@tangle-network/agent-runtime 0.109.1 → 0.109.2

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -3096,13 +3096,12 @@ function requireSpend(rolled, id, root) {
3096
3096
  /**
3097
3097
  * promotionGate — the statistical promotion decision over a holdout benchmark: does the
3098
3098
  * candidate strategy beat the incumbent on held-out tasks by a margin the task noise
3099
- * cannot fake? The statistics are the substrate's (`heldoutSignificance`): a SEEDED
3100
- * paired bootstrap over per-task (candidate incumbent) deltas deterministic verdict,
3101
- * a minimum-evidence floor, and the CI lower bound must clear `deltaThreshold`. A raw
3102
- * h1>h0 point comparison on m≈8 holdout tasks certifies false champions at near
3103
- * coin-flip rates; this gate is the instrument-grade replacement.
3099
+ * cannot fake? The statistics come from `heldoutSignificance`, which chooses the
3100
+ * decision interval that matches the outcome shape, applies a minimum-evidence floor,
3101
+ * and requires the interval's lower bound to clear `deltaThreshold`. A raw h1>h0 point
3102
+ * comparison on m≈8 holdout tasks certifies false champions at near coin-flip rates.
3104
3103
  */
3105
- /** Statistical promotion decision over a holdout benchmark: a seeded paired bootstrap (`heldoutSignificance`) whose CI lower bound must clear `deltaThreshold`. */
3104
+ /** Statistical promotion decision over a holdout benchmark using the outcome-appropriate interval selected by `heldoutSignificance`. */
3106
3105
  function promotionGate(opts) {
3107
3106
  const mode = opts.mode ?? "superiority";
3108
3107
  if (opts.candidate === opts.incumbent) return {
@@ -3151,8 +3150,8 @@ function promotionGate(opts) {
3151
3150
  const lift = {
3152
3151
  mean: sig.bootstrap.mean,
3153
3152
  median: sig.bootstrap.median,
3154
- low: sig.bootstrap.low,
3155
- high: sig.bootstrap.high
3153
+ low: sig.decision.low,
3154
+ high: sig.decision.high
3156
3155
  };
3157
3156
  const latSig = heldoutSignificance({
3158
3157
  before: incMs,
@@ -3168,8 +3167,8 @@ function promotionGate(opts) {
3168
3167
  const latency = {
3169
3168
  mean: latSig.bootstrap.mean,
3170
3169
  median: latSig.bootstrap.median,
3171
- low: latSig.bootstrap.low,
3172
- high: latSig.bootstrap.high
3170
+ low: latSig.decision.low,
3171
+ high: latSig.decision.high
3173
3172
  };
3174
3173
  if (mode === "superiority") {
3175
3174
  if (sig.fewRuns) return {
@@ -3222,8 +3221,8 @@ function promotionGate(opts) {
3222
3221
  const costSavings = {
3223
3222
  mean: costSig.bootstrap.mean,
3224
3223
  median: costSig.bootstrap.median,
3225
- low: costSig.bootstrap.low,
3226
- high: costSig.bootstrap.high
3224
+ low: costSig.decision.low,
3225
+ high: costSig.decision.high
3227
3226
  };
3228
3227
  if (scoreSig.fewRuns) return {
3229
3228
  promoted: false,
@@ -5413,4 +5412,4 @@ function tail(s) {
5413
5412
  //#endregion
5414
5413
  export { defineLeaderboard as $, trajectoryReport as A, flatWidenGate as B, strategyAuthorContract as C, sanitizeMcpToolSchema as Ct, runBenchmark as D, secretEnvOfMcpServer as Dt, printBenchmarkReport as E, resolveSecretEnv as Et, registerShape as F, verify as G, panel as H, FileCorpus as I, buildSteerContext as J, widen as K, InMemoryCorpus as L, runPersonified as M, builtinShapes as N, promotionGate as O, createShapeRegistry as P, harvestCorpus as Q, renderCorpusToInstructions as R, authorStrategy as S, createMcpEnvironment as St, openSandboxRun as T, mcpSecretEnvMetadataKey as Tt, pipeline as U, loopUntil as V, selectValidWinner as W, registryScopeAnalyst as X, createScopeAnalyst as Y, inProcessSandboxClient as Z, discriminatingMeans as _, auditIntent as _t, localShell as a, loopCampaignDispatch as at, selectChampion as b, connectStdioMcp as bt, createVerifierEnvironment as c, deterministicCompletion as ct, worktreeFanout as d, leaderboard as dt, dumbDriver as et, analyzeTrace as f, pairwiseSignificance as ft, streamAgentTurn as g, renderPairwiseMarkdown as gt, collectAgentTurn as h, renderLeaderboardSvg as ht, jjWorkspace as i, inlineSandboxClient as it, definePersona as j, equalKOnCost as k, failuresAnalyst as l, sentinelCompletion as lt, runCoderChecks as m, renderLeaderboardMarkdown as mt, makeFinding$1 as n, resolveSandboxClient as nt, runInWorkspace as o, loopDispatch as ot, patchDelivered as p, renderLeaderboardHtml as pt, assertTraceDerivedFindings as q, gitWorkspace as r, localSandboxClient as rt, createWaterfallCollector as s, completionAuthorizes as st, computeFindingId$1 as t, naiveDriver as tt, superviseSurface as u, stopSentinel as ut, pickChampion as v, defaultAuditorInstruction as vt, SandboxRunAbortError as w, envKeyProvider as wt, assertStrategyContract as x, materializeLocalMcp as xt, runStrategyEvolution as y, McpSpawnFault as yt, fanout as z };
5415
5414
 
5416
- //# sourceMappingURL=runtime-BVMyqgct.js.map
5415
+ //# sourceMappingURL=runtime-dpJZmZtt.js.map