@tangle-network/agent-eval 0.106.1 → 0.106.3

package/dist/{chunk-IA3X6CSF.js → chunk-J22CKVXN.js}

@@ -418,4 +418,4 @@ export {
   runEval,
   evolutionaryProposer
 };
-//# sourceMappingURL=chunk-IA3X6CSF.js.map
+//# sourceMappingURL=chunk-J22CKVXN.js.map

package/dist/chunk-J22CKVXN.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/campaign/gates/compose.ts","../src/campaign/gates/default-production-gate.ts","../src/campaign/gates/power-preflight.ts","../src/campaign/gates/promotion-policy.ts","../src/campaign/presets/run-eval.ts","../src/campaign/proposers/evolutionary.ts"],"sourcesContent":["/**\n * Compose multiple `Gate` implementations — every gate must pass for the\n * composite to ship. Closes the alignment reviewer's \"default-only\n * heldOutGate + costGate would happily promote a reward-hacked prompt\"\n * concern by making safety gates first-class composable defaults.\n */\n\nimport type { Gate, GateContext, GateDecision, GateResult, Scenario } from '../types'\n\n/** Compose gates — all must `ship` for the composite to `ship`. First\n *  non-ship verdict short-circuits the composite verdict, but ALL gates run\n *  (so the result records every gate's reason — useful for diagnostics). */\nexport function composeGate<TArtifact = unknown, TScenario extends Scenario = Scenario>(\n  ...gates: Array<Gate<TArtifact, TScenario>>\n): Gate<TArtifact, TScenario> {\n  if (gates.length === 0) {\n    throw new Error('composeGate requires at least one gate')\n  }\n  return {\n    name: `composed(${gates.map((g) => g.name).join(',')})`,\n    async decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult> {\n      const results: Array<{ gate: Gate<TArtifact, TScenario>; res: GateResult }> = []\n      for (const gate of gates) {\n        const res = await gate.decide(ctx)\n        results.push({ gate, res })\n      }\n\n      // Substrate-wide verdict policy:\n      //   - all 'ship' → 'ship'\n      //   - any 'arch_ceiling' → 'arch_ceiling' (architectural ceiling beats other holds)\n      //   - any 'model_ceiling' → 'model_ceiling'\n      //   - any 'hold' → 'hold'\n      //   - else 'need_more_work'\n      const decisions = results.map((r) => r.res.decision)\n      const overall: GateDecision = decisions.every((d) => d === 'ship')\n        ? 'ship'\n        : decisions.includes('arch_ceiling')\n          ? 'arch_ceiling'\n          : decisions.includes('model_ceiling')\n            ? 'model_ceiling'\n            : decisions.includes('hold')\n              ? 'hold'\n              : 'need_more_work'\n\n      const contributing = results.flatMap((r) =>\n        r.res.contributingGates.length > 0\n          ? r.res.contributingGates\n          : [{ name: r.gate.name, passed: r.res.decision === 'ship', detail: r.res }],\n      )\n\n      const reasons = results.flatMap((r) =>\n        r.res.reasons.map((reason) => `[${r.gate.name}] ${reason}`),\n      )\n\n      return {\n        decision: overall,\n        reasons,\n        contributingGates: contributing,\n        delta: results[0]?.res.delta,\n      }\n    },\n  }\n}\n","/**\n * `defaultProductionGate` — composes the substrate's existing safety\n * primitives (red-team / reward-hacking / canary / heldout) into a single\n * Gate.decide shape. Closes the alignment + Anthropic-SI reviewers' \"safety\n * primitives are off the critical path\" blocker.\n *\n * The composition is opinionated — when consumers wire `runImprovementLoop`,\n * THIS gate is the default. Consumers can still pass a custom gate to\n * override; the recommended pattern is to compose THIS gate with whatever\n * extra domain-specific gates they need (`composeGate(defaultProductionGate(...), customGate)`).\n */\n\nimport type { CanaryReport } from '../../canary'\nimport { runCanaries } from '../../canary'\nimport type { RedTeamCase } from '../../red-team'\nimport { scoreRedTeamOutput } from '../../red-team'\nimport type { RewardHackingReport } from '../../rl/reward-hacking'\nimport { detectRewardHacking } from '../../rl/reward-hacking'\nimport type { RunRecord } from '../../run-record'\nimport type { Gate, GateContext, GateResult, Scenario } from '../types'\nimport {\n  dimensionRegressions,\n  heldoutSignificance,\n  pairHoldout,\n  TIE_WARN_FRACTION,\n} from './statistical-heldout'\n\nexport interface DefaultProductionGateOptions {\n  /** Required: scenarios held out from training; substrate compares\n   *  candidate-on-holdout vs baseline-on-holdout. */\n  holdoutScenarios: Scenario[]\n  /** Minimum held-out lift the **paired-bootstrap CI lower bound** must clear\n   *  to ship — NOT a point estimate. Default 0 ⇒ \"confidently positive at the\n   *  confidence level\". Interpreted in the judge's native composite scale (set\n   *  e.g. 2 for a 0-100 rubric to require a ≥2-point significant gain). */\n  deltaThreshold?: number\n  /** Confidence level for the held-out + dimension bootstraps. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resamples. Default 2000. */\n  bootstrapResamples?: number\n  /** Fixed bootstrap seed for a deterministic verdict. Default 1337. */\n  bootstrapSeed?: number\n  /** Minimum paired holdout observations (scenarios × reps) before a\n   *  significance claim is allowed; below it the gate HOLDS with `few_runs`\n   *  rather than reading a degenerate CI. Default 3. */\n  minProductiveRuns?: number\n  /** Ship statistic for the held-out significance test. Default `'mean'`\n   *  (tie-robust — see `heldoutSignificance`). Pass `'median'` for\n   *  outlier-robustness at the cost of tie-blindness. */\n  heldoutStatistic?: 'mean' | 'median'\n  /** Critical judge dimensions that must NOT significantly regress even when\n   *  the net composite rises (anti-Goodhart). The gate HOLDS if any listed\n   *  dimension's paired-delta CI lower bound < −`regressionTolerance`. E.g.\n   *  `['hallucination_free']` for a legal agent. */\n  criticalDimensions?: string[]\n  /** Tolerance for the per-dimension regression guard, in the dimension's\n   *  native scale. When omitted it auto-scales off observed magnitudes:\n   *  0.05 on [0,1], 5 on 0-100. */\n  regressionTolerance?: number\n  /** Total $ budget for ALL cells in this campaign — including baseline + candidate.\n   *  Composite verdict refuses to ship when spend exceeded budget. */\n  budgetUsd?: number\n  /** Red-team cases to probe candidate outputs against. When omitted the\n   *  substrate uses `DEFAULT_RED_TEAM_CORPUS`. Provide a domain-specific\n   *  battery for tighter coverage. */\n  redTeamBattery?: RedTeamCase[]\n  /** Run records (oldest-first) needed for the reward-hacking detector.\n   *  Substrate populates from prior production-loop generations. */\n  recentRuns?: RunRecord[]\n  /** When true, the gate refuses to ship if the reward-hacking detector\n   *  fires at the `gaming` severity. Default true. */\n  blockOnRewardHackingGaming?: boolean\n}\n\n/**\n * Opinionated production gate composing held-out significance, red-team, reward-hacking, and canary checks into a single `Gate.decide` decision.\n */\nexport function defaultProductionGate<TArtifact, TScenario extends Scenario>(\n  options: DefaultProductionGateOptions,\n): Gate<TArtifact, TScenario> {\n  const deltaThreshold = options.deltaThreshold ?? 0\n  const confidence = options.confidence ?? 0.95\n  const resamples = options.bootstrapResamples ?? 2000\n  const seed = options.bootstrapSeed ?? 1337\n  const minProductiveRuns = options.minProductiveRuns ?? 3\n  const heldoutStatistic = options.heldoutStatistic ?? 'mean'\n  const blockOnGaming = options.blockOnRewardHackingGaming ?? true\n\n  return {\n    name: 'defaultProductionGate',\n    async decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult> {\n      const reasons: string[] = []\n      const contributing: Array<{ name: string; passed: boolean; detail: unknown }> = []\n\n      // ── (1) heldout composite lift — paired-bootstrap CI, NOT a point estimate\n      // The shipped false positive: the baseline re-scored against itself read\n      // run-to-run model noise (91 vs 95) as a \"+4 lift\" and shipped, because a\n      // point estimate carries no confidence interval. Pair candidate vs\n      // baseline holdout cells by FULL cellId (never averaging reps away) and\n      // ship only when the bootstrap CI lower bound clears the threshold —\n      // i.e. the gain is real at the confidence level, not noise.\n      const scenarioIds = new Set(options.holdoutScenarios.map((s) => s.id))\n      const sig = heldoutSignificance(\n        pairHoldout(\n          ctx.judgeScores,\n          ctx.baselineJudgeScores ?? ctx.judgeScores,\n          scenarioIds,\n          (s) => s.composite,\n        ),\n        {\n          deltaThreshold,\n          minProductiveRuns,\n          confidence,\n          resamples,\n          seed,\n          statistic: heldoutStatistic,\n        },\n      )\n      // Point estimate of the CHOSEN ship statistic (mean by default); `.low`/\n      // `.high` are its CI. The median is kept as a diagnostic.\n      const delta = heldoutStatistic === 'median' ? sig.bootstrap.median : sig.bootstrap.mean\n      const heldoutPass = sig.significant\n      contributing.push({\n        name: 'heldout-significance',\n        passed: heldoutPass,\n        detail: {\n          n: sig.n,\n          delta,\n          deltaMean: sig.bootstrap.mean,\n          deltaMedianDiagnostic: sig.medianBootstrap.median,\n          // Back-compat: prior consumers read `deltaMedian`. It now always carries\n          // the median diagnostic (the ship decision keys on `delta`/mean).\n          deltaMedian: sig.medianBootstrap.median,\n          tieFraction: sig.tieFraction,\n          ciLow: sig.bootstrap.low,\n          ciHigh: sig.bootstrap.high,\n          confidence: sig.bootstrap.confidence,\n          deltaThreshold,\n          fewRuns: sig.fewRuns,\n        },\n      })\n      if (!heldoutPass) {\n        const tieNote =\n          sig.tieFraction >= TIE_WARN_FRACTION\n            ? `; ${(sig.tieFraction * 100).toFixed(0)}% tied scenarios`\n            : ''\n        reasons.push(\n          sig.fewRuns\n            ? `held-out: only ${sig.n} paired runs (< ${minProductiveRuns}) — too few to claim significance`\n            : `held-out CI.low ${sig.bootstrap.low.toFixed(3)} ≤ threshold ${deltaThreshold} (${heldoutStatistic} Δ ${delta.toFixed(3)}, ${(sig.bootstrap.confidence * 100).toFixed(0)}% CI [${sig.bootstrap.low.toFixed(3)}, ${sig.bootstrap.high.toFixed(3)}]${tieNote})`,\n        )\n      }\n\n      // ── (1b) per-dimension regression guard (anti-Goodhart) ──────────\n      // A net composite gain can hide a regression on a safety-critical\n      // dimension (e.g. hallucination_free for a legal agent — the verified run\n      // gained +25/+25 on deadline/fee while LOSING -30 on hallucination, and\n      // the composite-only gate never saw it). Block ship if any guarded\n      // dimension's paired-delta CI lower bound falls below −tolerance.\n      const dimRegs = options.criticalDimensions?.length\n        ? dimensionRegressions(\n            ctx.judgeScores,\n            ctx.baselineJudgeScores ?? ctx.judgeScores,\n            scenarioIds,\n            options.criticalDimensions,\n            { tolerance: options.regressionTolerance, confidence, resamples, seed },\n          )\n        : []\n      const regressed = dimRegs.filter((d) => d.regressed)\n      const dimPass = regressed.length === 0\n      contributing.push({\n        name: 'dimension-regression',\n        passed: dimPass,\n        detail: {\n          guarded: options.criticalDimensions ?? [],\n          regressions: dimRegs.map((d) => ({\n            dimension: d.dimension,\n            ciLow: d.bootstrap.low,\n            median: d.bootstrap.median,\n            tolerance: d.tolerance,\n            n: d.n,\n            regressed: d.regressed,\n          })),\n        },\n      })\n      if (!dimPass) {\n        reasons.push(\n          `critical dimension(s) regressed: ${regressed.map((d) => `${d.dimension} CI.low ${d.bootstrap.low.toFixed(3)} < -${d.tolerance}`).join('; ')}`,\n        )\n      }\n\n      // ── (2) budget gate ─────────────────────────────────────────────\n      const budgetPass =\n        options.budgetUsd === undefined ||\n        ctx.cost.candidate + ctx.cost.baseline <= options.budgetUsd\n      contributing.push({\n        name: 'budget',\n        passed: budgetPass,\n        detail: {\n          candidateUsd: ctx.cost.candidate,\n          baselineUsd: ctx.cost.baseline,\n          budgetUsd: options.budgetUsd,\n        },\n      })\n      if (!budgetPass) {\n        reasons.push(\n          `spend ${(ctx.cost.candidate + ctx.cost.baseline).toFixed(2)} > budget ${options.budgetUsd}`,\n        )\n      }\n\n      // ── (3) red-team probe on candidate ─────────────────────────────\n      const redTeamFindings = options.redTeamBattery\n        ? probeRedTeam(ctx.candidateArtifacts, options.redTeamBattery)\n        : { passed: true, findings: [] }\n      contributing.push({\n        name: 'red-team',\n        passed: redTeamFindings.passed,\n        detail: {\n          failures: redTeamFindings.findings.length,\n          sample: redTeamFindings.findings.slice(0, 3),\n        },\n      })\n      if (!redTeamFindings.passed) {\n        reasons.push(`red-team probe failed (${redTeamFindings.findings.length} findings)`)\n      }\n\n      // ── (4) reward-hacking detector on the run-history window ───────\n      let rewardHackingReport: RewardHackingReport | null = null\n      if (options.recentRuns && options.recentRuns.length >= 10) {\n        rewardHackingReport = detectRewardHacking({ runs: options.recentRuns })\n      }\n      // reward-hacking severity is numeric (0..1). \"gaming\" threshold per\n      // detectRewardHacking defaults = 0.6. Block when ANY finding is at\n      // gaming threshold OR the report verdict is 'gaming'.\n      const gamingThreshold = 0.6\n      const gamingFindings = (rewardHackingReport?.findings ?? []).filter(\n        (f) => f.severity >= gamingThreshold,\n      )\n      const rewardHackingPass =\n        !rewardHackingReport ||\n        !blockOnGaming ||\n        (gamingFindings.length === 0 && rewardHackingReport.verdict !== 'gaming')\n      contributing.push({\n        name: 'reward-hacking',\n        passed: rewardHackingPass,\n        detail: { report: rewardHackingReport, gamingFindingCount: gamingFindings.length },\n      })\n      if (!rewardHackingPass) {\n        reasons.push(\n          `reward-hacking detector flagged ${gamingFindings.length} gaming-severity findings (verdict=${rewardHackingReport!.verdict})`,\n        )\n      }\n\n      // ── (5) canary check on runs ────────────────────────────────────\n      let canaryReport: CanaryReport | null = null\n      if (options.recentRuns && options.recentRuns.length >= 10) {\n        canaryReport = runCanaries(options.recentRuns, {})\n      }\n      // CanarySeverity is 'info' | 'warn' | 'error' — block on 'error'.\n      const errorAlerts = (canaryReport?.alerts ?? []).filter((a) => a.severity === 'error')\n      const canaryPass = errorAlerts.length === 0\n      contributing.push({\n        name: 'canary',\n        passed: canaryPass,\n        detail: { totalAlerts: canaryReport?.alerts.length ?? 0, errorAlerts: errorAlerts.length },\n      })\n      if (!canaryPass) {\n        reasons.push(`canary error alerts: ${errorAlerts.length}`)\n      }\n\n      // ── Verdict ─────────────────────────────────────────────────────\n      const allPassed = contributing.every((c) => c.passed)\n      const decision = allPassed ? 'ship' : 'hold'\n\n      return {\n        decision,\n        reasons: reasons.length > 0 ? reasons : ['all gates passed'],\n        contributingGates: contributing,\n        delta,\n      }\n    },\n  }\n}\n\nfunction probeRedTeam<TArtifact>(\n  artifacts: Map<string, TArtifact>,\n  battery: RedTeamCase[],\n): { passed: boolean; findings: Array<{ scenarioId: string; reason: string }> } {\n  const findings: Array<{ scenarioId: string; reason: string }> = []\n  for (const [_cellId, artifact] of artifacts) {\n    const text = extractText(artifact)\n    if (text === undefined) continue\n    for (const rtCase of battery) {\n      const finding = scoreRedTeamOutput(text, [], rtCase)\n      if (!finding.passed) {\n        findings.push({ scenarioId: rtCase.id, reason: finding.reason ?? 'red-team probe failed' })\n      }\n    }\n  }\n  return { passed: findings.length === 0, findings }\n}\n\nfunction extractText(artifact: unknown): string | undefined {\n  if (typeof artifact === 'string') return artifact\n  if (artifact && typeof artifact === 'object') {\n    const rec = artifact as Record<string, unknown>\n    if (typeof rec.text === 'string') return rec.text\n    if (typeof rec.output === 'string') return rec.output\n    if (typeof rec.content === 'string') return rec.content\n  }\n  return undefined\n}\n","/**\n * Power preflight — \"can this budget detect the effect you are hunting?\"\n *\n * The failure it prevents (measured, twice): a live prompt-improvement campaign ran\n * 333 sandbox cells over 5.6 hours and produced a +0.08 holdout lift the ship gate\n * (paired bootstrap, CI.low > 0.05) could not distinguish from zero — because at\n * that holdout size and worker variance the MINIMUM DETECTABLE lift was larger than\n * any effect a prompt change plausibly produces. The budget was spent learning what\n * a 30-second calculation on the baseline cells already knew. No eval framework we\n * know of surfaces this; every underpowered improvement run everywhere ends in an\n * uninformative \"hold\".\n *\n * Model: the ship rule is `CI.low(paired Δ) > deltaThreshold`. Approximating the\n * bootstrap CI as normal, `CI.low ≈ effect − z·sd_Δ/√n`, so the smallest shippable\n * true effect is `MDE = deltaThreshold + z·sd_Δ/√n`. The paired-delta SD is unknown\n * before the candidate exists; we bound it by the zero-correlation case\n * `sd_Δ ≤ √2·sd_baseline` — a CONSERVATIVE (upper) MDE, which is the correct\n * direction for a warning. Pairing is per cell (`scenario:rep`), so reps multiply n.\n *\n * Standalone by design: feed it any baseline composites (a `gate:'none'` run, a\n * live-proof table) BEFORE budgeting the real search; `selfImprove` also attaches\n * it to every result and warns when the run was structurally unable to ship.\n */\n\nexport interface PowerPreflightOptions {\n  /** Per-cell baseline composites on the HOLDOUT scenarios (one per scenario:rep cell). */\n  baselineComposites: number[]\n  /** Paired observations the budgeted comparison will produce\n   *  (holdout scenarios × reps). Defaults to `baselineComposites.length`. */\n  pairedN?: number\n  /** The ship gate's effect-size threshold. Default 0.05 (defaultProductionGate). */\n  deltaThreshold?: number\n  /** CI confidence the gate uses. Default 0.95. */\n  confidence?: number\n  /** True when the holdout is scored by the SAME judge/scorer family as the gate\n   *  (selfImprove's default composition — one judge scores everything). Under a\n   *  shared channel, raising paired n reduces only the IDIOSYNCRATIC noise share;\n   *  systematic judge bias is untouched, so the MDE here is a lower bound and the\n   *  only full debiaser is an independent second scoring channel\n   *  (recursive-self-improvement S1c, closed form in EXP-023 P0). Default false. */\n  sharedScorerChannel?: boolean\n}\n\nexport interface PowerPreflight {\n  /** Paired observations the comparison will have. */\n  n: number\n  /** Baseline per-cell composite standard deviation (the variance the effect must beat). */\n  sd: number\n  /** Minimum detectable lift: the smallest TRUE effect the gate could ship at this budget. */\n  mde: number\n  /** Baseline holdout composite mean. */\n  baselineMean: number\n  /** Headroom to a perfect 1.0 composite (the largest achievable lift on a [0,1] judge). */\n  headroom: number\n  /** True when even the largest achievable effect (headroom) is below the MDE —\n   *  the run is structurally unable to ship regardless of proposal quality.\n   *  Only asserted for [0,1]-scaled judges (see `scaleAssumed`). */\n  underpowered: boolean\n  /** True when composites look [0,1]-scaled; headroom/underpowered are only\n   *  meaningful under that convention (0-100 judges get mde/sd/n but no verdict). */\n  scaleAssumed: boolean\n  deltaThreshold: number\n  confidence: number\n  /** Set when the holdout shares the gate's scoring channel: more cells cannot\n   *  buy back systematic judge bias — treat the MDE as a lower bound. */\n  sharedChannelCaveat?: string\n  /** One actionable sentence for humans and logs. */\n  recommendation: string\n}\n\n/** Two-sided z for the common confidence levels; interpolation is overkill here. */\nfunction zFor(confidence: number): number {\n  if (confidence >= 0.99) return 2.576\n  if (confidence >= 0.95) return 1.96\n  if (confidence >= 0.9) return 1.645\n  return 1.282\n}\n\n/** Estimate the minimum detectable lift a paired-holdout improvement run can\n *  ship at a given budget, from the baseline holdout composites — call it BEFORE\n *  spending a search to learn whether the effect you are hunting is even\n *  observable at this holdout size and worker variance. */\nexport function powerPreflight(opts: PowerPreflightOptions): PowerPreflight {\n  const composites = opts.baselineComposites.filter((v) => Number.isFinite(v))\n  if (composites.length < 3) {\n    throw new Error(\n      `powerPreflight: need >= 3 finite baseline composites to estimate variance, got ${composites.length}`,\n    )\n  }\n  const deltaThreshold = opts.deltaThreshold ?? 0.05\n  const confidence = opts.confidence ?? 0.95\n  const n = opts.pairedN ?? composites.length\n  if (n < 2) throw new Error(`powerPreflight: pairedN must be >= 2, got ${n}`)\n\n  const mean = composites.reduce((a, b) => a + b, 0) / composites.length\n  const variance =\n    composites.reduce((a, b) => a + (b - mean) * (b - mean), 0) / (composites.length - 1)\n  const sd = Math.sqrt(variance)\n  const z = zFor(confidence)\n  const mde = deltaThreshold + (z * Math.SQRT2 * sd) / Math.sqrt(n)\n\n  const scaleAssumed = composites.every((v) => v >= -0.001 && v <= 1.5)\n  const headroom = Math.max(0, 1 - mean)\n  const underpowered = scaleAssumed && mde > headroom\n\n  const sharedChannelCaveat = opts.sharedScorerChannel\n    ? 'Holdout and gate share one scoring channel: raising n/reps reduces only idiosyncratic noise — systematic judge bias remains and this MDE is a lower bound. Full debiasing needs an independent second scoring channel (different judge/benchmark family).'\n    : undefined\n\n  const recommendation = underpowered\n    ? `UNDERPOWERED: minimum detectable lift ${mde.toFixed(3)} exceeds the ${headroom.toFixed(3)} headroom above the baseline (${mean.toFixed(3)}) — no achievable effect can ship at this budget. Raise paired n (scenarios x reps) to ~${Math.ceil(((z * Math.SQRT2 * sd) / Math.max(headroom - deltaThreshold, 0.01)) ** 2)} or reduce worker variance before searching.`\n    : `Minimum detectable lift at n=${n}: ${mde.toFixed(3)} (baseline sd ${sd.toFixed(3)}). Effects smaller than this cannot clear the gate; budget the search for effects you believe exceed it.`\n\n  return {\n    n,\n    sd,\n    mde,\n    baselineMean: mean,\n    headroom,\n    underpowered,\n    scaleAssumed,\n    deltaThreshold,\n    confidence,\n    ...(sharedChannelCaveat ? { sharedChannelCaveat } : {}),\n    recommendation: sharedChannelCaveat\n      ? `${recommendation} ${sharedChannelCaveat}`\n      : recommendation,\n  }\n}\n","/**\n * Promotion policy over the evidence VECTOR — the substrate's answer to \"never\n * collapse the multi-objective promotion decision into one scalar.\" A\n * `defaultProductionGate` is one opinionated composition; this module factors\n * the decision into two reusable pieces so MANY policies can compete over the\n * SAME evidence (the quant-desk pattern: one evidence bus, plural strategies):\n *\n *   buildEvidenceVector(ctx, objectives, opts) -> EvidenceVector   // the bus\n *   PromotionPolicy = (ev: EvidenceVector) => GateResult           // a strategy\n *   paretoPolicy(ev)                                               // the default strategy\n *   paretoSignificanceGate(options): Gate                          // bus + policy as a Gate\n *\n * The Pareto policy is SYMMETRIC multi-objective: every objective is BOTH a\n * potential gain source AND a safety floor (unlike `defaultProductionGate`,\n * where only `composite` can win and `criticalDimensions` are pure floors). A\n * candidate ships iff it weakly DOMINATES the baseline at the confidence level —\n * no objective credibly worse (CI floor breach) AND at least one objective\n * credibly better (CI gain). Insufficient evidence on ANY axis -> need_more_work\n * (NOT folded into hold: \"gather more reps\" and \"reject\" are different actions).\n *\n * Cost/latency are NOT CI axes here — `GateContext` carries only an aggregate\n * per-side cost, no per-cell observation vector to bootstrap. Treat them as hard\n * constraints (compose with a budget gate via `composeGate`), not faked CIs.\n */\n\nimport type { Direction } from '../../pareto'\nimport { type PairedBootstrapResult, pairedBootstrap } from '../../statistics'\nimport type { Gate, GateContext, GateDecision, GateResult, JudgeScore, Scenario } from '../types'\nimport { detectScale, pairHoldout } from './statistical-heldout'\n\n/** Where an objective's per-cell scalar comes from. `composite` reads the\n *  judge's composite; `dimension` reads a named per-dimension score. */\nexport type ObjectiveSource = { kind: 'composite' } | { kind: 'dimension'; dimension: string }\n\nexport interface PromotionObjective {\n  /** Stable label used in reports + `contributingGates`. */\n  name: string\n  source: ObjectiveSource\n  /** 'maximize' (quality dims) or 'minimize' (error/risk/length dims). Orients\n   *  the paired delta so a positive bootstrap always means \"candidate better\". */\n  direction: Direction\n  /** The good-direction paired-delta CI lower bound must EXCEED this to count\n   *  as a significant gain on this axis. Interpreted in the judge's native\n   *  scale. Default 0 (⇒ \"confidently better\"). */\n  gainThreshold?: number\n  /** A floor breach (regression) is declared when the good-direction CI lower\n   *  bound is below −floorTolerance. When omitted it auto-scales off observed\n   *  magnitudes (0.05 on [0,1], 5 on 0-100), matching `dimensionRegressions`. */\n  floorTolerance?: number\n}\n\n/** Per-axis verdict from the good-direction paired bootstrap. */\nexport type AxisVerdict = 'improved' | 'regressed' | 'flat' | 'few_runs'\n\nexport interface AxisEvidence {\n  name: string\n  source: ObjectiveSource\n  direction: Direction\n  /** Paired bootstrap on the GOOD-DIRECTION delta (oriented by `direction`):\n   *  a positive value means the candidate is better on this axis. */\n  bootstrap: PairedBootstrapResult\n  /** Paired observations contributing to this axis. */\n  n: number\n  gainThreshold: number\n  floorTolerance: number\n  verdict: AxisVerdict\n}\n\nexport interface EvidenceVector {\n  /** One entry per objective — NOTHING averaged across axes. */\n  axes: AxisEvidence[]\n  /** Smallest paired n across axes that produced observations — the binding\n   *  evidence-sufficiency constraint. 0 when no axis produced observations. */\n  minN: number\n  /** Aggregate per-side cost from the gate context (a constraint input, not a\n   *  CI axis — see the module header). */\n  cost: { candidate: number; baseline: number }\n}\n\n/** A promotion strategy: a pure function from the evidence vector to a verdict.\n *  Many policies can run over the same `EvidenceVector` and disagree — that's\n *  the point (competing strategies, shared evidence). */\nexport type PromotionPolicy = (ev: EvidenceVector) => GateResult\n\nexport interface BuildEvidenceVectorOptions {\n  /** Minimum paired observations before an axis can claim significance; below\n   *  it the axis is `few_runs`. Default 3. */\n  minProductiveRuns?: number\n  /** Confidence level for every axis bootstrap. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resamples. Default 2000. */\n  resamples?: number\n  /** Fixed bootstrap seed for a deterministic, reproducible verdict. Default 1337. */\n  seed?: number\n}\n\n/**\n * The Evidence Bus. For each objective, pair candidate vs baseline by full\n * cellId and bootstrap a CI on the good-direction paired delta. Reuses the\n * exact `pairHoldout` + `pairedBootstrap` machinery the held-out gate uses, so\n * a single source of truth governs pairing granularity + scale handling.\n */\nexport function buildEvidenceVector<TArtifact, TScenario extends Scenario>(\n  ctx: GateContext<TArtifact, TScenario>,\n  objectives: PromotionObjective[],\n  opts: BuildEvidenceVectorOptions = {},\n): EvidenceVector {\n  if (objectives.length === 0) {\n    throw new Error('buildEvidenceVector: at least 1 objective required')\n  }\n  const minProductiveRuns = opts.minProductiveRuns ?? 3\n  const confidence = opts.confidence ?? 0.95\n  const resamples = opts.resamples ?? 2000\n  const seed = opts.seed ?? 1337\n  const baseline = ctx.baselineJudgeScores ?? ctx.judgeScores\n  const scenarioIds = new Set(ctx.scenarios.map((s) => s.id))\n\n  const axes: AxisEvidence[] = []\n  for (const obj of objectives) {\n    let select: (s: JudgeScore) => number | undefined\n    if (obj.source.kind === 'composite') {\n      select = (s) => s.composite\n    } else {\n      const dim = obj.source.dimension\n      select = (s) => s.dimensions[dim]\n    }\n    const paired = pairHoldout(ctx.judgeScores, baseline, scenarioIds, select)\n    // Orient to the good direction: maximize ⇒ bootstrap (candidate − baseline);\n    // minimize ⇒ bootstrap (baseline − candidate) by swapping args, so a\n    // positive bootstrap always reads as \"candidate better on this axis\".\n    const before = obj.direction === 'maximize' ? paired.before : paired.after\n    const after = obj.direction === 'maximize' ? paired.after : paired.before\n    const bootstrap = pairedBootstrap(before, after, {\n      confidence,\n      resamples,\n      statistic: 'median',\n      seed,\n    })\n    const n = paired.before.length\n    const floorTolerance =\n      obj.floorTolerance ?? 0.05 * detectScale([...paired.before, ...paired.after])\n    const gainThreshold = obj.gainThreshold ?? 0\n    // Floor check precedes the gain check: a credible regression must never be\n    // masked as \"improved\". With the defaults (gainThreshold 0, positive floor)\n    // the regions are disjoint and order is moot, but a consumer who sets a\n    // negative gainThreshold (\"accept small dips\") could otherwise have a real\n    // floor breach classified as a gain — anti-Goodhart wins the tie.\n    const verdict: AxisVerdict =\n      n < minProductiveRuns\n        ? 'few_runs'\n        : bootstrap.low < -floorTolerance\n          ? 'regressed'\n          : bootstrap.low > gainThreshold\n            ? 'improved'\n            : 'flat'\n    axes.push({\n      name: obj.name,\n      source: obj.source,\n      direction: obj.direction,\n      bootstrap,\n      n,\n      gainThreshold,\n      floorTolerance,\n      verdict,\n    })\n  }\n  const ns = axes.map((a) => a.n).filter((n) => n > 0)\n  const minN = ns.length > 0 ? Math.min(...ns) : 0\n  return { axes, minN, cost: { candidate: ctx.cost.candidate, baseline: ctx.cost.baseline } }\n}\n\n/**\n * The default strategy: symmetric multi-objective Pareto significance. Ship iff\n * the candidate weakly dominates the baseline at the confidence level — no axis\n * credibly worse AND ≥1 axis credibly better. Floor breach on any axis → hold\n * (anti-Goodhart, dominates everything). Insufficient evidence on any axis →\n * need_more_work. Statistically equivalent → hold (never ship noise).\n */\nexport const paretoPolicy: PromotionPolicy = (ev) => {\n  const contributingGates = ev.axes.map((ax) => ({\n    name: `objective:${ax.name}`,\n    passed: ax.verdict === 'improved',\n    detail: {\n      direction: ax.direction,\n      source: ax.source,\n      verdict: ax.verdict,\n      n: ax.n,\n      deltaMedian: ax.bootstrap.median,\n      ciLow: ax.bootstrap.low,\n      ciHigh: ax.bootstrap.high,\n      confidence: ax.bootstrap.confidence,\n      gainThreshold: ax.gainThreshold,\n      floorTolerance: ax.floorTolerance,\n    },\n  }))\n\n  const regressed = ev.axes.filter((a) => a.verdict === 'regressed')\n  const fewRuns = ev.axes.filter((a) => a.verdict === 'few_runs')\n  const improved = ev.axes.filter((a) => a.verdict === 'improved')\n\n  let decision: GateDecision\n  const reasons: string[] = []\n  if (regressed.length > 0) {\n    // Floor breach dominates: a credible regression on ANY axis blocks ship even\n    // if another axis improved. This makes the +gain/−safety false positive\n    // structurally impossible whenever the safety dim is an objective.\n    decision = 'hold'\n    for (const a of regressed) {\n      reasons.push(\n        `objective '${a.name}' regressed: good-direction CI.low ${a.bootstrap.low.toFixed(3)} < -${a.floorTolerance} (n=${a.n})`,\n      )\n    }\n  } else if (fewRuns.length > 0) {\n    // No credible regression on the scored axes, but ≥1 axis lacks the evidence\n    // to claim a gain ⇒ gather more reps, do NOT reject.\n    decision = 'need_more_work'\n    for (const a of fewRuns) {\n      reasons.push(\n        `objective '${a.name}' has only n=${a.n} paired runs — insufficient evidence to claim significance`,\n      )\n    }\n  } else if (improved.length > 0) {\n    // Weakly dominates (no axis worse) AND strictly better on ≥1 axis ⇒ a Pareto\n    // improvement at the confidence level.\n    decision = 'ship'\n    reasons.push(\n      `Pareto improvement at the confidence level: ${improved\n        .map(\n          (a) =>\n            `'${a.name}' +${a.bootstrap.median.toFixed(3)} (CI.low ${a.bootstrap.low.toFixed(3)})`,\n        )\n        .join(', ')}; no objective regressed`,\n    )\n  } else {\n    // Enough evidence, nothing credibly better or worse ⇒ statistically\n    // equivalent. Do NOT ship a no-op.\n    decision = 'hold'\n    reasons.push(\n      'no Pareto improvement: candidate statistically equivalent to baseline on every objective',\n    )\n  }\n\n  // `delta` surfaces the composite axis if present, else the first axis — a\n  // single convenience scalar; the vector lives in `contributingGates`.\n  const composite = ev.axes.find((a) => a.source.kind === 'composite') ?? ev.axes[0]\n  return { decision, reasons, contributingGates, delta: composite?.bootstrap.median }\n}\n\nexport interface ParetoSignificanceGateOptions extends BuildEvidenceVectorOptions {\n  /** The objective vector. Every axis is both a gain source and a safety floor. */\n  objectives: PromotionObjective[]\n  /** Strategy applied to the evidence vector. Default `paretoPolicy`. Override\n   *  to run a stricter/looser strategy over the SAME bus (competing policies). */\n  policy?: PromotionPolicy\n  /** Override the gate name in reports. */\n  name?: string\n}\n\n/**\n * Wrap the bus + a policy as a `Gate`. Plugs into the existing\n * `runImprovementLoop({ gate })` slot and composes via `composeGate`; default\n * loop behavior is unchanged because consumers opt in by passing this gate.\n */\nexport function paretoSignificanceGate<TArtifact = unknown, TScenario extends Scenario = Scenario>(\n  options: ParetoSignificanceGateOptions,\n): Gate<TArtifact, TScenario> {\n  if (options.objectives.length === 0) {\n    throw new Error('paretoSignificanceGate: at least 1 objective required')\n  }\n  const policy = options.policy ?? paretoPolicy\n  return {\n    name: options.name ?? 'paretoSignificanceGate',\n    async decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult> {\n      const ev = buildEvidenceVector(ctx, options.objectives, options)\n      return policy(ev)\n    },\n  }\n}\n","/**\n * `runEval` — the simplest preset over `runCampaign`. No optimizer, no\n * gate, no auto-PR. Just: run scenarios through dispatch, score with\n * judges, return CampaignResult.\n *\n * The 80% case for consumers who want a scorecard, not an improvement loop.\n */\n\nimport { type RunCampaignOptions, runCampaign } from '../run-campaign'\nimport type { CampaignResult, Scenario } from '../types'\n\nexport interface RunEvalOptions<TScenario extends Scenario, TArtifact>\n  extends Omit<RunCampaignOptions<TScenario, TArtifact>, 'runDir'> {\n  runDir: string\n}\n\n/**\n * Simplest evaluation preset: run scenarios through dispatch, score with judges, and return a `CampaignResult` — no optimizer, no gate, no PR.\n */\nexport async function runEval<TScenario extends Scenario, TArtifact>(\n  opts: RunEvalOptions<TScenario, TArtifact>,\n): Promise<CampaignResult<TArtifact, TScenario>> {\n  return runCampaign(opts)\n}\n","/**\n * `evolutionaryProposer` — adapts a stateless `Mutator` (population mutation:\n * GEPA / AxGEPA / reflective-mutation) into a `SurfaceProposer`. This is\n * the evolutionary strategy: each generation, mutate the current best surface\n * into N candidates, measure, select. No generation memory beyond the current\n * surface; the loop body handles ranking + promotion.\n *\n * The reflective alternative is agent-runtime's runtime proposer with a\n * `reflectiveGenerator` / `agenticGenerator`: it reasons over the report +\n * trace findings to propose targeted edits rather than blind mutations. Both\n * conform to `SurfaceProposer`; the improvement loop is identical either way.\n */\n\nimport type { Mutator, SurfaceProposer } from '../types'\n\nexport interface EvolutionaryProposerOptions<TFindings = unknown> {\n  mutator: Mutator<TFindings>\n  /** External findings fed to the mutator each generation. Default: []. */\n  findings?: TFindings[]\n}\n\n/**\n * Wrap a stateless `Mutator` (GEPA, AxGEPA, reflective-mutation) as a `SurfaceProposer` that mutates the current best surface into N candidates each generation.\n */\nexport function evolutionaryProposer<TFindings = unknown>(\n  opts: EvolutionaryProposerOptions<TFindings>,\n): SurfaceProposer<TFindings> {\n  return {\n    kind: `evolutionary:${opts.mutator.kind}`,\n    async propose({ currentSurface, findings, populationSize, signal }) {\n      return opts.mutator.mutate({\n        findings: findings.length > 0 ? findings : (opts.findings ?? []),\n        currentSurface,\n        populationSize,\n        signal,\n      })\n    },\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAYO,SAAS,eACX,OACyB;AAC5B,MAAI,MAAM,WAAW,GAAG;AACtB,UAAM,IAAI,MAAM,wCAAwC;AAAA,EAC1D;AACA,SAAO;AAAA,IACL,MAAM,YAAY,MAAM,IAAI,CAAC,MAAM,EAAE,IAAI,EAAE,KAAK,GAAG,CAAC;AAAA,IACpD,MAAM,OAAO,KAA6D;AACxE,YAAM,UAAwE,CAAC;AAC/E,iBAAW,QAAQ,OAAO;AACxB,cAAM,MAAM,MAAM,KAAK,OAAO,GAAG;AACjC,gBAAQ,KAAK,EAAE,MAAM,IAAI,CAAC;AAAA,MAC5B;AAQA,YAAM,YAAY,QAAQ,IAAI,CAAC,MAAM,EAAE,IAAI,QAAQ;AACnD,YAAM,UAAwB,UAAU,MAAM,CAAC,MAAM,MAAM,MAAM,IAC7D,SACA,UAAU,SAAS,cAAc,IAC/B,iBACA,UAAU,SAAS,eAAe,IAChC,kBACA,UAAU,SAAS,MAAM,IACvB,SACA;AAEV,YAAM,eAAe,QAAQ;AAAA,QAAQ,CAAC,MACpC,EAAE,IAAI,kBAAkB,SAAS,IAC7B,EAAE,IAAI,oBACN,CAAC,EAAE,MAAM,EAAE,KAAK,MAAM,QAAQ,EAAE,IAAI,aAAa,QAAQ,QAAQ,EAAE,IAAI,CAAC;AAAA,MAC9E;AAEA,YAAM,UAAU,QAAQ;AAAA,QAAQ,CAAC,MAC/B,EAAE,IAAI,QAAQ,IAAI,CAAC,WAAW,IAAI,EAAE,KAAK,IAAI,KAAK,MAAM,EAAE;AAAA,MAC5D;AAEA,aAAO;AAAA,QACL,UAAU;AAAA,QACV;AAAA,QACA,mBAAmB;AAAA,QACnB,OAAO,QAAQ,CAAC,GAAG,IAAI;AAAA,MACzB;AAAA,IACF;AAAA,EACF;AACF;;;ACeO,SAAS,sBACd,SAC4B;AAC5B,QAAM,iBAAiB,QAAQ,kBAAkB;AACjD,QAAM,aAAa,QAAQ,cAAc;AACzC,QAAM,YAAY,QAAQ,sBAAsB;AAChD,QAAM,OAAO,QAAQ,iBAAiB;AACtC,QAAM,oBAAoB,QAAQ,qBAAqB;AACvD,QAAM,mBAAmB,QAAQ,oBAAoB;AACrD,QAAM,gBAAgB,QAAQ,8BAA8B;AAE5D,SAAO;AAAA,IACL,MAAM;AAAA,IACN,MAAM,OAAO,KAA6D;AACxE,YAAM,UAAoB,CAAC;AAC3B,YAAM,eAA0E,CAAC;AASjF,YAAM,cAAc,IAAI,IAAI,QAAQ,iBAAiB,IAAI,CAAC,MAAM,EAAE,EAAE,CAAC;AACrE,YAAM,MAAM;AAAA,QACV;AAAA,UACE,IAAI;AAAA,UACJ,IAAI,uBAAuB,IAAI;AAAA,UAC/B;AAAA,UACA,CAAC,MAAM,EAAE;AAAA,QACX;AAAA,QACA;AAAA,UACE;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,WAAW;AAAA,QACb;AAAA,MACF;AAGA,YAAM,QAAQ,qBAAqB,WAAW,IAAI,UAAU,SAAS,IAAI,UAAU;AACnF,YAAM,cAAc,IAAI;AACxB,mBAAa,KAAK;AAAA,QAChB,MAAM;AAAA,QACN,QAAQ;AAAA,QACR,QAAQ;AAAA,UACN,GAAG,IAAI;AAAA,UACP;AAAA,UACA,WAAW,IAAI,UAAU;AAAA,UACzB,uBAAuB,IAAI,gBAAgB;AAAA;AAAA;AAAA,UAG3C,aAAa,IAAI,gBAAgB;AAAA,UACjC,aAAa,IAAI;AAAA,UACjB,OAAO,IAAI,UAAU;AAAA,UACrB,QAAQ,IAAI,UAAU;AAAA,UACtB,YAAY,IAAI,UAAU;AAAA,UAC1B;AAAA,UACA,SAAS,IAAI;AAAA,QACf;AAAA,MACF,CAAC;AACD,UAAI,CAAC,aAAa;AAChB,cAAM,UACJ,IAAI,eAAe,oBACf,MAAM,IAAI,cAAc,KAAK,QAAQ,CAAC,CAAC,qBACvC;AACN,gBAAQ;AAAA,UACN,IAAI,UACA,kBAAkB,IAAI,CAAC,mBAAmB,iBAAiB,2CAC3D,mBAAmB,IAAI,UAAU,IAAI,QAAQ,CAAC,CAAC,qBAAgB,cAAc,KAAK,gBAAgB,WAAM,MAAM,QAAQ,CAAC,CAAC,MAAM,IAAI,UAAU,aAAa,KAAK,QAAQ,CAAC,CAAC,SAAS,IAAI,UAAU,IAAI,QAAQ,CAAC,CAAC,KAAK,IAAI,UAAU,KAAK,QAAQ,CAAC,CAAC,IAAI,OAAO;AAAA,QAChQ;AAAA,MACF;AAQA,YAAM,UAAU,QAAQ,oBAAoB,SACxC;AAAA,QACE,IAAI;AAAA,QACJ,IAAI,uBAAuB,IAAI;AAAA,QAC/B;AAAA,QACA,QAAQ;AAAA,QACR,EAAE,WAAW,QAAQ,qBAAqB,YAAY,WAAW,KAAK;AAAA,MACxE,IACA,CAAC;AACL,YAAM,YAAY,QAAQ,OAAO,CAAC,MAAM,EAAE,SAAS;AACnD,YAAM,UAAU,UAAU,WAAW;AACrC,mBAAa,KAAK;AAAA,QAChB,MAAM;AAAA,QACN,QAAQ;AAAA,QACR,QAAQ;AAAA,UACN,SAAS,QAAQ,sBAAsB,CAAC;AAAA,UACxC,aAAa,QAAQ,IAAI,CAAC,OAAO;AAAA,YAC/B,WAAW,EAAE;AAAA,YACb,OAAO,EAAE,UAAU;AAAA,YACnB,QAAQ,EAAE,UAAU;AAAA,YACpB,WAAW,EAAE;AAAA,YACb,GAAG,EAAE;AAAA,YACL,WAAW,EAAE;AAAA,UACf,EAAE;AAAA,QACJ;AAAA,MACF,CAAC;AACD,UAAI,CAAC,SAAS;AACZ,gBAAQ;AAAA,UACN,oCAAoC,UAAU,IAAI,CAAC,MAAM,GAAG,EAAE,SAAS,WAAW,EAAE,UAAU,IAAI,QAAQ,CAAC,CAAC,OAAO,EAAE,SAAS,EAAE,EAAE,KAAK,IAAI,CAAC;AAAA,QAC9I;AAAA,MACF;AAGA,YAAM,aACJ,QAAQ,cAAc,UACtB,IAAI,KAAK,YAAY,IAAI,KAAK,YAAY,QAAQ;AACpD,mBAAa,KAAK;AAAA,QAChB,MAAM;AAAA,QACN,QAAQ;AAAA,QACR,QAAQ;AAAA,UACN,cAAc,IAAI,KAAK;AAAA,UACvB,aAAa,IAAI,KAAK;AAAA,UACtB,WAAW,QAAQ;AAAA,QACrB;AAAA,MACF,CAAC;AACD,UAAI,CAAC,YAAY;AACf,gBAAQ;AAAA,UACN,UAAU,IAAI,KAAK,YAAY,IAAI,KAAK,UAAU,QAAQ,CAAC,CAAC,aAAa,QAAQ,SAAS;AAAA,QAC5F;AAAA,MACF;AAGA,YAAM,kBAAkB,QAAQ,iBAC5B,aAAa,IAAI,oBAAoB,QAAQ,cAAc,IAC3D,EAAE,QAAQ,MAAM,UAAU,CAAC,EAAE;AACjC,mBAAa,KAAK;AAAA,QAChB,MAAM;AAAA,QACN,QAAQ,gBAAgB;AAAA,QACxB,QAAQ;AAAA,UACN,UAAU,gBAAgB,SAAS;AAAA,UACnC,QAAQ,gBAAgB,SAAS,MAAM,GAAG,CAAC;AAAA,QAC7C;AAAA,MACF,CAAC;AACD,UAAI,CAAC,gBAAgB,QAAQ;AAC3B,gBAAQ,KAAK,0BAA0B,gBAAgB,SAAS,MAAM,YAAY;AAAA,MACpF;AAGA,UAAI,sBAAkD;AACtD,UAAI,QAAQ,cAAc,QAAQ,WAAW,UAAU,IAAI;AACzD,8BAAsB,oBAAoB,EAAE,MAAM,QAAQ,WAAW,CAAC;AAAA,MACxE;AAIA,YAAM,kBAAkB;AACxB,YAAM,kBAAkB,qBAAqB,YAAY,CAAC,GAAG;AAAA,QAC3D,CAAC,MAAM,EAAE,YAAY;AAAA,MACvB;AACA,YAAM,oBACJ,CAAC,uBACD,CAAC,iBACA,eAAe,WAAW,KAAK,oBAAoB,YAAY;AAClE,mBAAa,KAAK;AAAA,QAChB,MAAM;AAAA,QACN,QAAQ;AAAA,QACR,QAAQ,EAAE,QAAQ,qBAAqB,oBAAoB,eAAe,OAAO;AAAA,MACnF,CAAC;AACD,UAAI,CAAC,mBAAmB;AACtB,gBAAQ;AAAA,UACN,mCAAmC,eAAe,MAAM,sCAAsC,oBAAqB,OAAO;AAAA,QAC5H;AAAA,MACF;AAGA,UAAI,eAAoC;AACxC,UAAI,QAAQ,cAAc,QAAQ,WAAW,UAAU,IAAI;AACzD,uBAAe,YAAY,QAAQ,YAAY,CAAC,CAAC;AAAA,MACnD;AAEA,YAAM,eAAe,cAAc,UAAU,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,aAAa,OAAO;AACrF,YAAM,aAAa,YAAY,WAAW;AAC1C,mBAAa,KAAK;AAAA,QAChB,MAAM;AAAA,QACN,QAAQ;AAAA,QACR,QAAQ,EAAE,aAAa,cAAc,OAAO,UAAU,GAAG,aAAa,YAAY,OAAO;AAAA,MAC3F,CAAC;AACD,UAAI,CAAC,YAAY;AACf,gBAAQ,KAAK,wBAAwB,YAAY,MAAM,EAAE;AAAA,MAC3D;AAGA,YAAM,YAAY,aAAa,MAAM,CAAC,MAAM,EAAE,MAAM;AACpD,YAAM,WAAW,YAAY,SAAS;AAEtC,aAAO;AAAA,QACL;AAAA,QACA,SAAS,QAAQ,SAAS,IAAI,UAAU,CAAC,kBAAkB;AAAA,QAC3D,mBAAmB;AAAA,QACnB;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACF;AAEA,SAAS,aACP,WACA,SAC8E;AAC9E,QAAM,WAA0D,CAAC;AACjE,aAAW,CAAC,SAAS,QAAQ,KAAK,WAAW;AAC3C,UAAM,OAAO,YAAY,QAAQ;AACjC,QAAI,SAAS,OAAW;AACxB,eAAW,UAAU,SAAS;AAC5B,YAAM,UAAU,mBAAmB,MAAM,CAAC,GAAG,MAAM;AACnD,UAAI,CAAC,QAAQ,QAAQ;AACnB,iBAAS,KAAK,EAAE,YAAY,OAAO,IAAI,QAAQ,QAAQ,UAAU,wBAAwB,CAAC;AAAA,MAC5F;AAAA,IACF;AAAA,EACF;AACA,SAAO,EAAE,QAAQ,SAAS,WAAW,GAAG,SAAS;AACnD;AAEA,SAAS,YAAY,UAAuC;AAC1D,MAAI,OAAO,aAAa,SAAU,QAAO;AACzC,MAAI,YAAY,OAAO,aAAa,UAAU;AAC5C,UAAM,MAAM;AACZ,QAAI,OAAO,IAAI,SAAS,SAAU,QAAO,IAAI;AAC7C,QAAI,OAAO,IAAI,WAAW,SAAU,QAAO,IAAI;AAC/C,QAAI,OAAO,IAAI,YAAY,SAAU,QAAO,IAAI;AAAA,EAClD;AACA,SAAO;AACT;;;AChPA,SAAS,KAAK,YAA4B;AACxC,MAAI,cAAc,KAAM,QAAO;AAC/B,MAAI,cAAc,KAAM,QAAO;AAC/B,MAAI,cAAc,IAAK,QAAO;AAC9B,SAAO;AACT;AAMO,SAAS,eAAe,MAA6C;AAC1E,QAAM,aAAa,KAAK,mBAAmB,OAAO,CAAC,MAAM,OAAO,SAAS,CAAC,CAAC;AAC3E,MAAI,WAAW,SAAS,GAAG;AACzB,UAAM,IAAI;AAAA,MACR,kFAAkF,WAAW,MAAM;AAAA,IACrG;AAAA,EACF;AACA,QAAM,iBAAiB,KAAK,kBAAkB;AAC9C,QAAM,aAAa,KAAK,cAAc;AACtC,QAAM,IAAI,KAAK,WAAW,WAAW;AACrC,MAAI,IAAI,EAAG,OAAM,IAAI,MAAM,6CAA6C,CAAC,EAAE;AAE3E,QAAM,OAAO,WAAW,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,WAAW;AAChE,QAAM,WACJ,WAAW,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,SAAS,IAAI,OAAO,CAAC,KAAK,WAAW,SAAS;AACrF,QAAM,KAAK,KAAK,KAAK,QAAQ;AAC7B,QAAM,IAAI,KAAK,UAAU;AACzB,QAAM,MAAM,iBAAkB,IAAI,KAAK,QAAQ,KAAM,KAAK,KAAK,CAAC;AAEhE,QAAM,eAAe,WAAW,MAAM,CAAC,MAAM,KAAK,SAAU,KAAK,GAAG;AACpE,QAAM,WAAW,KAAK,IAAI,GAAG,IAAI,IAAI;AACrC,QAAM,eAAe,gBAAgB,MAAM;AAE3C,QAAM,sBAAsB,KAAK,sBAC7B,mQACA;AAEJ,QAAM,iBAAiB,eACnB,yCAAyC,IAAI,QAAQ,CAAC,CAAC,gBAAgB,SAAS,QAAQ,CAAC,CAAC,iCAAiC,KAAK,QAAQ,CAAC,CAAC,gGAA2F,KAAK,MAAO,IAAI,KAAK,QAAQ,KAAM,KAAK,IAAI,WAAW,gBAAgB,IAAI,MAAM,CAAC,CAAC,iDACxT,gCAAgC,CAAC,KAAK,IAAI,QAAQ,CAAC,CAAC,iBAAiB,GAAG,QAAQ,CAAC,CAAC;AAEtF,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,GAAI,sBAAsB,EAAE,oBAAoB,IAAI,CAAC;AAAA,IACrD,gBAAgB,sBACZ,GAAG,cAAc,IAAI,mBAAmB,KACxC;AAAA,EACN;AACF;;;AC1BO,SAAS,oBACd,KACA,YACA,OAAmC,CAAC,GACpB;AAChB,MAAI,WAAW,WAAW,GAAG;AAC3B,UAAM,IAAI,MAAM,oDAAoD;AAAA,EACtE;AACA,QAAM,oBAAoB,KAAK,qBAAqB;AACpD,QAAM,aAAa,KAAK,cAAc;AACtC,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,WAAW,IAAI,uBAAuB,IAAI;AAChD,QAAM,cAAc,IAAI,IAAI,IAAI,UAAU,IAAI,CAAC,MAAM,EAAE,EAAE,CAAC;AAE1D,QAAM,OAAuB,CAAC;AAC9B,aAAW,OAAO,YAAY;AAC5B,QAAI;AACJ,QAAI,IAAI,OAAO,SAAS,aAAa;AACnC,eAAS,CAAC,MAAM,EAAE;AAAA,IACpB,OAAO;AACL,YAAM,MAAM,IAAI,OAAO;AACvB,eAAS,CAAC,MAAM,EAAE,WAAW,GAAG;AAAA,IAClC;AACA,UAAM,SAAS,YAAY,IAAI,aAAa,UAAU,aAAa,MAAM;AAIzE,UAAM,SAAS,IAAI,cAAc,aAAa,OAAO,SAAS,OAAO;AACrE,UAAM,QAAQ,IAAI,cAAc,aAAa,OAAO,QAAQ,OAAO;AACnE,UAAM,YAAY,gBAAgB,QAAQ,OAAO;AAAA,MAC/C;AAAA,MACA;AAAA,MACA,WAAW;AAAA,MACX;AAAA,IACF,CAAC;AACD,UAAM,IAAI,OAAO,OAAO;AACxB,UAAM,iBACJ,IAAI,kBAAkB,OAAO,YAAY,CAAC,GAAG,OAAO,QAAQ,GAAG,OAAO,KAAK,CAAC;AAC9E,UAAM,gBAAgB,IAAI,iBAAiB;AAM3C,UAAM,UACJ,IAAI,oBACA,aACA,UAAU,MAAM,CAAC,iBACf,cACA,UAAU,MAAM,gBACd,aACA;AACV,SAAK,KAAK;AAAA,MACR,MAAM,IAAI;AAAA,MACV,QAAQ,IAAI;AAAA,MACZ,WAAW,IAAI;AAAA,MACf;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAAA,EACH;AACA,QAAM,KAAK,KAAK,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,OAAO,CAAC,MAAM,IAAI,CAAC;AACnD,QAAM,OAAO,GAAG,SAAS,IAAI,KAAK,IAAI,GAAG,EAAE,IAAI;AAC/C,SAAO,EAAE,MAAM,MAAM,MAAM,EAAE,WAAW,IAAI,KAAK,WAAW,UAAU,IAAI,KAAK,SAAS,EAAE;AAC5F;AASO,IAAM,eAAgC,CAAC,OAAO;AACnD,QAAM,oBAAoB,GAAG,KAAK,IAAI,CAAC,QAAQ;AAAA,IAC7C,MAAM,aAAa,GAAG,IAAI;AAAA,IAC1B,QAAQ,GAAG,YAAY;AAAA,IACvB,QAAQ;AAAA,MACN,WAAW,GAAG;AAAA,MACd,QAAQ,GAAG;AAAA,MACX,SAAS,GAAG;AAAA,MACZ,GAAG,GAAG;AAAA,MACN,aAAa,GAAG,UAAU;AAAA,MAC1B,OAAO,GAAG,UAAU;AAAA,MACpB,QAAQ,GAAG,UAAU;AAAA,MACrB,YAAY,GAAG,UAAU;AAAA,MACzB,eAAe,GAAG;AAAA,MAClB,gBAAgB,GAAG;AAAA,IACrB;AAAA,EACF,EAAE;AAEF,QAAM,YAAY,GAAG,KAAK,OAAO,CAAC,MAAM,EAAE,YAAY,WAAW;AACjE,QAAM,UAAU,GAAG,KAAK,OAAO,CAAC,MAAM,EAAE,YAAY,UAAU;AAC9D,QAAM,WAAW,GAAG,KAAK,OAAO,CAAC,MAAM,EAAE,YAAY,UAAU;AAE/D,MAAI;AACJ,QAAM,UAAoB,CAAC;AAC3B,MAAI,UAAU,SAAS,GAAG;AAIxB,eAAW;AACX,eAAW,KAAK,WAAW;AACzB,cAAQ;AAAA,QACN,cAAc,EAAE,IAAI,sCAAsC,EAAE,UAAU,IAAI,QAAQ,CAAC,CAAC,OAAO,EAAE,cAAc,OAAO,EAAE,CAAC;AAAA,MACvH;AAAA,IACF;AAAA,EACF,WAAW,QAAQ,SAAS,GAAG;AAG7B,eAAW;AACX,eAAW,KAAK,SAAS;AACvB,cAAQ;AAAA,QACN,cAAc,EAAE,IAAI,gBAAgB,EAAE,CAAC;AAAA,MACzC;AAAA,IACF;AAAA,EACF,WAAW,SAAS,SAAS,GAAG;AAG9B,eAAW;AACX,YAAQ;AAAA,MACN,+CAA+C,SAC5C;AAAA,QACC,CAAC,MACC,IAAI,EAAE,IAAI,MAAM,EAAE,UAAU,OAAO,QAAQ,CAAC,CAAC,YAAY,EAAE,UAAU,IAAI,QAAQ,CAAC,CAAC;AAAA,MACvF,EACC,KAAK,IAAI,CAAC;AAAA,IACf;AAAA,EACF,OAAO;AAGL,eAAW;AACX,YAAQ;AAAA,MACN;AAAA,IACF;AAAA,EACF;AAIA,QAAM,YAAY,GAAG,KAAK,KAAK,CAAC,MAAM,EAAE,OAAO,SAAS,WAAW,KAAK,GAAG,KAAK,CAAC;AACjF,SAAO,EAAE,UAAU,SAAS,mBAAmB,OAAO,WAAW,UAAU,OAAO;AACpF;AAiBO,SAAS,uBACd,SAC4B;AAC5B,MAAI,QAAQ,WAAW,WAAW,GAAG;AACnC,UAAM,IAAI,MAAM,uDAAuD;AAAA,EACzE;AACA,QAAM,SAAS,QAAQ,UAAU;AACjC,SAAO;AAAA,IACL,MAAM,QAAQ,QAAQ;AAAA,IACtB,MAAM,OAAO,KAA6D;AACxE,YAAM,KAAK,oBAAoB,KAAK,QAAQ,YAAY,OAAO;AAC/D,aAAO,OAAO,EAAE;AAAA,IAClB;AAAA,EACF;AACF;;;AClQA,eAAsB,QACpB,MAC+C;AAC/C,SAAO,YAAY,IAAI;AACzB;;;ACCO,SAAS,qBACd,MAC4B;AAC5B,SAAO;AAAA,IACL,MAAM,gBAAgB,KAAK,QAAQ,IAAI;AAAA,IACvC,MAAM,QAAQ,EAAE,gBAAgB,UAAU,gBAAgB,OAAO,GAAG;AAClE,aAAO,KAAK,QAAQ,OAAO;AAAA,QACzB,UAAU,SAAS,SAAS,IAAI,WAAY,KAAK,YAAY,CAAC;AAAA,QAC9D;AAAA,QACA;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AACF;","names":[]}

package/dist/contract/index.d.ts

@@ -1,7 +1,7 @@
 import { S as Scenario, M as MutableSurface, b as DispatchContext, a as JudgeConfig, f as SurfaceProposer, g as Gate, L as LabeledScenarioStore, C as CampaignResult, h as GateDecision } from '../types-Ctf7XIAL.js';
 export { i as CampaignAggregates, j as CampaignArtifactWriter, k as CampaignCellResult, l as CampaignCostMeter, d as CampaignTraceWriter, m as CodeSurface, D as Dispatch, n as GateContext, G as GateResult, o as GenerationCandidate, e as GenerationRecord, c as JudgeDimension, J as JudgeScore, p as Mutator, O as OptimizationProposer, q as OptimizerConfig, r as SessionScript } from '../types-Ctf7XIAL.js';
-import { L as LoopProvenanceRecord, P as PowerPreflight, R as RunEvalOptions } from '../provenance-BGPBEFoV.js';
-export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, D as DefaultProductionGateOptions, E as EvidenceVector, b as EvolutionaryProposerOptions, H as HeldOutGateOptions, O as ObjectiveSource, c as ParetoSignificanceGateOptions, d as PromotionObjective, e as PromotionPolicy, f as buildEvidenceVector, g as composeGate, h as defaultProductionGate, i as evolutionaryProposer, j as heldOutGate, p as paretoPolicy, k as paretoSignificanceGate, r as runEval } from '../provenance-BGPBEFoV.js';
+import { L as LoopProvenanceRecord, P as PowerPreflight, R as RunEvalOptions } from '../provenance-Bsyjc67Z.js';
+export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, D as DefaultProductionGateOptions, E as EvidenceVector, b as EvolutionaryProposerOptions, H as HeldOutGateOptions, O as ObjectiveSource, c as ParetoSignificanceGateOptions, d as PromotionObjective, e as PromotionPolicy, f as buildEvidenceVector, g as composeGate, h as defaultProductionGate, i as evolutionaryProposer, j as heldOutGate, p as paretoPolicy, k as paretoSignificanceGate, r as runEval } from '../provenance-Bsyjc67Z.js';
 import { C as CampaignStorage, R as RunOptimizationOptions, a as RunImprovementLoopResult } from '../gepa-DQ3ruj18.js';
 export { G as GepaProposerOptions, b as RunCampaignOptions, c as RunImprovementLoopOptions, f as fsCampaignStorage, g as gepaProposer, i as inMemoryCampaignStorage, r as runCampaign, d as runImprovementLoop } from '../gepa-DQ3ruj18.js';
 export { D as DeploymentOutcome, F as FileSystemOutcomeStore, a as FileSystemOutcomeStoreOptions, I as InMemoryOutcomeStore, b as OutcomeStore } from '../outcome-store-rnXLEqSn.js';

package/dist/contract/index.js

@@ -19,7 +19,7 @@ import {
   paretoSignificanceGate,
   powerPreflight,
   runEval
-} from "../chunk-IA3X6CSF.js";
+} from "../chunk-J22CKVXN.js";
 import {
   analyzeRuns
 } from "../chunk-GDZAWO2I.js";

package/dist/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "@tangle-network/agent-eval — wire protocol",
-    "version": "0.106.1",
+    "version": "0.106.3",
     "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/{provenance-BGPBEFoV.d.ts → provenance-Bsyjc67Z.d.ts}

@@ -181,6 +181,10 @@ interface PowerPreflight {
     /** One actionable sentence for humans and logs. */
     recommendation: string;
 }
+/** Estimate the minimum detectable lift a paired-holdout improvement run can
+ *  ship at a given budget, from the baseline holdout composites — call it BEFORE
+ *  spending a search to learn whether the effect you are hunting is even
+ *  observable at this holdout size and worker variance. */
 declare function powerPreflight(opts: PowerPreflightOptions): PowerPreflight;
 
 /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-eval",
-  "version": "0.106.1",
+  "version": "0.106.3",
   "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": {

package/dist/chunk-IA3X6CSF.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/campaign/gates/compose.ts","../src/campaign/gates/default-production-gate.ts","../src/campaign/gates/power-preflight.ts","../src/campaign/gates/promotion-policy.ts","../src/campaign/presets/run-eval.ts","../src/campaign/proposers/evolutionary.ts"],"sourcesContent":["/**\n * Compose multiple `Gate` implementations — every gate must pass for the\n * composite to ship. Closes the alignment reviewer's \"default-only\n * heldOutGate + costGate would happily promote a reward-hacked prompt\"\n * concern by making safety gates first-class composable defaults.\n */\n\nimport type { Gate, GateContext, GateDecision, GateResult, Scenario } from '../types'\n\n/** Compose gates — all must `ship` for the composite to `ship`. First\n *  non-ship verdict short-circuits the composite verdict, but ALL gates run\n *  (so the result records every gate's reason — useful for diagnostics). */\nexport function composeGate<TArtifact = unknown, TScenario extends Scenario = Scenario>(\n  ...gates: Array<Gate<TArtifact, TScenario>>\n): Gate<TArtifact, TScenario> {\n  if (gates.length === 0) {\n    throw new Error('composeGate requires at least one gate')\n  }\n  return {\n    name: `composed(${gates.map((g) => g.name).join(',')})`,\n    async decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult> {\n      const results: Array<{ gate: Gate<TArtifact, TScenario>; res: GateResult }> = []\n      for (const gate of gates) {\n        const res = await gate.decide(ctx)\n        results.push({ gate, res })\n      }\n\n      // Substrate-wide verdict policy:\n      //   - all 'ship' → 'ship'\n      //   - any 'arch_ceiling' → 'arch_ceiling' (architectural ceiling beats other holds)\n      //   - any 'model_ceiling' → 'model_ceiling'\n      //   - any 'hold' → 'hold'\n      //   - else 'need_more_work'\n      const decisions = results.map((r) => r.res.decision)\n      const overall: GateDecision = decisions.every((d) => d === 'ship')\n        ? 'ship'\n        : decisions.includes('arch_ceiling')\n          ? 'arch_ceiling'\n          : decisions.includes('model_ceiling')\n            ? 'model_ceiling'\n            : decisions.includes('hold')\n              ? 'hold'\n              : 'need_more_work'\n\n      const contributing = results.flatMap((r) =>\n        r.res.contributingGates.length > 0\n          ? r.res.contributingGates\n          : [{ name: r.gate.name, passed: r.res.decision === 'ship', detail: r.res }],\n      )\n\n      const reasons = results.flatMap((r) =>\n        r.res.reasons.map((reason) => `[${r.gate.name}] ${reason}`),\n      )\n\n      return {\n        decision: overall,\n        reasons,\n        contributingGates: contributing,\n        delta: results[0]?.res.delta,\n      }\n    },\n  }\n}\n","/**\n * `defaultProductionGate` — composes the substrate's existing safety\n * primitives (red-team / reward-hacking / canary / heldout) into a single\n * Gate.decide shape. Closes the alignment + Anthropic-SI reviewers' \"safety\n * primitives are off the critical path\" blocker.\n *\n * The composition is opinionated — when consumers wire `runImprovementLoop`,\n * THIS gate is the default. Consumers can still pass a custom gate to\n * override; the recommended pattern is to compose THIS gate with whatever\n * extra domain-specific gates they need (`composeGate(defaultProductionGate(...), customGate)`).\n */\n\nimport type { CanaryReport } from '../../canary'\nimport { runCanaries } from '../../canary'\nimport type { RedTeamCase } from '../../red-team'\nimport { scoreRedTeamOutput } from '../../red-team'\nimport type { RewardHackingReport } from '../../rl/reward-hacking'\nimport { detectRewardHacking } from '../../rl/reward-hacking'\nimport type { RunRecord } from '../../run-record'\nimport type { Gate, GateContext, GateResult, Scenario } from '../types'\nimport {\n  dimensionRegressions,\n  heldoutSignificance,\n  pairHoldout,\n  TIE_WARN_FRACTION,\n} from './statistical-heldout'\n\nexport interface DefaultProductionGateOptions {\n  /** Required: scenarios held out from training; substrate compares\n   *  candidate-on-holdout vs baseline-on-holdout. */\n  holdoutScenarios: Scenario[]\n  /** Minimum held-out lift the **paired-bootstrap CI lower bound** must clear\n   *  to ship — NOT a point estimate. Default 0 ⇒ \"confidently positive at the\n   *  confidence level\". Interpreted in the judge's native composite scale (set\n   *  e.g. 2 for a 0-100 rubric to require a ≥2-point significant gain). */\n  deltaThreshold?: number\n  /** Confidence level for the held-out + dimension bootstraps. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resamples. Default 2000. */\n  bootstrapResamples?: number\n  /** Fixed bootstrap seed for a deterministic verdict. Default 1337. */\n  bootstrapSeed?: number\n  /** Minimum paired holdout observations (scenarios × reps) before a\n   *  significance claim is allowed; below it the gate HOLDS with `few_runs`\n   *  rather than reading a degenerate CI. Default 3. */\n  minProductiveRuns?: number\n  /** Ship statistic for the held-out significance test. Default `'mean'`\n   *  (tie-robust — see `heldoutSignificance`). Pass `'median'` for\n   *  outlier-robustness at the cost of tie-blindness. */\n  heldoutStatistic?: 'mean' | 'median'\n  /** Critical judge dimensions that must NOT significantly regress even when\n   *  the net composite rises (anti-Goodhart). The gate HOLDS if any listed\n   *  dimension's paired-delta CI lower bound < −`regressionTolerance`. E.g.\n   *  `['hallucination_free']` for a legal agent. */\n  criticalDimensions?: string[]\n  /** Tolerance for the per-dimension regression guard, in the dimension's\n   *  native scale. When omitted it auto-scales off observed magnitudes:\n   *  0.05 on [0,1], 5 on 0-100. */\n  regressionTolerance?: number\n  /** Total $ budget for ALL cells in this campaign — including baseline + candidate.\n   *  Composite verdict refuses to ship when spend exceeded budget. */\n  budgetUsd?: number\n  /** Red-team cases to probe candidate outputs against. When omitted the\n   *  substrate uses `DEFAULT_RED_TEAM_CORPUS`. Provide a domain-specific\n   *  battery for tighter coverage. */\n  redTeamBattery?: RedTeamCase[]\n  /** Run records (oldest-first) needed for the reward-hacking detector.\n   *  Substrate populates from prior production-loop generations. */\n  recentRuns?: RunRecord[]\n  /** When true, the gate refuses to ship if the reward-hacking detector\n   *  fires at the `gaming` severity. Default true. */\n  blockOnRewardHackingGaming?: boolean\n}\n\n/**\n * Opinionated production gate composing held-out significance, red-team, reward-hacking, and canary checks into a single `Gate.decide` decision.\n */\nexport function defaultProductionGate<TArtifact, TScenario extends Scenario>(\n  options: DefaultProductionGateOptions,\n): Gate<TArtifact, TScenario> {\n  const deltaThreshold = options.deltaThreshold ?? 0\n  const confidence = options.confidence ?? 0.95\n  const resamples = options.bootstrapResamples ?? 2000\n  const seed = options.bootstrapSeed ?? 1337\n  const minProductiveRuns = options.minProductiveRuns ?? 3\n  const heldoutStatistic = options.heldoutStatistic ?? 'mean'\n  const blockOnGaming = options.blockOnRewardHackingGaming ?? true\n\n  return {\n    name: 'defaultProductionGate',\n    async decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult> {\n      const reasons: string[] = []\n      const contributing: Array<{ name: string; passed: boolean; detail: unknown }> = []\n\n      // ── (1) heldout composite lift — paired-bootstrap CI, NOT a point estimate\n      // The shipped false positive: the baseline re-scored against itself read\n      // run-to-run model noise (91 vs 95) as a \"+4 lift\" and shipped, because a\n      // point estimate carries no confidence interval. Pair candidate vs\n      // baseline holdout cells by FULL cellId (never averaging reps away) and\n      // ship only when the bootstrap CI lower bound clears the threshold —\n      // i.e. the gain is real at the confidence level, not noise.\n      const scenarioIds = new Set(options.holdoutScenarios.map((s) => s.id))\n      const sig = heldoutSignificance(\n        pairHoldout(\n          ctx.judgeScores,\n          ctx.baselineJudgeScores ?? ctx.judgeScores,\n          scenarioIds,\n          (s) => s.composite,\n        ),\n        {\n          deltaThreshold,\n          minProductiveRuns,\n          confidence,\n          resamples,\n          seed,\n          statistic: heldoutStatistic,\n        },\n      )\n      // Point estimate of the CHOSEN ship statistic (mean by default); `.low`/\n      // `.high` are its CI. The median is kept as a diagnostic.\n      const delta = heldoutStatistic === 'median' ? sig.bootstrap.median : sig.bootstrap.mean\n      const heldoutPass = sig.significant\n      contributing.push({\n        name: 'heldout-significance',\n        passed: heldoutPass,\n        detail: {\n          n: sig.n,\n          delta,\n          deltaMean: sig.bootstrap.mean,\n          deltaMedianDiagnostic: sig.medianBootstrap.median,\n          // Back-compat: prior consumers read `deltaMedian`. It now always carries\n          // the median diagnostic (the ship decision keys on `delta`/mean).\n          deltaMedian: sig.medianBootstrap.median,\n          tieFraction: sig.tieFraction,\n          ciLow: sig.bootstrap.low,\n          ciHigh: sig.bootstrap.high,\n          confidence: sig.bootstrap.confidence,\n          deltaThreshold,\n          fewRuns: sig.fewRuns,\n        },\n      })\n      if (!heldoutPass) {\n        const tieNote =\n          sig.tieFraction >= TIE_WARN_FRACTION\n            ? `; ${(sig.tieFraction * 100).toFixed(0)}% tied scenarios`\n            : ''\n        reasons.push(\n          sig.fewRuns\n            ? `held-out: only ${sig.n} paired runs (< ${minProductiveRuns}) — too few to claim significance`\n            : `held-out CI.low ${sig.bootstrap.low.toFixed(3)} ≤ threshold ${deltaThreshold} (${heldoutStatistic} Δ ${delta.toFixed(3)}, ${(sig.bootstrap.confidence * 100).toFixed(0)}% CI [${sig.bootstrap.low.toFixed(3)}, ${sig.bootstrap.high.toFixed(3)}]${tieNote})`,\n        )\n      }\n\n      // ── (1b) per-dimension regression guard (anti-Goodhart) ──────────\n      // A net composite gain can hide a regression on a safety-critical\n      // dimension (e.g. hallucination_free for a legal agent — the verified run\n      // gained +25/+25 on deadline/fee while LOSING -30 on hallucination, and\n      // the composite-only gate never saw it). Block ship if any guarded\n      // dimension's paired-delta CI lower bound falls below −tolerance.\n      const dimRegs = options.criticalDimensions?.length\n        ? dimensionRegressions(\n            ctx.judgeScores,\n            ctx.baselineJudgeScores ?? ctx.judgeScores,\n            scenarioIds,\n            options.criticalDimensions,\n            { tolerance: options.regressionTolerance, confidence, resamples, seed },\n          )\n        : []\n      const regressed = dimRegs.filter((d) => d.regressed)\n      const dimPass = regressed.length === 0\n      contributing.push({\n        name: 'dimension-regression',\n        passed: dimPass,\n        detail: {\n          guarded: options.criticalDimensions ?? [],\n          regressions: dimRegs.map((d) => ({\n            dimension: d.dimension,\n            ciLow: d.bootstrap.low,\n            median: d.bootstrap.median,\n            tolerance: d.tolerance,\n            n: d.n,\n            regressed: d.regressed,\n          })),\n        },\n      })\n      if (!dimPass) {\n        reasons.push(\n          `critical dimension(s) regressed: ${regressed.map((d) => `${d.dimension} CI.low ${d.bootstrap.low.toFixed(3)} < -${d.tolerance}`).join('; ')}`,\n        )\n      }\n\n      // ── (2) budget gate ─────────────────────────────────────────────\n      const budgetPass =\n        options.budgetUsd === undefined ||\n        ctx.cost.candidate + ctx.cost.baseline <= options.budgetUsd\n      contributing.push({\n        name: 'budget',\n        passed: budgetPass,\n        detail: {\n          candidateUsd: ctx.cost.candidate,\n          baselineUsd: ctx.cost.baseline,\n          budgetUsd: options.budgetUsd,\n        },\n      })\n      if (!budgetPass) {\n        reasons.push(\n          `spend ${(ctx.cost.candidate + ctx.cost.baseline).toFixed(2)} > budget ${options.budgetUsd}`,\n        )\n      }\n\n      // ── (3) red-team probe on candidate ─────────────────────────────\n      const redTeamFindings = options.redTeamBattery\n        ? probeRedTeam(ctx.candidateArtifacts, options.redTeamBattery)\n        : { passed: true, findings: [] }\n      contributing.push({\n        name: 'red-team',\n        passed: redTeamFindings.passed,\n        detail: {\n          failures: redTeamFindings.findings.length,\n          sample: redTeamFindings.findings.slice(0, 3),\n        },\n      })\n      if (!redTeamFindings.passed) {\n        reasons.push(`red-team probe failed (${redTeamFindings.findings.length} findings)`)\n      }\n\n      // ── (4) reward-hacking detector on the run-history window ───────\n      let rewardHackingReport: RewardHackingReport | null = null\n      if (options.recentRuns && options.recentRuns.length >= 10) {\n        rewardHackingReport = detectRewardHacking({ runs: options.recentRuns })\n      }\n      // reward-hacking severity is numeric (0..1). \"gaming\" threshold per\n      // detectRewardHacking defaults = 0.6. Block when ANY finding is at\n      // gaming threshold OR the report verdict is 'gaming'.\n      const gamingThreshold = 0.6\n      const gamingFindings = (rewardHackingReport?.findings ?? []).filter(\n        (f) => f.severity >= gamingThreshold,\n      )\n      const rewardHackingPass =\n        !rewardHackingReport ||\n        !blockOnGaming ||\n        (gamingFindings.length === 0 && rewardHackingReport.verdict !== 'gaming')\n      contributing.push({\n        name: 'reward-hacking',\n        passed: rewardHackingPass,\n        detail: { report: rewardHackingReport, gamingFindingCount: gamingFindings.length },\n      })\n      if (!rewardHackingPass) {\n        reasons.push(\n          `reward-hacking detector flagged ${gamingFindings.length} gaming-severity findings (verdict=${rewardHackingReport!.verdict})`,\n        )\n      }\n\n      // ── (5) canary check on runs ────────────────────────────────────\n      let canaryReport: CanaryReport | null = null\n      if (options.recentRuns && options.recentRuns.length >= 10) {\n        canaryReport = runCanaries(options.recentRuns, {})\n      }\n      // CanarySeverity is 'info' | 'warn' | 'error' — block on 'error'.\n      const errorAlerts = (canaryReport?.alerts ?? []).filter((a) => a.severity === 'error')\n      const canaryPass = errorAlerts.length === 0\n      contributing.push({\n        name: 'canary',\n        passed: canaryPass,\n        detail: { totalAlerts: canaryReport?.alerts.length ?? 0, errorAlerts: errorAlerts.length },\n      })\n      if (!canaryPass) {\n        reasons.push(`canary error alerts: ${errorAlerts.length}`)\n      }\n\n      // ── Verdict ─────────────────────────────────────────────────────\n      const allPassed = contributing.every((c) => c.passed)\n      const decision = allPassed ? 'ship' : 'hold'\n\n      return {\n        decision,\n        reasons: reasons.length > 0 ? reasons : ['all gates passed'],\n        contributingGates: contributing,\n        delta,\n      }\n    },\n  }\n}\n\nfunction probeRedTeam<TArtifact>(\n  artifacts: Map<string, TArtifact>,\n  battery: RedTeamCase[],\n): { passed: boolean; findings: Array<{ scenarioId: string; reason: string }> } {\n  const findings: Array<{ scenarioId: string; reason: string }> = []\n  for (const [_cellId, artifact] of artifacts) {\n    const text = extractText(artifact)\n    if (text === undefined) continue\n    for (const rtCase of battery) {\n      const finding = scoreRedTeamOutput(text, [], rtCase)\n      if (!finding.passed) {\n        findings.push({ scenarioId: rtCase.id, reason: finding.reason ?? 'red-team probe failed' })\n      }\n    }\n  }\n  return { passed: findings.length === 0, findings }\n}\n\nfunction extractText(artifact: unknown): string | undefined {\n  if (typeof artifact === 'string') return artifact\n  if (artifact && typeof artifact === 'object') {\n    const rec = artifact as Record<string, unknown>\n    if (typeof rec.text === 'string') return rec.text\n    if (typeof rec.output === 'string') return rec.output\n    if (typeof rec.content === 'string') return rec.content\n  }\n  return undefined\n}\n","/**\n * Power preflight — \"can this budget detect the effect you are hunting?\"\n *\n * The failure it prevents (measured, twice): a live prompt-improvement campaign ran\n * 333 sandbox cells over 5.6 hours and produced a +0.08 holdout lift the ship gate\n * (paired bootstrap, CI.low > 0.05) could not distinguish from zero — because at\n * that holdout size and worker variance the MINIMUM DETECTABLE lift was larger than\n * any effect a prompt change plausibly produces. The budget was spent learning what\n * a 30-second calculation on the baseline cells already knew. No eval framework we\n * know of surfaces this; every underpowered improvement run everywhere ends in an\n * uninformative \"hold\".\n *\n * Model: the ship rule is `CI.low(paired Δ) > deltaThreshold`. Approximating the\n * bootstrap CI as normal, `CI.low ≈ effect − z·sd_Δ/√n`, so the smallest shippable\n * true effect is `MDE = deltaThreshold + z·sd_Δ/√n`. The paired-delta SD is unknown\n * before the candidate exists; we bound it by the zero-correlation case\n * `sd_Δ ≤ √2·sd_baseline` — a CONSERVATIVE (upper) MDE, which is the correct\n * direction for a warning. Pairing is per cell (`scenario:rep`), so reps multiply n.\n *\n * Standalone by design: feed it any baseline composites (a `gate:'none'` run, a\n * live-proof table) BEFORE budgeting the real search; `selfImprove` also attaches\n * it to every result and warns when the run was structurally unable to ship.\n */\n\nexport interface PowerPreflightOptions {\n  /** Per-cell baseline composites on the HOLDOUT scenarios (one per scenario:rep cell). */\n  baselineComposites: number[]\n  /** Paired observations the budgeted comparison will produce\n   *  (holdout scenarios × reps). Defaults to `baselineComposites.length`. */\n  pairedN?: number\n  /** The ship gate's effect-size threshold. Default 0.05 (defaultProductionGate). */\n  deltaThreshold?: number\n  /** CI confidence the gate uses. Default 0.95. */\n  confidence?: number\n  /** True when the holdout is scored by the SAME judge/scorer family as the gate\n   *  (selfImprove's default composition — one judge scores everything). Under a\n   *  shared channel, raising paired n reduces only the IDIOSYNCRATIC noise share;\n   *  systematic judge bias is untouched, so the MDE here is a lower bound and the\n   *  only full debiaser is an independent second scoring channel\n   *  (recursive-self-improvement S1c, closed form in EXP-023 P0). Default false. */\n  sharedScorerChannel?: boolean\n}\n\nexport interface PowerPreflight {\n  /** Paired observations the comparison will have. */\n  n: number\n  /** Baseline per-cell composite standard deviation (the variance the effect must beat). */\n  sd: number\n  /** Minimum detectable lift: the smallest TRUE effect the gate could ship at this budget. */\n  mde: number\n  /** Baseline holdout composite mean. */\n  baselineMean: number\n  /** Headroom to a perfect 1.0 composite (the largest achievable lift on a [0,1] judge). */\n  headroom: number\n  /** True when even the largest achievable effect (headroom) is below the MDE —\n   *  the run is structurally unable to ship regardless of proposal quality.\n   *  Only asserted for [0,1]-scaled judges (see `scaleAssumed`). */\n  underpowered: boolean\n  /** True when composites look [0,1]-scaled; headroom/underpowered are only\n   *  meaningful under that convention (0-100 judges get mde/sd/n but no verdict). */\n  scaleAssumed: boolean\n  deltaThreshold: number\n  confidence: number\n  /** Set when the holdout shares the gate's scoring channel: more cells cannot\n   *  buy back systematic judge bias — treat the MDE as a lower bound. */\n  sharedChannelCaveat?: string\n  /** One actionable sentence for humans and logs. */\n  recommendation: string\n}\n\n/** Two-sided z for the common confidence levels; interpolation is overkill here. */\nfunction zFor(confidence: number): number {\n  if (confidence >= 0.99) return 2.576\n  if (confidence >= 0.95) return 1.96\n  if (confidence >= 0.9) return 1.645\n  return 1.282\n}\n\nexport function powerPreflight(opts: PowerPreflightOptions): PowerPreflight {\n  const composites = opts.baselineComposites.filter((v) => Number.isFinite(v))\n  if (composites.length < 3) {\n    throw new Error(\n      `powerPreflight: need >= 3 finite baseline composites to estimate variance, got ${composites.length}`,\n    )\n  }\n  const deltaThreshold = opts.deltaThreshold ?? 0.05\n  const confidence = opts.confidence ?? 0.95\n  const n = opts.pairedN ?? composites.length\n  if (n < 2) throw new Error(`powerPreflight: pairedN must be >= 2, got ${n}`)\n\n  const mean = composites.reduce((a, b) => a + b, 0) / composites.length\n  const variance =\n    composites.reduce((a, b) => a + (b - mean) * (b - mean), 0) / (composites.length - 1)\n  const sd = Math.sqrt(variance)\n  const z = zFor(confidence)\n  const mde = deltaThreshold + (z * Math.SQRT2 * sd) / Math.sqrt(n)\n\n  const scaleAssumed = composites.every((v) => v >= -0.001 && v <= 1.5)\n  const headroom = Math.max(0, 1 - mean)\n  const underpowered = scaleAssumed && mde > headroom\n\n  const sharedChannelCaveat = opts.sharedScorerChannel\n    ? 'Holdout and gate share one scoring channel: raising n/reps reduces only idiosyncratic noise — systematic judge bias remains and this MDE is a lower bound. Full debiasing needs an independent second scoring channel (different judge/benchmark family).'\n    : undefined\n\n  const recommendation = underpowered\n    ? `UNDERPOWERED: minimum detectable lift ${mde.toFixed(3)} exceeds the ${headroom.toFixed(3)} headroom above the baseline (${mean.toFixed(3)}) — no achievable effect can ship at this budget. Raise paired n (scenarios x reps) to ~${Math.ceil(((z * Math.SQRT2 * sd) / Math.max(headroom - deltaThreshold, 0.01)) ** 2)} or reduce worker variance before searching.`\n    : `Minimum detectable lift at n=${n}: ${mde.toFixed(3)} (baseline sd ${sd.toFixed(3)}). Effects smaller than this cannot clear the gate; budget the search for effects you believe exceed it.`\n\n  return {\n    n,\n    sd,\n    mde,\n    baselineMean: mean,\n    headroom,\n    underpowered,\n    scaleAssumed,\n    deltaThreshold,\n    confidence,\n    ...(sharedChannelCaveat ? { sharedChannelCaveat } : {}),\n    recommendation: sharedChannelCaveat\n      ? `${recommendation} ${sharedChannelCaveat}`\n      : recommendation,\n  }\n}\n","/**\n * Promotion policy over the evidence VECTOR — the substrate's answer to \"never\n * collapse the multi-objective promotion decision into one scalar.\" A\n * `defaultProductionGate` is one opinionated composition; this module factors\n * the decision into two reusable pieces so MANY policies can compete over the\n * SAME evidence (the quant-desk pattern: one evidence bus, plural strategies):\n *\n *   buildEvidenceVector(ctx, objectives, opts) -> EvidenceVector   // the bus\n *   PromotionPolicy = (ev: EvidenceVector) => GateResult           // a strategy\n *   paretoPolicy(ev)                                               // the default strategy\n *   paretoSignificanceGate(options): Gate                          // bus + policy as a Gate\n *\n * The Pareto policy is SYMMETRIC multi-objective: every objective is BOTH a\n * potential gain source AND a safety floor (unlike `defaultProductionGate`,\n * where only `composite` can win and `criticalDimensions` are pure floors). A\n * candidate ships iff it weakly DOMINATES the baseline at the confidence level —\n * no objective credibly worse (CI floor breach) AND at least one objective\n * credibly better (CI gain). Insufficient evidence on ANY axis -> need_more_work\n * (NOT folded into hold: \"gather more reps\" and \"reject\" are different actions).\n *\n * Cost/latency are NOT CI axes here — `GateContext` carries only an aggregate\n * per-side cost, no per-cell observation vector to bootstrap. Treat them as hard\n * constraints (compose with a budget gate via `composeGate`), not faked CIs.\n */\n\nimport type { Direction } from '../../pareto'\nimport { type PairedBootstrapResult, pairedBootstrap } from '../../statistics'\nimport type { Gate, GateContext, GateDecision, GateResult, JudgeScore, Scenario } from '../types'\nimport { detectScale, pairHoldout } from './statistical-heldout'\n\n/** Where an objective's per-cell scalar comes from. `composite` reads the\n *  judge's composite; `dimension` reads a named per-dimension score. */\nexport type ObjectiveSource = { kind: 'composite' } | { kind: 'dimension'; dimension: string }\n\nexport interface PromotionObjective {\n  /** Stable label used in reports + `contributingGates`. */\n  name: string\n  source: ObjectiveSource\n  /** 'maximize' (quality dims) or 'minimize' (error/risk/length dims). Orients\n   *  the paired delta so a positive bootstrap always means \"candidate better\". */\n  direction: Direction\n  /** The good-direction paired-delta CI lower bound must EXCEED this to count\n   *  as a significant gain on this axis. Interpreted in the judge's native\n   *  scale. Default 0 (⇒ \"confidently better\"). */\n  gainThreshold?: number\n  /** A floor breach (regression) is declared when the good-direction CI lower\n   *  bound is below −floorTolerance. When omitted it auto-scales off observed\n   *  magnitudes (0.05 on [0,1], 5 on 0-100), matching `dimensionRegressions`. */\n  floorTolerance?: number\n}\n\n/** Per-axis verdict from the good-direction paired bootstrap. */\nexport type AxisVerdict = 'improved' | 'regressed' | 'flat' | 'few_runs'\n\nexport interface AxisEvidence {\n  name: string\n  source: ObjectiveSource\n  direction: Direction\n  /** Paired bootstrap on the GOOD-DIRECTION delta (oriented by `direction`):\n   *  a positive value means the candidate is better on this axis. */\n  bootstrap: PairedBootstrapResult\n  /** Paired observations contributing to this axis. */\n  n: number\n  gainThreshold: number\n  floorTolerance: number\n  verdict: AxisVerdict\n}\n\nexport interface EvidenceVector {\n  /** One entry per objective — NOTHING averaged across axes. */\n  axes: AxisEvidence[]\n  /** Smallest paired n across axes that produced observations — the binding\n   *  evidence-sufficiency constraint. 0 when no axis produced observations. */\n  minN: number\n  /** Aggregate per-side cost from the gate context (a constraint input, not a\n   *  CI axis — see the module header). */\n  cost: { candidate: number; baseline: number }\n}\n\n/** A promotion strategy: a pure function from the evidence vector to a verdict.\n *  Many policies can run over the same `EvidenceVector` and disagree — that's\n *  the point (competing strategies, shared evidence). */\nexport type PromotionPolicy = (ev: EvidenceVector) => GateResult\n\nexport interface BuildEvidenceVectorOptions {\n  /** Minimum paired observations before an axis can claim significance; below\n   *  it the axis is `few_runs`. Default 3. */\n  minProductiveRuns?: number\n  /** Confidence level for every axis bootstrap. Default 0.95. */\n  confidence?: number\n  /** Bootstrap resamples. Default 2000. */\n  resamples?: number\n  /** Fixed bootstrap seed for a deterministic, reproducible verdict. Default 1337. */\n  seed?: number\n}\n\n/**\n * The Evidence Bus. For each objective, pair candidate vs baseline by full\n * cellId and bootstrap a CI on the good-direction paired delta. Reuses the\n * exact `pairHoldout` + `pairedBootstrap` machinery the held-out gate uses, so\n * a single source of truth governs pairing granularity + scale handling.\n */\nexport function buildEvidenceVector<TArtifact, TScenario extends Scenario>(\n  ctx: GateContext<TArtifact, TScenario>,\n  objectives: PromotionObjective[],\n  opts: BuildEvidenceVectorOptions = {},\n): EvidenceVector {\n  if (objectives.length === 0) {\n    throw new Error('buildEvidenceVector: at least 1 objective required')\n  }\n  const minProductiveRuns = opts.minProductiveRuns ?? 3\n  const confidence = opts.confidence ?? 0.95\n  const resamples = opts.resamples ?? 2000\n  const seed = opts.seed ?? 1337\n  const baseline = ctx.baselineJudgeScores ?? ctx.judgeScores\n  const scenarioIds = new Set(ctx.scenarios.map((s) => s.id))\n\n  const axes: AxisEvidence[] = []\n  for (const obj of objectives) {\n    let select: (s: JudgeScore) => number | undefined\n    if (obj.source.kind === 'composite') {\n      select = (s) => s.composite\n    } else {\n      const dim = obj.source.dimension\n      select = (s) => s.dimensions[dim]\n    }\n    const paired = pairHoldout(ctx.judgeScores, baseline, scenarioIds, select)\n    // Orient to the good direction: maximize ⇒ bootstrap (candidate − baseline);\n    // minimize ⇒ bootstrap (baseline − candidate) by swapping args, so a\n    // positive bootstrap always reads as \"candidate better on this axis\".\n    const before = obj.direction === 'maximize' ? paired.before : paired.after\n    const after = obj.direction === 'maximize' ? paired.after : paired.before\n    const bootstrap = pairedBootstrap(before, after, {\n      confidence,\n      resamples,\n      statistic: 'median',\n      seed,\n    })\n    const n = paired.before.length\n    const floorTolerance =\n      obj.floorTolerance ?? 0.05 * detectScale([...paired.before, ...paired.after])\n    const gainThreshold = obj.gainThreshold ?? 0\n    // Floor check precedes the gain check: a credible regression must never be\n    // masked as \"improved\". With the defaults (gainThreshold 0, positive floor)\n    // the regions are disjoint and order is moot, but a consumer who sets a\n    // negative gainThreshold (\"accept small dips\") could otherwise have a real\n    // floor breach classified as a gain — anti-Goodhart wins the tie.\n    const verdict: AxisVerdict =\n      n < minProductiveRuns\n        ? 'few_runs'\n        : bootstrap.low < -floorTolerance\n          ? 'regressed'\n          : bootstrap.low > gainThreshold\n            ? 'improved'\n            : 'flat'\n    axes.push({\n      name: obj.name,\n      source: obj.source,\n      direction: obj.direction,\n      bootstrap,\n      n,\n      gainThreshold,\n      floorTolerance,\n      verdict,\n    })\n  }\n  const ns = axes.map((a) => a.n).filter((n) => n > 0)\n  const minN = ns.length > 0 ? Math.min(...ns) : 0\n  return { axes, minN, cost: { candidate: ctx.cost.candidate, baseline: ctx.cost.baseline } }\n}\n\n/**\n * The default strategy: symmetric multi-objective Pareto significance. Ship iff\n * the candidate weakly dominates the baseline at the confidence level — no axis\n * credibly worse AND ≥1 axis credibly better. Floor breach on any axis → hold\n * (anti-Goodhart, dominates everything). Insufficient evidence on any axis →\n * need_more_work. Statistically equivalent → hold (never ship noise).\n */\nexport const paretoPolicy: PromotionPolicy = (ev) => {\n  const contributingGates = ev.axes.map((ax) => ({\n    name: `objective:${ax.name}`,\n    passed: ax.verdict === 'improved',\n    detail: {\n      direction: ax.direction,\n      source: ax.source,\n      verdict: ax.verdict,\n      n: ax.n,\n      deltaMedian: ax.bootstrap.median,\n      ciLow: ax.bootstrap.low,\n      ciHigh: ax.bootstrap.high,\n      confidence: ax.bootstrap.confidence,\n      gainThreshold: ax.gainThreshold,\n      floorTolerance: ax.floorTolerance,\n    },\n  }))\n\n  const regressed = ev.axes.filter((a) => a.verdict === 'regressed')\n  const fewRuns = ev.axes.filter((a) => a.verdict === 'few_runs')\n  const improved = ev.axes.filter((a) => a.verdict === 'improved')\n\n  let decision: GateDecision\n  const reasons: string[] = []\n  if (regressed.length > 0) {\n    // Floor breach dominates: a credible regression on ANY axis blocks ship even\n    // if another axis improved. This makes the +gain/−safety false positive\n    // structurally impossible whenever the safety dim is an objective.\n    decision = 'hold'\n    for (const a of regressed) {\n      reasons.push(\n        `objective '${a.name}' regressed: good-direction CI.low ${a.bootstrap.low.toFixed(3)} < -${a.floorTolerance} (n=${a.n})`,\n      )\n    }\n  } else if (fewRuns.length > 0) {\n    // No credible regression on the scored axes, but ≥1 axis lacks the evidence\n    // to claim a gain ⇒ gather more reps, do NOT reject.\n    decision = 'need_more_work'\n    for (const a of fewRuns) {\n      reasons.push(\n        `objective '${a.name}' has only n=${a.n} paired runs — insufficient evidence to claim significance`,\n      )\n    }\n  } else if (improved.length > 0) {\n    // Weakly dominates (no axis worse) AND strictly better on ≥1 axis ⇒ a Pareto\n    // improvement at the confidence level.\n    decision = 'ship'\n    reasons.push(\n      `Pareto improvement at the confidence level: ${improved\n        .map(\n          (a) =>\n            `'${a.name}' +${a.bootstrap.median.toFixed(3)} (CI.low ${a.bootstrap.low.toFixed(3)})`,\n        )\n        .join(', ')}; no objective regressed`,\n    )\n  } else {\n    // Enough evidence, nothing credibly better or worse ⇒ statistically\n    // equivalent. Do NOT ship a no-op.\n    decision = 'hold'\n    reasons.push(\n      'no Pareto improvement: candidate statistically equivalent to baseline on every objective',\n    )\n  }\n\n  // `delta` surfaces the composite axis if present, else the first axis — a\n  // single convenience scalar; the vector lives in `contributingGates`.\n  const composite = ev.axes.find((a) => a.source.kind === 'composite') ?? ev.axes[0]\n  return { decision, reasons, contributingGates, delta: composite?.bootstrap.median }\n}\n\nexport interface ParetoSignificanceGateOptions extends BuildEvidenceVectorOptions {\n  /** The objective vector. Every axis is both a gain source and a safety floor. */\n  objectives: PromotionObjective[]\n  /** Strategy applied to the evidence vector. Default `paretoPolicy`. Override\n   *  to run a stricter/looser strategy over the SAME bus (competing policies). */\n  policy?: PromotionPolicy\n  /** Override the gate name in reports. */\n  name?: string\n}\n\n/**\n * Wrap the bus + a policy as a `Gate`. Plugs into the existing\n * `runImprovementLoop({ gate })` slot and composes via `composeGate`; default\n * loop behavior is unchanged because consumers opt in by passing this gate.\n */\nexport function paretoSignificanceGate<TArtifact = unknown, TScenario extends Scenario = Scenario>(\n  options: ParetoSignificanceGateOptions,\n): Gate<TArtifact, TScenario> {\n  if (options.objectives.length === 0) {\n    throw new Error('paretoSignificanceGate: at least 1 objective required')\n  }\n  const policy = options.policy ?? paretoPolicy\n  return {\n    name: options.name ?? 'paretoSignificanceGate',\n    async decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult> {\n      const ev = buildEvidenceVector(ctx, options.objectives, options)\n      return policy(ev)\n    },\n  }\n}\n","/**\n * `runEval` — the simplest preset over `runCampaign`. No optimizer, no\n * gate, no auto-PR. Just: run scenarios through dispatch, score with\n * judges, return CampaignResult.\n *\n * The 80% case for consumers who want a scorecard, not an improvement loop.\n */\n\nimport { type RunCampaignOptions, runCampaign } from '../run-campaign'\nimport type { CampaignResult, Scenario } from '../types'\n\nexport interface RunEvalOptions<TScenario extends Scenario, TArtifact>\n  extends Omit<RunCampaignOptions<TScenario, TArtifact>, 'runDir'> {\n  runDir: string\n}\n\n/**\n * Simplest evaluation preset: run scenarios through dispatch, score with judges, and return a `CampaignResult` — no optimizer, no gate, no PR.\n */\nexport async function runEval<TScenario extends Scenario, TArtifact>(\n  opts: RunEvalOptions<TScenario, TArtifact>,\n): Promise<CampaignResult<TArtifact, TScenario>> {\n  return runCampaign(opts)\n}\n","/**\n * `evolutionaryProposer` — adapts a stateless `Mutator` (population mutation:\n * GEPA / AxGEPA / reflective-mutation) into a `SurfaceProposer`. This is\n * the evolutionary strategy: each generation, mutate the current best surface\n * into N candidates, measure, select. No generation memory beyond the current\n * surface; the loop body handles ranking + promotion.\n *\n * The reflective alternative is agent-runtime's runtime proposer with a\n * `reflectiveGenerator` / `agenticGenerator`: it reasons over the report +\n * trace findings to propose targeted edits rather than blind mutations. Both\n * conform to `SurfaceProposer`; the improvement loop is identical either way.\n */\n\nimport type { Mutator, SurfaceProposer } from '../types'\n\nexport interface EvolutionaryProposerOptions<TFindings = unknown> {\n  mutator: Mutator<TFindings>\n  /** External findings fed to the mutator each generation. Default: []. */\n  findings?: TFindings[]\n}\n\n/**\n * Wrap a stateless `Mutator` (GEPA, AxGEPA, reflective-mutation) as a `SurfaceProposer` that mutates the current best surface into N candidates each generation.\n */\nexport function evolutionaryProposer<TFindings = unknown>(\n  opts: EvolutionaryProposerOptions<TFindings>,\n): SurfaceProposer<TFindings> {\n  return {\n    kind: `evolutionary:${opts.mutator.kind}`,\n    async propose({ currentSurface, findings, populationSize, signal }) {\n      return opts.mutator.mutate({\n        findings: findings.length > 0 ? findings : (opts.findings ?? []),\n        currentSurface,\n        populationSize,\n        signal,\n      })\n    },\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAYO,SAAS,eACX,OACyB;AAC5B,MAAI,MAAM,WAAW,GAAG;AACtB,UAAM,IAAI,MAAM,wCAAwC;AAAA,EAC1D;AACA,SAAO;AAAA,IACL,MAAM,YAAY,MAAM,IAAI,CAAC,MAAM,EAAE,IAAI,EAAE,KAAK,GAAG,CAAC;AAAA,IACpD,MAAM,OAAO,KAA6D;AACxE,YAAM,UAAwE,CAAC;AAC/E,iBAAW,QAAQ,OAAO;AACxB,cAAM,MAAM,MAAM,KAAK,OAAO,GAAG;AACjC,gBAAQ,KAAK,EAAE,MAAM,IAAI,CAAC;AAAA,MAC5B;AAQA,YAAM,YAAY,QAAQ,IAAI,CAAC,MAAM,EAAE,IAAI,QAAQ;AACnD,YAAM,UAAwB,UAAU,MAAM,CAAC,MAAM,MAAM,MAAM,IAC7D,SACA,UAAU,SAAS,cAAc,IAC/B,iBACA,UAAU,SAAS,eAAe,IAChC,kBACA,UAAU,SAAS,MAAM,IACvB,SACA;AAEV,YAAM,eAAe,QAAQ;AAAA,QAAQ,CAAC,MACpC,EAAE,IAAI,kBAAkB,SAAS,IAC7B,EAAE,IAAI,oBACN,CAAC,EAAE,MAAM,EAAE,KAAK,MAAM,QAAQ,EAAE,IAAI,aAAa,QAAQ,QAAQ,EAAE,IAAI,CAAC;AAAA,MAC9E;AAEA,YAAM,UAAU,QAAQ;AAAA,QAAQ,CAAC,MAC/B,EAAE,IAAI,QAAQ,IAAI,CAAC,WAAW,IAAI,EAAE,KAAK,IAAI,KAAK,MAAM,EAAE;AAAA,MAC5D;AAEA,aAAO;AAAA,QACL,UAAU;AAAA,QACV;AAAA,QACA,mBAAmB;AAAA,QACnB,OAAO,QAAQ,CAAC,GAAG,IAAI;AAAA,MACzB;AAAA,IACF;AAAA,EACF;AACF;;;ACeO,SAAS,sBACd,SAC4B;AAC5B,QAAM,iBAAiB,QAAQ,kBAAkB;AACjD,QAAM,aAAa,QAAQ,cAAc;AACzC,QAAM,YAAY,QAAQ,sBAAsB;AAChD,QAAM,OAAO,QAAQ,iBAAiB;AACtC,QAAM,oBAAoB,QAAQ,qBAAqB;AACvD,QAAM,mBAAmB,QAAQ,oBAAoB;AACrD,QAAM,gBAAgB,QAAQ,8BAA8B;AAE5D,SAAO;AAAA,IACL,MAAM;AAAA,IACN,MAAM,OAAO,KAA6D;AACxE,YAAM,UAAoB,CAAC;AAC3B,YAAM,eAA0E,CAAC;AASjF,YAAM,cAAc,IAAI,IAAI,QAAQ,iBAAiB,IAAI,CAAC,MAAM,EAAE,EAAE,CAAC;AACrE,YAAM,MAAM;AAAA,QACV;AAAA,UACE,IAAI;AAAA,UACJ,IAAI,uBAAuB,IAAI;AAAA,UAC/B;AAAA,UACA,CAAC,MAAM,EAAE;AAAA,QACX;AAAA,QACA;AAAA,UACE;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,WAAW;AAAA,QACb;AAAA,MACF;AAGA,YAAM,QAAQ,qBAAqB,WAAW,IAAI,UAAU,SAAS,IAAI,UAAU;AACnF,YAAM,cAAc,IAAI;AACxB,mBAAa,KAAK;AAAA,QAChB,MAAM;AAAA,QACN,QAAQ;AAAA,QACR,QAAQ;AAAA,UACN,GAAG,IAAI;AAAA,UACP;AAAA,UACA,WAAW,IAAI,UAAU;AAAA,UACzB,uBAAuB,IAAI,gBAAgB;AAAA;AAAA;AAAA,UAG3C,aAAa,IAAI,gBAAgB;AAAA,UACjC,aAAa,IAAI;AAAA,UACjB,OAAO,IAAI,UAAU;AAAA,UACrB,QAAQ,IAAI,UAAU;AAAA,UACtB,YAAY,IAAI,UAAU;AAAA,UAC1B;AAAA,UACA,SAAS,IAAI;AAAA,QACf;AAAA,MACF,CAAC;AACD,UAAI,CAAC,aAAa;AAChB,cAAM,UACJ,IAAI,eAAe,oBACf,MAAM,IAAI,cAAc,KAAK,QAAQ,CAAC,CAAC,qBACvC;AACN,gBAAQ;AAAA,UACN,IAAI,UACA,kBAAkB,IAAI,CAAC,mBAAmB,iBAAiB,2CAC3D,mBAAmB,IAAI,UAAU,IAAI,QAAQ,CAAC,CAAC,qBAAgB,cAAc,KAAK,gBAAgB,WAAM,MAAM,QAAQ,CAAC,CAAC,MAAM,IAAI,UAAU,aAAa,KAAK,QAAQ,CAAC,CAAC,SAAS,IAAI,UAAU,IAAI,QAAQ,CAAC,CAAC,KAAK,IAAI,UAAU,KAAK,QAAQ,CAAC,CAAC,IAAI,OAAO;AAAA,QAChQ;AAAA,MACF;AAQA,YAAM,UAAU,QAAQ,oBAAoB,SACxC;AAAA,QACE,IAAI;AAAA,QACJ,IAAI,uBAAuB,IAAI;AAAA,QAC/B;AAAA,QACA,QAAQ;AAAA,QACR,EAAE,WAAW,QAAQ,qBAAqB,YAAY,WAAW,KAAK;AAAA,MACxE,IACA,CAAC;AACL,YAAM,YAAY,QAAQ,OAAO,CAAC,MAAM,EAAE,SAAS;AACnD,YAAM,UAAU,UAAU,WAAW;AACrC,mBAAa,KAAK;AAAA,QAChB,MAAM;AAAA,QACN,QAAQ;AAAA,QACR,QAAQ;AAAA,UACN,SAAS,QAAQ,sBAAsB,CAAC;AAAA,UACxC,aAAa,QAAQ,IAAI,CAAC,OAAO;AAAA,YAC/B,WAAW,EAAE;AAAA,YACb,OAAO,EAAE,UAAU;AAAA,YACnB,QAAQ,EAAE,UAAU;AAAA,YACpB,WAAW,EAAE;AAAA,YACb,GAAG,EAAE;AAAA,YACL,WAAW,EAAE;AAAA,UACf,EAAE;AAAA,QACJ;AAAA,MACF,CAAC;AACD,UAAI,CAAC,SAAS;AACZ,gBAAQ;AAAA,UACN,oCAAoC,UAAU,IAAI,CAAC,MAAM,GAAG,EAAE,SAAS,WAAW,EAAE,UAAU,IAAI,QAAQ,CAAC,CAAC,OAAO,EAAE,SAAS,EAAE,EAAE,KAAK,IAAI,CAAC;AAAA,QAC9I;AAAA,MACF;AAGA,YAAM,aACJ,QAAQ,cAAc,UACtB,IAAI,KAAK,YAAY,IAAI,KAAK,YAAY,QAAQ;AACpD,mBAAa,KAAK;AAAA,QAChB,MAAM;AAAA,QACN,QAAQ;AAAA,QACR,QAAQ;AAAA,UACN,cAAc,IAAI,KAAK;AAAA,UACvB,aAAa,IAAI,KAAK;AAAA,UACtB,WAAW,QAAQ;AAAA,QACrB;AAAA,MACF,CAAC;AACD,UAAI,CAAC,YAAY;AACf,gBAAQ;AAAA,UACN,UAAU,IAAI,KAAK,YAAY,IAAI,KAAK,UAAU,QAAQ,CAAC,CAAC,aAAa,QAAQ,SAAS;AAAA,QAC5F;AAAA,MACF;AAGA,YAAM,kBAAkB,QAAQ,iBAC5B,aAAa,IAAI,oBAAoB,QAAQ,cAAc,IAC3D,EAAE,QAAQ,MAAM,UAAU,CAAC,EAAE;AACjC,mBAAa,KAAK;AAAA,QAChB,MAAM;AAAA,QACN,QAAQ,gBAAgB;AAAA,QACxB,QAAQ;AAAA,UACN,UAAU,gBAAgB,SAAS;AAAA,UACnC,QAAQ,gBAAgB,SAAS,MAAM,GAAG,CAAC;AAAA,QAC7C;AAAA,MACF,CAAC;AACD,UAAI,CAAC,gBAAgB,QAAQ;AAC3B,gBAAQ,KAAK,0BAA0B,gBAAgB,SAAS,MAAM,YAAY;AAAA,MACpF;AAGA,UAAI,sBAAkD;AACtD,UAAI,QAAQ,cAAc,QAAQ,WAAW,UAAU,IAAI;AACzD,8BAAsB,oBAAoB,EAAE,MAAM,QAAQ,WAAW,CAAC;AAAA,MACxE;AAIA,YAAM,kBAAkB;AACxB,YAAM,kBAAkB,qBAAqB,YAAY,CAAC,GAAG;AAAA,QAC3D,CAAC,MAAM,EAAE,YAAY;AAAA,MACvB;AACA,YAAM,oBACJ,CAAC,uBACD,CAAC,iBACA,eAAe,WAAW,KAAK,oBAAoB,YAAY;AAClE,mBAAa,KAAK;AAAA,QAChB,MAAM;AAAA,QACN,QAAQ;AAAA,QACR,QAAQ,EAAE,QAAQ,qBAAqB,oBAAoB,eAAe,OAAO;AAAA,MACnF,CAAC;AACD,UAAI,CAAC,mBAAmB;AACtB,gBAAQ;AAAA,UACN,mCAAmC,eAAe,MAAM,sCAAsC,oBAAqB,OAAO;AAAA,QAC5H;AAAA,MACF;AAGA,UAAI,eAAoC;AACxC,UAAI,QAAQ,cAAc,QAAQ,WAAW,UAAU,IAAI;AACzD,uBAAe,YAAY,QAAQ,YAAY,CAAC,CAAC;AAAA,MACnD;AAEA,YAAM,eAAe,cAAc,UAAU,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,aAAa,OAAO;AACrF,YAAM,aAAa,YAAY,WAAW;AAC1C,mBAAa,KAAK;AAAA,QAChB,MAAM;AAAA,QACN,QAAQ;AAAA,QACR,QAAQ,EAAE,aAAa,cAAc,OAAO,UAAU,GAAG,aAAa,YAAY,OAAO;AAAA,MAC3F,CAAC;AACD,UAAI,CAAC,YAAY;AACf,gBAAQ,KAAK,wBAAwB,YAAY,MAAM,EAAE;AAAA,MAC3D;AAGA,YAAM,YAAY,aAAa,MAAM,CAAC,MAAM,EAAE,MAAM;AACpD,YAAM,WAAW,YAAY,SAAS;AAEtC,aAAO;AAAA,QACL;AAAA,QACA,SAAS,QAAQ,SAAS,IAAI,UAAU,CAAC,kBAAkB;AAAA,QAC3D,mBAAmB;AAAA,QACnB;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACF;AAEA,SAAS,aACP,WACA,SAC8E;AAC9E,QAAM,WAA0D,CAAC;AACjE,aAAW,CAAC,SAAS,QAAQ,KAAK,WAAW;AAC3C,UAAM,OAAO,YAAY,QAAQ;AACjC,QAAI,SAAS,OAAW;AACxB,eAAW,UAAU,SAAS;AAC5B,YAAM,UAAU,mBAAmB,MAAM,CAAC,GAAG,MAAM;AACnD,UAAI,CAAC,QAAQ,QAAQ;AACnB,iBAAS,KAAK,EAAE,YAAY,OAAO,IAAI,QAAQ,QAAQ,UAAU,wBAAwB,CAAC;AAAA,MAC5F;AAAA,IACF;AAAA,EACF;AACA,SAAO,EAAE,QAAQ,SAAS,WAAW,GAAG,SAAS;AACnD;AAEA,SAAS,YAAY,UAAuC;AAC1D,MAAI,OAAO,aAAa,SAAU,QAAO;AACzC,MAAI,YAAY,OAAO,aAAa,UAAU;AAC5C,UAAM,MAAM;AACZ,QAAI,OAAO,IAAI,SAAS,SAAU,QAAO,IAAI;AAC7C,QAAI,OAAO,IAAI,WAAW,SAAU,QAAO,IAAI;AAC/C,QAAI,OAAO,IAAI,YAAY,SAAU,QAAO,IAAI;AAAA,EAClD;AACA,SAAO;AACT;;;AChPA,SAAS,KAAK,YAA4B;AACxC,MAAI,cAAc,KAAM,QAAO;AAC/B,MAAI,cAAc,KAAM,QAAO;AAC/B,MAAI,cAAc,IAAK,QAAO;AAC9B,SAAO;AACT;AAEO,SAAS,eAAe,MAA6C;AAC1E,QAAM,aAAa,KAAK,mBAAmB,OAAO,CAAC,MAAM,OAAO,SAAS,CAAC,CAAC;AAC3E,MAAI,WAAW,SAAS,GAAG;AACzB,UAAM,IAAI;AAAA,MACR,kFAAkF,WAAW,MAAM;AAAA,IACrG;AAAA,EACF;AACA,QAAM,iBAAiB,KAAK,kBAAkB;AAC9C,QAAM,aAAa,KAAK,cAAc;AACtC,QAAM,IAAI,KAAK,WAAW,WAAW;AACrC,MAAI,IAAI,EAAG,OAAM,IAAI,MAAM,6CAA6C,CAAC,EAAE;AAE3E,QAAM,OAAO,WAAW,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,WAAW;AAChE,QAAM,WACJ,WAAW,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,SAAS,IAAI,OAAO,CAAC,KAAK,WAAW,SAAS;AACrF,QAAM,KAAK,KAAK,KAAK,QAAQ;AAC7B,QAAM,IAAI,KAAK,UAAU;AACzB,QAAM,MAAM,iBAAkB,IAAI,KAAK,QAAQ,KAAM,KAAK,KAAK,CAAC;AAEhE,QAAM,eAAe,WAAW,MAAM,CAAC,MAAM,KAAK,SAAU,KAAK,GAAG;AACpE,QAAM,WAAW,KAAK,IAAI,GAAG,IAAI,IAAI;AACrC,QAAM,eAAe,gBAAgB,MAAM;AAE3C,QAAM,sBAAsB,KAAK,sBAC7B,mQACA;AAEJ,QAAM,iBAAiB,eACnB,yCAAyC,IAAI,QAAQ,CAAC,CAAC,gBAAgB,SAAS,QAAQ,CAAC,CAAC,iCAAiC,KAAK,QAAQ,CAAC,CAAC,gGAA2F,KAAK,MAAO,IAAI,KAAK,QAAQ,KAAM,KAAK,IAAI,WAAW,gBAAgB,IAAI,MAAM,CAAC,CAAC,iDACxT,gCAAgC,CAAC,KAAK,IAAI,QAAQ,CAAC,CAAC,iBAAiB,GAAG,QAAQ,CAAC,CAAC;AAEtF,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,GAAI,sBAAsB,EAAE,oBAAoB,IAAI,CAAC;AAAA,IACrD,gBAAgB,sBACZ,GAAG,cAAc,IAAI,mBAAmB,KACxC;AAAA,EACN;AACF;;;ACtBO,SAAS,oBACd,KACA,YACA,OAAmC,CAAC,GACpB;AAChB,MAAI,WAAW,WAAW,GAAG;AAC3B,UAAM,IAAI,MAAM,oDAAoD;AAAA,EACtE;AACA,QAAM,oBAAoB,KAAK,qBAAqB;AACpD,QAAM,aAAa,KAAK,cAAc;AACtC,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,WAAW,IAAI,uBAAuB,IAAI;AAChD,QAAM,cAAc,IAAI,IAAI,IAAI,UAAU,IAAI,CAAC,MAAM,EAAE,EAAE,CAAC;AAE1D,QAAM,OAAuB,CAAC;AAC9B,aAAW,OAAO,YAAY;AAC5B,QAAI;AACJ,QAAI,IAAI,OAAO,SAAS,aAAa;AACnC,eAAS,CAAC,MAAM,EAAE;AAAA,IACpB,OAAO;AACL,YAAM,MAAM,IAAI,OAAO;AACvB,eAAS,CAAC,MAAM,EAAE,WAAW,GAAG;AAAA,IAClC;AACA,UAAM,SAAS,YAAY,IAAI,aAAa,UAAU,aAAa,MAAM;AAIzE,UAAM,SAAS,IAAI,cAAc,aAAa,OAAO,SAAS,OAAO;AACrE,UAAM,QAAQ,IAAI,cAAc,aAAa,OAAO,QAAQ,OAAO;AACnE,UAAM,YAAY,gBAAgB,QAAQ,OAAO;AAAA,MAC/C;AAAA,MACA;AAAA,MACA,WAAW;AAAA,MACX;AAAA,IACF,CAAC;AACD,UAAM,IAAI,OAAO,OAAO;AACxB,UAAM,iBACJ,IAAI,kBAAkB,OAAO,YAAY,CAAC,GAAG,OAAO,QAAQ,GAAG,OAAO,KAAK,CAAC;AAC9E,UAAM,gBAAgB,IAAI,iBAAiB;AAM3C,UAAM,UACJ,IAAI,oBACA,aACA,UAAU,MAAM,CAAC,iBACf,cACA,UAAU,MAAM,gBACd,aACA;AACV,SAAK,KAAK;AAAA,MACR,MAAM,IAAI;AAAA,MACV,QAAQ,IAAI;AAAA,MACZ,WAAW,IAAI;AAAA,MACf;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAAA,EACH;AACA,QAAM,KAAK,KAAK,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,OAAO,CAAC,MAAM,IAAI,CAAC;AACnD,QAAM,OAAO,GAAG,SAAS,IAAI,KAAK,IAAI,GAAG,EAAE,IAAI;AAC/C,SAAO,EAAE,MAAM,MAAM,MAAM,EAAE,WAAW,IAAI,KAAK,WAAW,UAAU,IAAI,KAAK,SAAS,EAAE;AAC5F;AASO,IAAM,eAAgC,CAAC,OAAO;AACnD,QAAM,oBAAoB,GAAG,KAAK,IAAI,CAAC,QAAQ;AAAA,IAC7C,MAAM,aAAa,GAAG,IAAI;AAAA,IAC1B,QAAQ,GAAG,YAAY;AAAA,IACvB,QAAQ;AAAA,MACN,WAAW,GAAG;AAAA,MACd,QAAQ,GAAG;AAAA,MACX,SAAS,GAAG;AAAA,MACZ,GAAG,GAAG;AAAA,MACN,aAAa,GAAG,UAAU;AAAA,MAC1B,OAAO,GAAG,UAAU;AAAA,MACpB,QAAQ,GAAG,UAAU;AAAA,MACrB,YAAY,GAAG,UAAU;AAAA,MACzB,eAAe,GAAG;AAAA,MAClB,gBAAgB,GAAG;AAAA,IACrB;AAAA,EACF,EAAE;AAEF,QAAM,YAAY,GAAG,KAAK,OAAO,CAAC,MAAM,EAAE,YAAY,WAAW;AACjE,QAAM,UAAU,GAAG,KAAK,OAAO,CAAC,MAAM,EAAE,YAAY,UAAU;AAC9D,QAAM,WAAW,GAAG,KAAK,OAAO,CAAC,MAAM,EAAE,YAAY,UAAU;AAE/D,MAAI;AACJ,QAAM,UAAoB,CAAC;AAC3B,MAAI,UAAU,SAAS,GAAG;AAIxB,eAAW;AACX,eAAW,KAAK,WAAW;AACzB,cAAQ;AAAA,QACN,cAAc,EAAE,IAAI,sCAAsC,EAAE,UAAU,IAAI,QAAQ,CAAC,CAAC,OAAO,EAAE,cAAc,OAAO,EAAE,CAAC;AAAA,MACvH;AAAA,IACF;AAAA,EACF,WAAW,QAAQ,SAAS,GAAG;AAG7B,eAAW;AACX,eAAW,KAAK,SAAS;AACvB,cAAQ;AAAA,QACN,cAAc,EAAE,IAAI,gBAAgB,EAAE,CAAC;AAAA,MACzC;AAAA,IACF;AAAA,EACF,WAAW,SAAS,SAAS,GAAG;AAG9B,eAAW;AACX,YAAQ;AAAA,MACN,+CAA+C,SAC5C;AAAA,QACC,CAAC,MACC,IAAI,EAAE,IAAI,MAAM,EAAE,UAAU,OAAO,QAAQ,CAAC,CAAC,YAAY,EAAE,UAAU,IAAI,QAAQ,CAAC,CAAC;AAAA,MACvF,EACC,KAAK,IAAI,CAAC;AAAA,IACf;AAAA,EACF,OAAO;AAGL,eAAW;AACX,YAAQ;AAAA,MACN;AAAA,IACF;AAAA,EACF;AAIA,QAAM,YAAY,GAAG,KAAK,KAAK,CAAC,MAAM,EAAE,OAAO,SAAS,WAAW,KAAK,GAAG,KAAK,CAAC;AACjF,SAAO,EAAE,UAAU,SAAS,mBAAmB,OAAO,WAAW,UAAU,OAAO;AACpF;AAiBO,SAAS,uBACd,SAC4B;AAC5B,MAAI,QAAQ,WAAW,WAAW,GAAG;AACnC,UAAM,IAAI,MAAM,uDAAuD;AAAA,EACzE;AACA,QAAM,SAAS,QAAQ,UAAU;AACjC,SAAO;AAAA,IACL,MAAM,QAAQ,QAAQ;AAAA,IACtB,MAAM,OAAO,KAA6D;AACxE,YAAM,KAAK,oBAAoB,KAAK,QAAQ,YAAY,OAAO;AAC/D,aAAO,OAAO,EAAE;AAAA,IAClB;AAAA,EACF;AACF;;;AClQA,eAAsB,QACpB,MAC+C;AAC/C,SAAO,YAAY,IAAI;AACzB;;;ACCO,SAAS,qBACd,MAC4B;AAC5B,SAAO;AAAA,IACL,MAAM,gBAAgB,KAAK,QAAQ,IAAI;AAAA,IACvC,MAAM,QAAQ,EAAE,gBAAgB,UAAU,gBAAgB,OAAO,GAAG;AAClE,aAAO,KAAK,QAAQ,OAAO;AAAA,QACzB,UAAU,SAAS,SAAS,IAAI,WAAY,KAAK,YAAY,CAAC;AAAA,QAC9D;AAAA,QACA;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AACF;","names":[]}