extends Scenario$1, TArtifact> {
     /** Composite mean across all scenarios, baseline run. When

package/dist/contract/index.js

@@ -40,7 +40,7 @@ import {
   skillOptOptimizationMethod,
   surfaceContentHash,
   surfaceHash
-} from "../chunk-VMUENW6F.js";
+} from "../chunk-7AN2E7BU.js";
 import {
   campaignSplitDigest,
   createRunCostLedger,
@@ -340,7 +340,8 @@ async function runSelfImprove(opts, costLedger, startedAt, runDir, storage) {
     labeledStore: opts.labeledStore,
     captureSource: opts.captureSource,
     analyzeGeneration: opts.analyzeGeneration,
-    findings: opts.findings
+    findings: opts.findings,
+    selectionRankKey: opts.selectionRankKey
   });
   const winnerSearch = holdoutDeferred ? winnerSearchCampaign(result) : void 0;
   const baseline = meanComposite(