@tangle-network/agent-eval 0.93.0 → 0.95.0

package/dist/rl.js

@@ -16,19 +16,19 @@ import {
 } from "./chunk-3RF76KTD.js";
 import {
   runEvalCampaign
-} from "./chunk-GSH6QNNS.js";
-import "./chunk-CVVHBFGN.js";
+} from "./chunk-X74V6ESX.js";
+import "./chunk-CWNP4DV4.js";
 import {
   rubricPredictiveValidity
 } from "./chunk-YRZ4M5GS.js";
 import {
   evaluateInterimReleaseConfidence
 } from "./chunk-MAZ26DC7.js";
-import "./chunk-CY6U5S3X.js";
+import "./chunk-URWVKRCS.js";
 import {
   benjaminiHochberg,
   wilcoxonSignedRank
-} from "./chunk-LMZQ2Z4U.js";
+} from "./chunk-56GC6VYO.js";
 import {
   observationsFromRunRecords,
   thompsonCurriculum,
@@ -36,7 +36,7 @@ import {
 } from "./chunk-VZSRQ272.js";
 import "./chunk-SBCB6VZY.js";
 import "./chunk-PC4UYEBM.js";
-import "./chunk-KWRRMR3J.js";
+import "./chunk-LO6IOIJ2.js";
 import "./chunk-TVVP3ZZQ.js";
 import "./chunk-VSMTAMNK.js";
 import {
@@ -44,6 +44,197 @@ import {
 } from "./chunk-3BFEG2F6.js";
 import "./chunk-PZ5AY32C.js";
 
+// src/rl/adaptation-eval.ts
+async function runAdaptationCurve(opts) {
+  const ks = opts.ks ?? [0, 1, 2, 4, 8, 16];
+  const reps = opts.reps ?? 3;
+  const passThreshold = opts.passThreshold ?? 0.5;
+  const sortedKs = [...ks].sort((a, b) => a - b);
+  const points = [];
+  for (const k of sortedKs) {
+    const perScenario = [];
+    const allScores = [];
+    let totalPasses = 0;
+    let totalAttempts = 0;
+    for (const scenario of opts.scenarios) {
+      const sid = scenario.scenarioId ?? `scenario-${opts.scenarios.indexOf(scenario)}`;
+      const scores = [];
+      let passes = 0;
+      for (let r = 0; r < reps; r++) {
+        const score = await opts.runner.run({ scenario, k, rep: r });
+        scores.push(score);
+        if (score >= passThreshold) passes++;
+        allScores.push(score);
+        if (score >= passThreshold) totalPasses++;
+        totalAttempts++;
+      }
+      const meanS = scores.reduce((s, v) => s + v, 0) / scores.length;
+      perScenario.push({ scenarioId: sid, meanScore: meanS, passes, total: scores.length });
+    }
+    const meanScore = allScores.reduce((s, v) => s + v, 0) / Math.max(1, allScores.length);
+    const variance = allScores.length < 2 ? 0 : allScores.reduce((s, v) => s + (v - meanScore) ** 2, 0) / (allScores.length - 1);
+    points.push({
+      k,
+      meanScore,
+      passRate: totalPasses / Math.max(1, totalAttempts),
+      std: Math.sqrt(variance),
+      n: allScores.length,
+      perScenario
+    });
+  }
+  const firstPassK2 = points.find((p) => p.passRate >= passThreshold)?.k ?? null;
+  const maxK = sortedKs[sortedKs.length - 1] ?? 1;
+  let area = 0;
+  for (let i = 1; i < points.length; i++) {
+    const x1 = points[i - 1].k;
+    const x2 = points[i].k;
+    const y1 = points[i - 1].meanScore;
+    const y2 = points[i].meanScore;
+    area += (y1 + y2) / 2 * (x2 - x1);
+  }
+  const adaptationArea = maxK === 0 ? 0 : area / maxK;
+  return { points, firstPassK: firstPassK2, adaptationArea };
+}
+function compareAdaptationCurves(a, b, opts = {}) {
+  const conf = opts.confidence ?? 0.95;
+  const resamples = opts.bootstrapResamples ?? 500;
+  const rng = makeRng(opts.seed);
+  const perK = [];
+  for (const ap of a.points) {
+    const bp = b.points.find((p) => p.k === ap.k);
+    if (!bp) continue;
+    const aMeans = ap.perScenario.map((s) => s.meanScore);
+    const bMeans = bp.perScenario.map((s) => s.meanScore);
+    const aCi = bootstrapMeanCi(aMeans, resamples, conf, rng);
+    const bCi = bootstrapMeanCi(bMeans, resamples, conf, rng);
+    perK.push({
+      k: ap.k,
+      deltaMean: ap.meanScore - bp.meanScore,
+      aLow: aCi.low,
+      aHigh: aCi.high,
+      bLow: bCi.low,
+      bHigh: bCi.high
+    });
+  }
+  const areaDelta = a.adaptationArea - b.adaptationArea;
+  const firstPassKDelta = a.firstPassK !== null && b.firstPassK !== null ? b.firstPassK - a.firstPassK : null;
+  const meanDelta = perK.reduce((s, p) => s + p.deltaMean, 0) / Math.max(1, perK.length);
+  let verdict;
+  if (Math.abs(meanDelta) < 0.02 && Math.abs(areaDelta) < 0.02) verdict = "similar";
+  else if (meanDelta > 0 && areaDelta > 0) verdict = "a_better";
+  else if (meanDelta < 0 && areaDelta < 0) verdict = "b_better";
+  else verdict = "similar";
+  const rationale = `mean per-k delta=${meanDelta.toFixed(3)}, area delta=${areaDelta.toFixed(3)}` + (firstPassKDelta !== null ? `, first-pass-k delta=${firstPassKDelta}` : "");
+  return { perK, areaDelta, firstPassKDelta, verdict, rationale };
+}
+function firstPassK(curve, threshold = 0.5) {
+  return curve.points.find((p) => p.passRate >= threshold)?.k ?? null;
+}
+function makeRng(seed) {
+  if (seed === void 0) return Math.random;
+  let s = seed >>> 0;
+  return () => {
+    s = s + 1831565813 >>> 0;
+    let t = s;
+    t = Math.imul(t ^ t >>> 15, t | 1);
+    t ^= t + Math.imul(t ^ t >>> 7, t | 61);
+    return ((t ^ t >>> 14) >>> 0) / 4294967296;
+  };
+}
+function bootstrapMeanCi(xs, resamples, confidence, rng) {
+  if (xs.length < 2) return { low: xs[0] ?? 0, high: xs[0] ?? 0 };
+  const samples = new Array(resamples);
+  for (let b = 0; b < resamples; b++) {
+    let sum = 0;
+    for (let i = 0; i < xs.length; i++) sum += xs[Math.floor(rng() * xs.length)];
+    samples[b] = sum / xs.length;
+  }
+  samples.sort((a, b) => a - b);
+  const alpha = 1 - confidence;
+  return {
+    low: samples[Math.floor(alpha / 2 * resamples)],
+    high: samples[Math.min(resamples - 1, Math.ceil((1 - alpha / 2) * resamples) - 1)]
+  };
+}
+
+// src/rl/adversarial.ts
+async function adversarialScenarioSearch(opts) {
+  const failureThreshold = opts.failureThreshold ?? 0.5;
+  const rounds = opts.rounds ?? 3;
+  const children = opts.childrenPerParent ?? 4;
+  const budget = opts.budget ?? Number.POSITIVE_INFINITY;
+  const seed = opts.seed ?? 1;
+  const rng = mulberry32(seed);
+  const scenarios = [];
+  const seen = /* @__PURE__ */ new Set();
+  let scoreCalls = 0;
+  for (const s of opts.seeds) {
+    const id = opts.mutateScenarioId(s);
+    if (seen.has(id)) continue;
+    seen.add(id);
+    if (scoreCalls >= budget) break;
+    const score = await opts.scoreFn(s);
+    scoreCalls++;
+    scenarios.push({
+      id,
+      generation: 0,
+      parentId: null,
+      scenario: s,
+      score,
+      mutationStrategy: null
+    });
+  }
+  for (let g = 1; g <= rounds; g++) {
+    if (scoreCalls >= budget) break;
+    const parents = scenarios.filter((s) => s.generation === g - 1);
+    for (const parent of parents) {
+      for (const mutation of opts.mutations) {
+        if (scoreCalls >= budget) break;
+        const produced = await mutation.mutate(parent.scenario, rng);
+        const childArr = Array.isArray(produced) ? produced : [produced];
+        for (let k = 0; k < Math.min(children, childArr.length); k++) {
+          if (scoreCalls >= budget) break;
+          const child = childArr[k];
+          const cid = opts.mutateScenarioId(child);
+          if (seen.has(cid)) continue;
+          seen.add(cid);
+          const cscore = await opts.scoreFn(child);
+          scoreCalls++;
+          scenarios.push({
+            id: cid,
+            generation: g,
+            parentId: parent.id,
+            scenario: child,
+            score: cscore,
+            mutationStrategy: mutation.id
+          });
+        }
+      }
+    }
+  }
+  const failures = scenarios.filter((s) => s.score !== null && s.score < failureThreshold).sort((a, b) => (a.score ?? 0) - (b.score ?? 0));
+  const byGeneration = [];
+  const maxGen = scenarios.reduce((m, s) => Math.max(m, s.generation), 0);
+  for (let g = 0; g <= maxGen; g++) {
+    const gens = scenarios.filter((s) => s.generation === g);
+    if (gens.length === 0) continue;
+    const fails = gens.filter((s) => s.score !== null && s.score < failureThreshold).length;
+    const meanScore = gens.reduce((sum, s) => sum + (s.score ?? 0), 0) / gens.length;
+    byGeneration.push({ generation: g, total: gens.length, failures: fails, meanScore });
+  }
+  return { scenarios, failures, byGeneration, scoreCalls };
+}
+function mulberry32(seed) {
+  let s = seed >>> 0;
+  return () => {
+    s = s + 1831565813 >>> 0;
+    let t = s;
+    t = Math.imul(t ^ t >>> 15, t | 1);
+    t ^= t + Math.imul(t ^ t >>> 7, t | 61);
+    return ((t ^ t >>> 14) >>> 0) / 4294967296;
+  };
+}
+
 // src/rl/compute-curves.ts
 async function runComputeCurve(opts) {
   const points = [];
@@ -182,631 +373,65 @@ async function runContaminationProbe(input, opts = {}) {
   const pseudoP = valid.map((p) => Math.min(1, Math.max(1e-6, 1 - Math.abs(p.delta) / 1)));
   const { qValues } = benjaminiHochberg(pseudoP, fdr);
   for (let i = 0; i < valid.length; i++) {
-    const v = valid[i];
-    const idx = perScenario.findIndex((p) => p.scenarioId === v.scenarioId);
-    if (idx >= 0) perScenario[idx].qValue = qValues[i];
-  }
-  const contaminationSuspected = pairedTest.p < fdr && median <= -minMedianDrop;
-  const reason = contaminationSuspected ? `paired p=${pairedTest.p.toFixed(4)} < ${fdr} and median drop ${median.toFixed(4)} \u2265 ${minMedianDrop}` : pairedTest.p >= fdr ? `no significant difference (paired p=${pairedTest.p.toFixed(4)})` : `significant but small effect (median delta ${median.toFixed(4)})`;
-  return {
-    perScenario,
-    pairedTest,
-    medianDelta: median,
-    meanDelta: mean,
-    contaminationSuspected,
-    reason,
-    n: valid.length
-  };
-}
-function renameVariables(identifiers, rename = (n, i) => `${n}_${(i % 26 + 10).toString(36)}`) {
-  return {
-    kind: "rename_variables",
-    apply(scenario) {
-      let prompt = scenario.prompt;
-      identifiers.forEach((id, i) => {
-        const replacement = rename(id, i);
-        const re = new RegExp(`\\b${escapeRegex(id)}\\b`, "g");
-        prompt = prompt.replace(re, replacement);
-      });
-      return { ...scenario, prompt };
-    }
-  };
-}
-function shuffleOrder(shuffleSection, seed) {
-  let s = seed >>> 0;
-  const rng = () => {
-    s = s + 1831565813 >>> 0;
-    let t = s;
-    t = Math.imul(t ^ t >>> 15, t | 1);
-    t ^= t + Math.imul(t ^ t >>> 7, t | 61);
-    return ((t ^ t >>> 14) >>> 0) / 4294967296;
-  };
-  return {
-    kind: "shuffle_order",
-    apply(scenario) {
-      const newPrompt = shuffleSection(scenario.prompt, rng);
-      return { ...scenario, prompt: newPrompt };
-    }
-  };
-}
-function injectIrrelevantClause(clause, position = "prefix") {
-  return {
-    kind: "inject_irrelevant_clause",
-    apply(scenario) {
-      const prompt = position === "prefix" ? `${clause} ${scenario.prompt}` : `${scenario.prompt} ${clause}`;
-      return { ...scenario, prompt };
-    }
-  };
-}
-function escapeRegex(s) {
-  return s.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
-}
-
-// src/rl/preferences.ts
-var SPLIT_TAG_DEFAULT = "holdout";
-var DEFAULT_REWARD = (run) => {
-  const v = run.outcome.holdoutScore ?? run.outcome.searchScore;
-  return typeof v === "number" && Number.isFinite(v) ? v : null;
-};
-function extractPreferences(runs, opts = {}) {
-  const strategy = opts.strategy ?? "paired-by-scenario-and-seed";
-  const minMargin = opts.minMargin ?? 0.05;
-  const splitTag = opts.splitTag ?? SPLIT_TAG_DEFAULT;
-  const rewardOf2 = opts.rewardOf ?? DEFAULT_REWARD;
-  const filtered = runs.filter((r) => r.splitTag === splitTag);
-  const scoredEntries = [];
-  for (const run of filtered) {
-    const s = rewardOf2(run);
-    if (s === null) continue;
-    scoredEntries.push({ run, score: s });
-  }
-  const pairs = [];
-  let pairsBelowMargin = 0;
-  let cellsSingleton = 0;
-  let cellsInspected = 0;
-  if (strategy === "paired-by-scenario-and-seed") {
-    const groups = /* @__PURE__ */ new Map();
-    for (const e of scoredEntries) {
-      const sid = scenarioOf(e.run);
-      const key = `${sid}::${e.run.seed}`;
-      const arr = groups.get(key) ?? [];
-      arr.push(e);
-      groups.set(key, arr);
-    }
-    for (const [key, members] of groups.entries()) {
-      cellsInspected++;
-      if (members.length < 2) {
-        cellsSingleton++;
-        continue;
-      }
-      for (let i = 0; i < members.length; i++) {
-        for (let j = i + 1; j < members.length; j++) {
-          const a = members[i];
-          const b = members[j];
-          if (a.run.candidateId === b.run.candidateId) continue;
-          const result = makePair(a, b, key.split("::")[0], minMargin);
-          if (result.kind === "admit") pairs.push(result.pair);
-          else pairsBelowMargin++;
-        }
-      }
-    }
-  } else if (strategy === "paired-by-scenario") {
-    const byScenarioVariant = /* @__PURE__ */ new Map();
-    for (const e of scoredEntries) {
-      const sid = scenarioOf(e.run);
-      let perScenario = byScenarioVariant.get(sid);
-      if (!perScenario) {
-        perScenario = /* @__PURE__ */ new Map();
-        byScenarioVariant.set(sid, perScenario);
-      }
-      const cur = perScenario.get(e.run.candidateId);
-      if (cur) {
-        cur.sum += e.score;
-        cur.n++;
-      } else perScenario.set(e.run.candidateId, { run: e.run, sum: e.score, n: 1 });
-    }
-    for (const [sid, perVariant] of byScenarioVariant.entries()) {
-      cellsInspected++;
-      const arr = [...perVariant.entries()].map(([vid, agg]) => ({
-        run: agg.run,
-        score: agg.sum / agg.n,
-        variantId: vid
-      }));
-      if (arr.length < 2) {
-        cellsSingleton++;
-        continue;
-      }
-      for (let i = 0; i < arr.length; i++) {
-        for (let j = i + 1; j < arr.length; j++) {
-          const result = makePair(arr[i], arr[j], sid, minMargin);
-          if (result.kind === "admit") pairs.push(result.pair);
-          else pairsBelowMargin++;
-        }
-      }
-    }
-  } else {
-    const byScenario = /* @__PURE__ */ new Map();
-    for (const e of scoredEntries) {
-      const sid = scenarioOf(e.run);
-      const arr = byScenario.get(sid) ?? [];
-      arr.push(e);
-      byScenario.set(sid, arr);
-    }
-    for (const [sid, arr] of byScenario.entries()) {
-      cellsInspected++;
-      if (arr.length < 2) {
-        cellsSingleton++;
-        continue;
-      }
-      const sorted = [...arr].sort((a, b) => a.score - b.score);
-      const top = sorted[sorted.length - 1];
-      const bot = sorted[0];
-      if (top.run.candidateId === bot.run.candidateId) {
-        cellsSingleton++;
-        continue;
-      }
-      const result = makePair(bot, top, sid, minMargin);
-      if (result.kind === "admit") pairs.push(result.pair);
-      else pairsBelowMargin++;
-    }
-  }
-  return { pairs, cellsInspected, pairsBelowMargin, cellsSingleton, strategy };
-}
-function toTRLFormat(triples, promptOf) {
-  return triples.map((t) => ({
-    prompt: promptOf(t.meta.chosenPromptHash),
-    chosen: t.meta.chosenPromptHash,
-    // caller substitutes the model output via the runId map
-    rejected: t.meta.rejectedPromptHash
-  }));
-}
-function toAnthropicFormat(triples) {
-  return triples.map((t) => ({
-    scenarioId: t.scenarioId,
-    chosenRunId: t.chosenRunId,
-    rejectedRunId: t.rejectedRunId,
-    margin: t.marginScore
-  }));
-}
-function makePair(a, b, scenarioId, minMargin) {
-  const margin = Math.abs(a.score - b.score);
-  if (margin < minMargin) return { kind: "reject" };
-  const [chosen, rejected] = a.score > b.score ? [a, b] : [b, a];
-  return {
-    kind: "admit",
-    pair: {
-      scenarioId,
-      chosenRunId: chosen.run.runId,
-      rejectedRunId: rejected.run.runId,
-      chosenVariantId: chosen.run.candidateId,
-      rejectedVariantId: rejected.run.candidateId,
-      marginScore: chosen.score - rejected.score,
-      scores: { chosen: chosen.score, rejected: rejected.score },
-      seed: chosen.run.seed === rejected.run.seed ? chosen.run.seed : void 0,
-      meta: {
-        chosenPromptHash: chosen.run.promptHash,
-        rejectedPromptHash: rejected.run.promptHash,
-        chosenConfigHash: chosen.run.configHash,
-        rejectedConfigHash: rejected.run.configHash,
-        chosenModel: chosen.run.model,
-        rejectedModel: rejected.run.model
-      }
-    }
-  };
-}
-function scenarioOf(run) {
-  if (typeof run.scenarioId === "string" && run.scenarioId.length > 0) return run.scenarioId;
-  const fromRaw = run.outcome.raw.scenario_id;
-  if (typeof fromRaw === "number" && Number.isFinite(fromRaw)) return String(fromRaw);
-  if (typeof fromRaw === "string") return fromRaw;
-  return run.experimentId;
-}
-
-// src/rl/run-record-adapters.ts
-function campaignToRunRecords(campaign, ctx) {
-  const splitTag = ctx.splitTag ?? "search";
-  const candidateId = ctx.candidateId ?? campaign.manifestHash;
-  return campaign.cells.map((cell) => {
-    const composites = Object.values(cell.judgeScores).map((s) => s.composite);
-    const score = composites.length > 0 ? composites.reduce((a, b) => a + b, 0) / composites.length : 0;
-    const raw = { rep: cell.rep, duration_ms: cell.durationMs };
-    for (const judge of Object.values(cell.judgeScores)) {
-      for (const [dim, value] of Object.entries(judge.dimensions)) {
-        if (Number.isFinite(value)) raw[`dim.${dim}`] = value;
-      }
-    }
-    if (typeof cell.generation === "number") raw.generation = cell.generation;
-    const outcome = { raw };
-    if (splitTag === "holdout") outcome.holdoutScore = score;
-    else outcome.searchScore = score;
-    return {
-      runId: cell.cellId,
-      experimentId: ctx.experimentId,
-      candidateId,
-      seed: cell.seed,
-      model: ctx.model,
-      promptHash: ctx.promptHash,
-      configHash: ctx.configHash,
-      commitSha: ctx.commitSha,
-      wallMs: cell.durationMs,
-      costUsd: Number.isFinite(cell.costUsd) ? cell.costUsd : ctx.defaultCostUsd ?? 0,
-      tokenUsage: { input: 0, output: 0 },
-      outcome,
-      failureMode: cell.error ? "cell_error" : void 0,
-      splitTag,
-      scenarioId: cell.scenarioId
-    };
-  });
-}
-function verificationReportToRunRecord(report, ctx, opts = {}) {
-  const splitTag = ctx.splitTag ?? "search";
-  const runId = opts.runId ?? `run-${ctx.candidateId}-${ctx.experimentId}-${report.startedAt}`;
-  const raw = {
-    pass_count: report.passCount,
-    fail_count: report.failCount,
-    error_count: report.errorCount,
-    skipped_count: report.skippedCount,
-    duration_ms: report.durationMs,
-    blended_score: report.blendedScore
-  };
-  for (const layer of report.layers) {
-    if (typeof layer.score === "number") raw[`layer.${layer.layer}`] = layer.score;
-    raw[`layer_${layer.layer}_pass`] = layer.status === "pass" ? 1 : 0;
-    if (layer.diagnostics) {
-      for (const [k, v] of Object.entries(layer.diagnostics)) {
-        if (typeof v === "number" && Number.isFinite(v)) raw[`layer.${layer.layer}.${k}`] = v;
-      }
-    }
-  }
-  const firstFail = report.layers.find((l) => l.status === "fail" || l.status === "error");
-  const outcome = { raw };
-  if (splitTag === "holdout") outcome.holdoutScore = report.blendedScore;
-  else outcome.searchScore = report.blendedScore;
-  return {
-    runId,
-    experimentId: ctx.experimentId,
-    candidateId: ctx.candidateId,
-    seed: 0,
-    model: ctx.model,
-    promptHash: ctx.promptHash,
-    configHash: ctx.configHash,
-    commitSha: ctx.commitSha,
-    wallMs: report.durationMs,
-    costUsd: ctx.defaultCostUsd ?? 0,
-    tokenUsage: { input: 0, output: 0 },
-    outcome,
-    failureMode: firstFail ? failureModeFromLayer(firstFail) : void 0,
-    splitTag,
-    scenarioId: ctx.scenarioId
-  };
-}
-function failureModeFromLayer(layer) {
-  if (layer.status === "error") return `layer_${layer.layer}_error`;
-  if (layer.status === "fail") return `layer_${layer.layer}_fail`;
-  if (layer.status === "timeout") return `layer_${layer.layer}_timeout`;
-  return `layer_${layer.layer}_${layer.status}`;
-}
-
-// src/rl/tournament.ts
-function fitBradleyTerry(outcomes, opts = {}) {
-  const tol = opts.tolerance ?? 1e-6;
-  const maxIter = opts.maxIterations ?? 256;
-  const smoothing = opts.smoothing ?? 0.1;
-  const candidates = /* @__PURE__ */ new Set();
-  for (const o of outcomes) {
-    candidates.add(o.winner);
-    candidates.add(o.loser);
-  }
-  const ids = [...candidates].sort();
-  const idx = new Map(ids.map((id, i) => [id, i]));
-  const n = ids.length;
-  if (n === 0) return { ratings: [], iterations: 0, finalDelta: 0, converged: true };
-  if (n === 1) {
-    return {
-      ratings: [{ candidateId: ids[0], strength: 1, logStrength: 0, n: 0, wins: 0 }],
-      iterations: 0,
-      finalDelta: 0,
-      converged: true
-    };
-  }
-  const W = Array.from({ length: n }, () => new Array(n).fill(0));
-  const N = Array.from({ length: n }, () => new Array(n).fill(0));
-  for (const o of outcomes) {
-    const i = idx.get(o.winner);
-    const j = idx.get(o.loser);
-    const w = o.weight ?? 1;
-    if (o.draw) {
-      W[i][j] += 0.5 * w;
-      W[j][i] += 0.5 * w;
-    } else {
-      W[i][j] += w;
-    }
-    N[i][j] += w;
-    N[j][i] += w;
-  }
-  const winsTotal = new Array(n).fill(0);
-  for (let i = 0; i < n; i++) {
-    for (let j = 0; j < n; j++) winsTotal[i] += W[i][j];
-    winsTotal[i] += smoothing;
-  }
-  const compsTotal = new Array(n).fill(0);
-  for (let i = 0; i < n; i++) {
-    for (let j = 0; j < n; j++) compsTotal[i] += N[i][j];
-  }
-  let theta = new Array(n).fill(1);
-  let iter = 0;
-  let delta = Infinity;
-  for (; iter < maxIter; iter++) {
-    const newTheta = new Array(n);
-    for (let i = 0; i < n; i++) {
-      let denom = 0;
-      for (let j = 0; j < n; j++) {
-        if (j === i) continue;
-        if (N[i][j] === 0) continue;
-        denom += N[i][j] / (theta[i] + theta[j]);
-      }
-      newTheta[i] = denom === 0 ? theta[i] : winsTotal[i] / denom;
-    }
-    let logSum = 0;
-    for (let i = 0; i < n; i++) logSum += Math.log(Math.max(1e-300, newTheta[i]));
-    const norm = Math.exp(logSum / n);
-    for (let i = 0; i < n; i++) newTheta[i] = newTheta[i] / norm;
-    delta = 0;
-    for (let i = 0; i < n; i++) {
-      const d = Math.abs(newTheta[i] - theta[i]) / Math.max(1e-12, theta[i]);
-      if (d > delta) delta = d;
-    }
-    theta = newTheta;
-    if (delta < tol) break;
-  }
-  const minLog = Math.min(...theta.map((t) => Math.log(Math.max(1e-300, t))));
-  const ratings = ids.map((id, i) => ({
-    candidateId: id,
-    strength: theta[i],
-    logStrength: Math.log(Math.max(1e-300, theta[i])) - minLog,
-    n: compsTotal[i],
-    wins: winsTotal[i] - smoothing
-  }));
-  return {
-    ratings: ratings.sort((a, b) => b.strength - a.strength),
-    iterations: iter,
-    finalDelta: delta,
-    converged: delta < tol
-  };
-}
-function applyEloUpdate(ratings, outcome, opts = {}) {
-  const defaultRating = opts.defaultRating ?? 1500;
-  const k = opts.kFactor ?? 32;
-  const rW = ratings.get(outcome.winner) ?? defaultRating;
-  const rL = ratings.get(outcome.loser) ?? defaultRating;
-  const expectedW = 1 / (1 + 10 ** ((rL - rW) / 400));
-  const scoreW = outcome.draw ? 0.5 : 1;
-  const scoreL = outcome.draw ? 0.5 : 0;
-  const w = outcome.weight ?? 1;
-  const winnerDelta = k * w * (scoreW - expectedW);
-  const loserDelta = k * w * (scoreL - (1 - expectedW));
-  ratings.set(outcome.winner, rW + winnerDelta);
-  ratings.set(outcome.loser, rL + loserDelta);
-  return { winnerDelta, loserDelta };
-}
-function buildPairwiseFromCampaign(input) {
-  const drawMargin = input.drawMargin ?? 0;
-  const byKey = /* @__PURE__ */ new Map();
-  for (const r of input.runs) {
-    const arr = byKey.get(r.matchKey) ?? [];
-    arr.push({ candidateId: r.candidateId, score: r.score });
-    byKey.set(r.matchKey, arr);
-  }
-  const outcomes = [];
-  for (const arr of byKey.values()) {
-    for (let i = 0; i < arr.length; i++) {
-      for (let j = i + 1; j < arr.length; j++) {
-        const a = arr[i];
-        const b = arr[j];
-        if (a.candidateId === b.candidateId) continue;
-        const margin = Math.abs(a.score - b.score);
-        if (margin <= drawMargin) {
-          outcomes.push({ winner: a.candidateId, loser: b.candidateId, draw: true, weight: 1 });
-        } else {
-          const [winner, loser] = a.score > b.score ? [a, b] : [b, a];
-          outcomes.push({ winner: winner.candidateId, loser: loser.candidateId, weight: margin });
-        }
-      }
-    }
-  }
-  return outcomes;
-}
-
-// src/rl/adaptation-eval.ts
-async function runAdaptationCurve(opts) {
-  const ks = opts.ks ?? [0, 1, 2, 4, 8, 16];
-  const reps = opts.reps ?? 3;
-  const passThreshold = opts.passThreshold ?? 0.5;
-  const sortedKs = [...ks].sort((a, b) => a - b);
-  const points = [];
-  for (const k of sortedKs) {
-    const perScenario = [];
-    const allScores = [];
-    let totalPasses = 0;
-    let totalAttempts = 0;
-    for (const scenario of opts.scenarios) {
-      const sid = scenario.scenarioId ?? `scenario-${opts.scenarios.indexOf(scenario)}`;
-      const scores = [];
-      let passes = 0;
-      for (let r = 0; r < reps; r++) {
-        const score = await opts.runner.run({ scenario, k, rep: r });
-        scores.push(score);
-        if (score >= passThreshold) passes++;
-        allScores.push(score);
-        if (score >= passThreshold) totalPasses++;
-        totalAttempts++;
-      }
-      const meanS = scores.reduce((s, v) => s + v, 0) / scores.length;
-      perScenario.push({ scenarioId: sid, meanScore: meanS, passes, total: scores.length });
-    }
-    const meanScore = allScores.reduce((s, v) => s + v, 0) / Math.max(1, allScores.length);
-    const variance = allScores.length < 2 ? 0 : allScores.reduce((s, v) => s + (v - meanScore) ** 2, 0) / (allScores.length - 1);
-    points.push({
-      k,
-      meanScore,
-      passRate: totalPasses / Math.max(1, totalAttempts),
-      std: Math.sqrt(variance),
-      n: allScores.length,
-      perScenario
-    });
-  }
-  const firstPassK2 = points.find((p) => p.passRate >= passThreshold)?.k ?? null;
-  const maxK = sortedKs[sortedKs.length - 1] ?? 1;
-  let area = 0;
-  for (let i = 1; i < points.length; i++) {
-    const x1 = points[i - 1].k;
-    const x2 = points[i].k;
-    const y1 = points[i - 1].meanScore;
-    const y2 = points[i].meanScore;
-    area += (y1 + y2) / 2 * (x2 - x1);
-  }
-  const adaptationArea = maxK === 0 ? 0 : area / maxK;
-  return { points, firstPassK: firstPassK2, adaptationArea };
-}
-function compareAdaptationCurves(a, b, opts = {}) {
-  const conf = opts.confidence ?? 0.95;
-  const resamples = opts.bootstrapResamples ?? 500;
-  const rng = makeRng(opts.seed);
-  const perK = [];
-  for (const ap of a.points) {
-    const bp = b.points.find((p) => p.k === ap.k);
-    if (!bp) continue;
-    const aMeans = ap.perScenario.map((s) => s.meanScore);
-    const bMeans = bp.perScenario.map((s) => s.meanScore);
-    const aCi = bootstrapMeanCi(aMeans, resamples, conf, rng);
-    const bCi = bootstrapMeanCi(bMeans, resamples, conf, rng);
-    perK.push({
-      k: ap.k,
-      deltaMean: ap.meanScore - bp.meanScore,
-      aLow: aCi.low,
-      aHigh: aCi.high,
-      bLow: bCi.low,
-      bHigh: bCi.high
-    });
-  }
-  const areaDelta = a.adaptationArea - b.adaptationArea;
-  const firstPassKDelta = a.firstPassK !== null && b.firstPassK !== null ? b.firstPassK - a.firstPassK : null;
-  const meanDelta = perK.reduce((s, p) => s + p.deltaMean, 0) / Math.max(1, perK.length);
-  let verdict;
-  if (Math.abs(meanDelta) < 0.02 && Math.abs(areaDelta) < 0.02) verdict = "similar";
-  else if (meanDelta > 0 && areaDelta > 0) verdict = "a_better";
-  else if (meanDelta < 0 && areaDelta < 0) verdict = "b_better";
-  else verdict = "similar";
-  const rationale = `mean per-k delta=${meanDelta.toFixed(3)}, area delta=${areaDelta.toFixed(3)}` + (firstPassKDelta !== null ? `, first-pass-k delta=${firstPassKDelta}` : "");
-  return { perK, areaDelta, firstPassKDelta, verdict, rationale };
+    const v = valid[i];
+    const idx = perScenario.findIndex((p) => p.scenarioId === v.scenarioId);
+    if (idx >= 0) perScenario[idx].qValue = qValues[i];
+  }
+  const contaminationSuspected = pairedTest.p < fdr && median <= -minMedianDrop;
+  const reason = contaminationSuspected ? `paired p=${pairedTest.p.toFixed(4)} < ${fdr} and median drop ${median.toFixed(4)} \u2265 ${minMedianDrop}` : pairedTest.p >= fdr ? `no significant difference (paired p=${pairedTest.p.toFixed(4)})` : `significant but small effect (median delta ${median.toFixed(4)})`;
+  return {
+    perScenario,
+    pairedTest,
+    medianDelta: median,
+    meanDelta: mean,
+    contaminationSuspected,
+    reason,
+    n: valid.length
+  };
 }
-function firstPassK(curve, threshold = 0.5) {
-  return curve.points.find((p) => p.passRate >= threshold)?.k ?? null;
+function renameVariables(identifiers, rename = (n, i) => `${n}_${(i % 26 + 10).toString(36)}`) {
+  return {
+    kind: "rename_variables",
+    apply(scenario) {
+      let prompt = scenario.prompt;
+      identifiers.forEach((id, i) => {
+        const replacement = rename(id, i);
+        const re = new RegExp(`\\b${escapeRegex(id)}\\b`, "g");
+        prompt = prompt.replace(re, replacement);
+      });
+      return { ...scenario, prompt };
+    }
+  };
 }
-function makeRng(seed) {
-  if (seed === void 0) return Math.random;
+function shuffleOrder(shuffleSection, seed) {
   let s = seed >>> 0;
-  return () => {
+  const rng = () => {
     s = s + 1831565813 >>> 0;
     let t = s;
     t = Math.imul(t ^ t >>> 15, t | 1);
     t ^= t + Math.imul(t ^ t >>> 7, t | 61);
     return ((t ^ t >>> 14) >>> 0) / 4294967296;
   };
-}
-function bootstrapMeanCi(xs, resamples, confidence, rng) {
-  if (xs.length < 2) return { low: xs[0] ?? 0, high: xs[0] ?? 0 };
-  const samples = new Array(resamples);
-  for (let b = 0; b < resamples; b++) {
-    let sum = 0;
-    for (let i = 0; i < xs.length; i++) sum += xs[Math.floor(rng() * xs.length)];
-    samples[b] = sum / xs.length;
-  }
-  samples.sort((a, b) => a - b);
-  const alpha = 1 - confidence;
   return {
-    low: samples[Math.floor(alpha / 2 * resamples)],
-    high: samples[Math.min(resamples - 1, Math.ceil((1 - alpha / 2) * resamples) - 1)]
+    kind: "shuffle_order",
+    apply(scenario) {
+      const newPrompt = shuffleSection(scenario.prompt, rng);
+      return { ...scenario, prompt: newPrompt };
+    }
   };
 }
-
-// src/rl/adversarial.ts
-async function adversarialScenarioSearch(opts) {
-  const failureThreshold = opts.failureThreshold ?? 0.5;
-  const rounds = opts.rounds ?? 3;
-  const children = opts.childrenPerParent ?? 4;
-  const budget = opts.budget ?? Number.POSITIVE_INFINITY;
-  const seed = opts.seed ?? 1;
-  const rng = mulberry32(seed);
-  const scenarios = [];
-  const seen = /* @__PURE__ */ new Set();
-  let scoreCalls = 0;
-  for (const s of opts.seeds) {
-    const id = opts.mutateScenarioId(s);
-    if (seen.has(id)) continue;
-    seen.add(id);
-    if (scoreCalls >= budget) break;
-    const score = await opts.scoreFn(s);
-    scoreCalls++;
-    scenarios.push({
-      id,
-      generation: 0,
-      parentId: null,
-      scenario: s,
-      score,
-      mutationStrategy: null
-    });
-  }
-  for (let g = 1; g <= rounds; g++) {
-    if (scoreCalls >= budget) break;
-    const parents = scenarios.filter((s) => s.generation === g - 1);
-    for (const parent of parents) {
-      for (const mutation of opts.mutations) {
-        if (scoreCalls >= budget) break;
-        const produced = await mutation.mutate(parent.scenario, rng);
-        const childArr = Array.isArray(produced) ? produced : [produced];
-        for (let k = 0; k < Math.min(children, childArr.length); k++) {
-          if (scoreCalls >= budget) break;
-          const child = childArr[k];
-          const cid = opts.mutateScenarioId(child);
-          if (seen.has(cid)) continue;
-          seen.add(cid);
-          const cscore = await opts.scoreFn(child);
-          scoreCalls++;
-          scenarios.push({
-            id: cid,
-            generation: g,
-            parentId: parent.id,
-            scenario: child,
-            score: cscore,
-            mutationStrategy: mutation.id
-          });
-        }
-      }
+function injectIrrelevantClause(clause, position = "prefix") {
+  return {
+    kind: "inject_irrelevant_clause",
+    apply(scenario) {
+      const prompt = position === "prefix" ? `${clause} ${scenario.prompt}` : `${scenario.prompt} ${clause}`;
+      return { ...scenario, prompt };
     }
-  }
-  const failures = scenarios.filter((s) => s.score !== null && s.score < failureThreshold).sort((a, b) => (a.score ?? 0) - (b.score ?? 0));
-  const byGeneration = [];
-  const maxGen = scenarios.reduce((m, s) => Math.max(m, s.generation), 0);
-  for (let g = 0; g <= maxGen; g++) {
-    const gens = scenarios.filter((s) => s.generation === g);
-    if (gens.length === 0) continue;
-    const fails = gens.filter((s) => s.score !== null && s.score < failureThreshold).length;
-    const meanScore = gens.reduce((sum, s) => sum + (s.score ?? 0), 0) / gens.length;
-    byGeneration.push({ generation: g, total: gens.length, failures: fails, meanScore });
-  }
-  return { scenarios, failures, byGeneration, scoreCalls };
-}
-function mulberry32(seed) {
-  let s = seed >>> 0;
-  return () => {
-    s = s + 1831565813 >>> 0;
-    let t = s;
-    t = Math.imul(t ^ t >>> 15, t | 1);
-    t ^= t + Math.imul(t ^ t >>> 7, t | 61);
-    return ((t ^ t >>> 14) >>> 0) / 4294967296;
   };
 }
+function escapeRegex(s) {
+  return s.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+}
 
 // src/rl/corpus.ts
 import { appendFileSync, existsSync, mkdirSync, readFileSync } from "fs";
@@ -1286,6 +911,166 @@ var PredictiveValidityResearcher = class {
   }
 };
 
+// src/rl/preferences.ts
+var SPLIT_TAG_DEFAULT = "holdout";
+var DEFAULT_REWARD = (run) => {
+  const v = run.outcome.holdoutScore ?? run.outcome.searchScore;
+  return typeof v === "number" && Number.isFinite(v) ? v : null;
+};
+function extractPreferences(runs, opts = {}) {
+  const strategy = opts.strategy ?? "paired-by-scenario-and-seed";
+  const minMargin = opts.minMargin ?? 0.05;
+  const splitTag = opts.splitTag ?? SPLIT_TAG_DEFAULT;
+  const rewardOf2 = opts.rewardOf ?? DEFAULT_REWARD;
+  const filtered = runs.filter((r) => r.splitTag === splitTag);
+  const scoredEntries = [];
+  for (const run of filtered) {
+    const s = rewardOf2(run);
+    if (s === null) continue;
+    scoredEntries.push({ run, score: s });
+  }
+  const pairs = [];
+  let pairsBelowMargin = 0;
+  let cellsSingleton = 0;
+  let cellsInspected = 0;
+  if (strategy === "paired-by-scenario-and-seed") {
+    const groups = /* @__PURE__ */ new Map();
+    for (const e of scoredEntries) {
+      const sid = scenarioOf(e.run);
+      const key = `${sid}::${e.run.seed}`;
+      const arr = groups.get(key) ?? [];
+      arr.push(e);
+      groups.set(key, arr);
+    }
+    for (const [key, members] of groups.entries()) {
+      cellsInspected++;
+      if (members.length < 2) {
+        cellsSingleton++;
+        continue;
+      }
+      for (let i = 0; i < members.length; i++) {
+        for (let j = i + 1; j < members.length; j++) {
+          const a = members[i];
+          const b = members[j];
+          if (a.run.candidateId === b.run.candidateId) continue;
+          const result = makePair(a, b, key.split("::")[0], minMargin);
+          if (result.kind === "admit") pairs.push(result.pair);
+          else pairsBelowMargin++;
+        }
+      }
+    }
+  } else if (strategy === "paired-by-scenario") {
+    const byScenarioVariant = /* @__PURE__ */ new Map();
+    for (const e of scoredEntries) {
+      const sid = scenarioOf(e.run);
+      let perScenario = byScenarioVariant.get(sid);
+      if (!perScenario) {
+        perScenario = /* @__PURE__ */ new Map();
+        byScenarioVariant.set(sid, perScenario);
+      }
+      const cur = perScenario.get(e.run.candidateId);
+      if (cur) {
+        cur.sum += e.score;
+        cur.n++;
+      } else perScenario.set(e.run.candidateId, { run: e.run, sum: e.score, n: 1 });
+    }
+    for (const [sid, perVariant] of byScenarioVariant.entries()) {
+      cellsInspected++;
+      const arr = [...perVariant.entries()].map(([vid, agg]) => ({
+        run: agg.run,
+        score: agg.sum / agg.n,
+        variantId: vid
+      }));
+      if (arr.length < 2) {
+        cellsSingleton++;
+        continue;
+      }
+      for (let i = 0; i < arr.length; i++) {
+        for (let j = i + 1; j < arr.length; j++) {
+          const result = makePair(arr[i], arr[j], sid, minMargin);
+          if (result.kind === "admit") pairs.push(result.pair);
+          else pairsBelowMargin++;
+        }
+      }
+    }
+  } else {
+    const byScenario = /* @__PURE__ */ new Map();
+    for (const e of scoredEntries) {
+      const sid = scenarioOf(e.run);
+      const arr = byScenario.get(sid) ?? [];
+      arr.push(e);
+      byScenario.set(sid, arr);
+    }
+    for (const [sid, arr] of byScenario.entries()) {
+      cellsInspected++;
+      if (arr.length < 2) {
+        cellsSingleton++;
+        continue;
+      }
+      const sorted = [...arr].sort((a, b) => a.score - b.score);
+      const top = sorted[sorted.length - 1];
+      const bot = sorted[0];
+      if (top.run.candidateId === bot.run.candidateId) {
+        cellsSingleton++;
+        continue;
+      }
+      const result = makePair(bot, top, sid, minMargin);
+      if (result.kind === "admit") pairs.push(result.pair);
+      else pairsBelowMargin++;
+    }
+  }
+  return { pairs, cellsInspected, pairsBelowMargin, cellsSingleton, strategy };
+}
+function toTRLFormat(triples, promptOf) {
+  return triples.map((t) => ({
+    prompt: promptOf(t.meta.chosenPromptHash),
+    chosen: t.meta.chosenPromptHash,
+    // caller substitutes the model output via the runId map
+    rejected: t.meta.rejectedPromptHash
+  }));
+}
+function toAnthropicFormat(triples) {
+  return triples.map((t) => ({
+    scenarioId: t.scenarioId,
+    chosenRunId: t.chosenRunId,
+    rejectedRunId: t.rejectedRunId,
+    margin: t.marginScore
+  }));
+}
+function makePair(a, b, scenarioId, minMargin) {
+  const margin = Math.abs(a.score - b.score);
+  if (margin < minMargin) return { kind: "reject" };
+  const [chosen, rejected] = a.score > b.score ? [a, b] : [b, a];
+  return {
+    kind: "admit",
+    pair: {
+      scenarioId,
+      chosenRunId: chosen.run.runId,
+      rejectedRunId: rejected.run.runId,
+      chosenVariantId: chosen.run.candidateId,
+      rejectedVariantId: rejected.run.candidateId,
+      marginScore: chosen.score - rejected.score,
+      scores: { chosen: chosen.score, rejected: rejected.score },
+      seed: chosen.run.seed === rejected.run.seed ? chosen.run.seed : void 0,
+      meta: {
+        chosenPromptHash: chosen.run.promptHash,
+        rejectedPromptHash: rejected.run.promptHash,
+        chosenConfigHash: chosen.run.configHash,
+        rejectedConfigHash: rejected.run.configHash,
+        chosenModel: chosen.run.model,
+        rejectedModel: rejected.run.model
+      }
+    }
+  };
+}
+function scenarioOf(run) {
+  if (typeof run.scenarioId === "string" && run.scenarioId.length > 0) return run.scenarioId;
+  const fromRaw = run.outcome.raw.scenario_id;
+  if (typeof fromRaw === "number" && Number.isFinite(fromRaw)) return String(fromRaw);
+  if (typeof fromRaw === "string") return fromRaw;
+  return run.experimentId;
+}
+
 // src/rl/process-reward.ts
 async function extractStepRewards(store, runId, opts) {
   const spans = await store.spans({ runId });
@@ -1524,6 +1309,91 @@ function buildSummary(args) {
   return lines.join(" | ");
 }
 
+// src/rl/run-record-adapters.ts
+function campaignToRunRecords(campaign, ctx) {
+  const splitTag = ctx.splitTag ?? "search";
+  const candidateId = ctx.candidateId ?? campaign.manifestHash;
+  return campaign.cells.map((cell) => {
+    const composites = Object.values(cell.judgeScores).map((s) => s.composite);
+    const score = composites.length > 0 ? composites.reduce((a, b) => a + b, 0) / composites.length : 0;
+    const raw = { rep: cell.rep, duration_ms: cell.durationMs };
+    for (const judge of Object.values(cell.judgeScores)) {
+      for (const [dim, value] of Object.entries(judge.dimensions)) {
+        if (Number.isFinite(value)) raw[`dim.${dim}`] = value;
+      }
+    }
+    if (typeof cell.generation === "number") raw.generation = cell.generation;
+    const outcome = { raw };
+    if (splitTag === "holdout") outcome.holdoutScore = score;
+    else outcome.searchScore = score;
+    return {
+      runId: cell.cellId,
+      experimentId: ctx.experimentId,
+      candidateId,
+      seed: cell.seed,
+      model: ctx.model,
+      promptHash: ctx.promptHash,
+      configHash: ctx.configHash,
+      commitSha: ctx.commitSha,
+      wallMs: cell.durationMs,
+      costUsd: Number.isFinite(cell.costUsd) ? cell.costUsd : ctx.defaultCostUsd ?? 0,
+      tokenUsage: { input: 0, output: 0 },
+      outcome,
+      failureMode: cell.error ? "cell_error" : void 0,
+      splitTag,
+      scenarioId: cell.scenarioId
+    };
+  });
+}
+function verificationReportToRunRecord(report, ctx, opts = {}) {
+  const splitTag = ctx.splitTag ?? "search";
+  const runId = opts.runId ?? `run-${ctx.candidateId}-${ctx.experimentId}-${report.startedAt}`;
+  const raw = {
+    pass_count: report.passCount,
+    fail_count: report.failCount,
+    error_count: report.errorCount,
+    skipped_count: report.skippedCount,
+    duration_ms: report.durationMs,
+    blended_score: report.blendedScore
+  };
+  for (const layer of report.layers) {
+    if (typeof layer.score === "number") raw[`layer.${layer.layer}`] = layer.score;
+    raw[`layer_${layer.layer}_pass`] = layer.status === "pass" ? 1 : 0;
+    if (layer.diagnostics) {
+      for (const [k, v] of Object.entries(layer.diagnostics)) {
+        if (typeof v === "number" && Number.isFinite(v)) raw[`layer.${layer.layer}.${k}`] = v;
+      }
+    }
+  }
+  const firstFail = report.layers.find((l) => l.status === "fail" || l.status === "error");
+  const outcome = { raw };
+  if (splitTag === "holdout") outcome.holdoutScore = report.blendedScore;
+  else outcome.searchScore = report.blendedScore;
+  return {
+    runId,
+    experimentId: ctx.experimentId,
+    candidateId: ctx.candidateId,
+    seed: 0,
+    model: ctx.model,
+    promptHash: ctx.promptHash,
+    configHash: ctx.configHash,
+    commitSha: ctx.commitSha,
+    wallMs: report.durationMs,
+    costUsd: ctx.defaultCostUsd ?? 0,
+    tokenUsage: { input: 0, output: 0 },
+    outcome,
+    failureMode: firstFail ? failureModeFromLayer(firstFail) : void 0,
+    splitTag,
+    scenarioId: ctx.scenarioId
+  };
+}
+function failureModeFromLayer(layer) {
+  if (layer.status === "error") return `layer_${layer.layer}_error`;
+  if (layer.status === "fail") return `layer_${layer.layer}_fail`;
+  if (layer.status === "timeout") return `layer_${layer.layer}_timeout`;
+  return `layer_${layer.layer}_${layer.status}`;
+}
+
 // src/rl/sim-fidelity.ts
 var ABSENT_CATEGORY = "(absent)";
 var DEFAULT_MIN_N_PER_FEATURE = 20;
@@ -1733,6 +1603,136 @@ function easyModeCheck(simulated, production, opts = {}) {
   const gap = simPassRate - prodPassRate;
   return { simPassRate, prodPassRate, gap, inflated: gap > tolerance };
 }
+
+// src/rl/tournament.ts
+function fitBradleyTerry(outcomes, opts = {}) {
+  const tol = opts.tolerance ?? 1e-6;
+  const maxIter = opts.maxIterations ?? 256;
+  const smoothing = opts.smoothing ?? 0.1;
+  const candidates = /* @__PURE__ */ new Set();
+  for (const o of outcomes) {
+    candidates.add(o.winner);
+    candidates.add(o.loser);
+  }
+  const ids = [...candidates].sort();
+  const idx = new Map(ids.map((id, i) => [id, i]));
+  const n = ids.length;
+  if (n === 0) return { ratings: [], iterations: 0, finalDelta: 0, converged: true };
+  if (n === 1) {
+    return {
+      ratings: [{ candidateId: ids[0], strength: 1, logStrength: 0, n: 0, wins: 0 }],
+      iterations: 0,
+      finalDelta: 0,
+      converged: true
+    };
+  }
+  const W = Array.from({ length: n }, () => new Array(n).fill(0));
+  const N = Array.from({ length: n }, () => new Array(n).fill(0));
+  for (const o of outcomes) {
+    const i = idx.get(o.winner);
+    const j = idx.get(o.loser);
+    const w = o.weight ?? 1;
+    if (o.draw) {
+      W[i][j] += 0.5 * w;
+      W[j][i] += 0.5 * w;
+    } else {
+      W[i][j] += w;
+    }
+    N[i][j] += w;
+    N[j][i] += w;
+  }
+  const winsTotal = new Array(n).fill(0);
+  for (let i = 0; i < n; i++) {
+    for (let j = 0; j < n; j++) winsTotal[i] += W[i][j];
+    winsTotal[i] += smoothing;
+  }
+  const compsTotal = new Array(n).fill(0);
+  for (let i = 0; i < n; i++) {
+    for (let j = 0; j < n; j++) compsTotal[i] += N[i][j];
+  }
+  let theta = new Array(n).fill(1);
+  let iter = 0;
+  let delta = Infinity;
+  for (; iter < maxIter; iter++) {
+    const newTheta = new Array(n);
+    for (let i = 0; i < n; i++) {
+      let denom = 0;
+      for (let j = 0; j < n; j++) {
+        if (j === i) continue;
+        if (N[i][j] === 0) continue;
+        denom += N[i][j] / (theta[i] + theta[j]);
+      }
+      newTheta[i] = denom === 0 ? theta[i] : winsTotal[i] / denom;
+    }
+    let logSum = 0;
+    for (let i = 0; i < n; i++) logSum += Math.log(Math.max(1e-300, newTheta[i]));
+    const norm = Math.exp(logSum / n);
+    for (let i = 0; i < n; i++) newTheta[i] = newTheta[i] / norm;
+    delta = 0;
+    for (let i = 0; i < n; i++) {
+      const d = Math.abs(newTheta[i] - theta[i]) / Math.max(1e-12, theta[i]);
+      if (d > delta) delta = d;
+    }
+    theta = newTheta;
+    if (delta < tol) break;
+  }
+  const minLog = Math.min(...theta.map((t) => Math.log(Math.max(1e-300, t))));
+  const ratings = ids.map((id, i) => ({
+    candidateId: id,
+    strength: theta[i],
+    logStrength: Math.log(Math.max(1e-300, theta[i])) - minLog,
+    n: compsTotal[i],
+    wins: winsTotal[i] - smoothing
+  }));
+  return {
+    ratings: ratings.sort((a, b) => b.strength - a.strength),
+    iterations: iter,
+    finalDelta: delta,
+    converged: delta < tol
+  };
+}
+function applyEloUpdate(ratings, outcome, opts = {}) {
+  const defaultRating = opts.defaultRating ?? 1500;
+  const k = opts.kFactor ?? 32;
+  const rW = ratings.get(outcome.winner) ?? defaultRating;
+  const rL = ratings.get(outcome.loser) ?? defaultRating;
+  const expectedW = 1 / (1 + 10 ** ((rL - rW) / 400));
+  const scoreW = outcome.draw ? 0.5 : 1;
+  const scoreL = outcome.draw ? 0.5 : 0;
+  const w = outcome.weight ?? 1;
+  const winnerDelta = k * w * (scoreW - expectedW);
+  const loserDelta = k * w * (scoreL - (1 - expectedW));
+  ratings.set(outcome.winner, rW + winnerDelta);
+  ratings.set(outcome.loser, rL + loserDelta);
+  return { winnerDelta, loserDelta };
+}
+function buildPairwiseFromCampaign(input) {
+  const drawMargin = input.drawMargin ?? 0;
+  const byKey = /* @__PURE__ */ new Map();
+  for (const r of input.runs) {
+    const arr = byKey.get(r.matchKey) ?? [];
+    arr.push({ candidateId: r.candidateId, score: r.score });
+    byKey.set(r.matchKey, arr);
+  }
+  const outcomes = [];
+  for (const arr of byKey.values()) {
+    for (let i = 0; i < arr.length; i++) {
+      for (let j = i + 1; j < arr.length; j++) {
+        const a = arr[i];
+        const b = arr[j];
+        if (a.candidateId === b.candidateId) continue;
+        const margin = Math.abs(a.score - b.score);
+        if (margin <= drawMargin) {
+          outcomes.push({ winner: a.candidateId, loser: b.candidateId, draw: true, weight: 1 });
+        } else {
+          const [winner, loser] = a.score > b.score ? [a, b] : [b, a];
+          outcomes.push({ winner: winner.candidateId, loser: loser.candidateId, weight: margin });
+        }
+      }
+    }
+  }
+  return outcomes;
+}
 export {
   ABSENT_CATEGORY,
   DEFAULT_MIN_N_PER_FEATURE,