@tangle-network/agent-eval 0.120.1 → 0.120.2

package/dist/chunk-MOXWMGPC.js

@@ -1,577 +0,0 @@
-import {
-  summaryTable
-} from "./chunk-DPZAEKA6.js";
-import {
-  VerificationError
-} from "./chunk-ONWEPEDO.js";
-
-// src/release-confidence.ts
-var DEFAULT_THRESHOLDS = {
-  requireCorpus: true,
-  minScenarioCount: 1,
-  minSearchRuns: 1,
-  minHoldoutRuns: 1,
-  requireHoldout: true,
-  minPassRate: 0.8,
-  minMeanScore: 0.7,
-  maxOverfitGap: 0.15,
-  maxMeanCostUsd: Number.POSITIVE_INFINITY,
-  maxP95WallMs: Number.POSITIVE_INFINITY,
-  requireAsiForFailures: true,
-  failureScoreThreshold: 0.5
-};
-function evaluateReleaseConfidence(input) {
-  const thresholds = { ...DEFAULT_THRESHOLDS, ...input.thresholds };
-  const candidateId = input.candidateId ?? null;
-  const runs = filterCandidate(input.runs ?? [], candidateId, input.baselineId);
-  const traces = filterTraceCandidate(input.traces ?? [], candidateId, input.baselineId);
-  const scenarios = input.scenarios ?? [];
-  const scenarioCount = input.dataset?.scenarioCount ?? scenarios.length;
-  const splitCounts = input.dataset?.splitCounts ?? countScenarioSplits(scenarios);
-  const searchScores = scoresFor(runs, "search");
-  const holdoutScores = scoresFor(runs, "holdout");
-  const allScores = [...searchScores, ...holdoutScores];
-  const traceScores = traces.map((t) => t.score).filter(isFiniteNumber);
-  const scoreUniverse = allScores.length > 0 ? allScores : traceScores;
-  const searchRuns = runs.filter((r) => r.splitTag === "search").length;
-  const holdoutRuns = runs.filter((r) => r.splitTag === "holdout").length;
-  const searchMeanScore = mean(searchScores);
-  const holdoutMeanScore = mean(holdoutScores);
-  const metrics = {
-    scenarioCount,
-    searchRuns,
-    holdoutRuns,
-    passRate: passRate(runs, traces, thresholds.failureScoreThreshold),
-    meanScore: mean(scoreUniverse),
-    searchMeanScore,
-    holdoutMeanScore,
-    overfitGap: safeDiff(searchMeanScore, holdoutMeanScore),
-    meanCostUsd: mean([
-      ...runs.map((r) => r.costUsd),
-      ...traces.map((t) => t.costUsd).filter(isFiniteNumber)
-    ]),
-    p95WallMs: percentile(
-      [...runs.map((r) => r.wallMs), ...traces.map((t) => t.durationMs).filter(isFiniteNumber)],
-      0.95
-    ),
-    failedRows: failedRows(runs, traces, thresholds.failureScoreThreshold).length,
-    failuresWithAsi: failedRows(runs, traces, thresholds.failureScoreThreshold).filter(
-      (row) => row.hasAsi
-    ).length,
-    singleShotTraces: traces.filter((t) => t.turnCount === 1).length,
-    multiShotTraces: traces.filter((t) => (t.turnCount ?? 0) > 1).length,
-    splitCounts,
-    domainCounts: countDomains(scenarios),
-    failureModeCounts: countFailureModes(runs, traces, thresholds.failureScoreThreshold),
-    responsibleSurfaceCounts: countResponsibleSurfaces(traces)
-  };
-  const issues = [];
-  checkCorpus(input, thresholds, metrics, issues);
-  checkQuality(thresholds, metrics, issues);
-  checkGeneralization(input.gateDecision ?? null, thresholds, metrics, issues);
-  checkDiagnostics(thresholds, metrics, issues);
-  checkEfficiency(thresholds, metrics, issues);
-  const axes = buildAxes(metrics, thresholds, input.gateDecision ?? null, issues);
-  const status = issues.some((i) => i.severity === "critical") ? "fail" : issues.length > 0 ? "warn" : "pass";
-  return {
-    target: input.target,
-    candidateId,
-    baselineId: input.baselineId ?? null,
-    status,
-    promote: status === "pass" && (input.gateDecision ? input.gateDecision.promote : true),
-    axes,
-    issues,
-    metrics,
-    dataset: input.dataset ?? null,
-    gateDecision: input.gateDecision ?? null,
-    summary: renderSummary(input.target, status, metrics, issues)
-  };
-}
-function assertReleaseConfidence(input) {
-  const scorecard = evaluateReleaseConfidence(input);
-  if (scorecard.status === "fail") {
-    throw new VerificationError(scorecard.summary);
-  }
-  return scorecard;
-}
-function filterCandidate(runs, candidateId, baselineId) {
-  if (candidateId) return runs.filter((r) => r.candidateId === candidateId);
-  if (baselineId) return runs.filter((r) => r.candidateId !== baselineId);
-  return [...runs];
-}
-function filterTraceCandidate(traces, candidateId, baselineId) {
-  if (candidateId)
-    return traces.filter((t) => t.candidateId === void 0 || t.candidateId === candidateId);
-  if (baselineId)
-    return traces.filter((t) => t.candidateId === void 0 || t.candidateId !== baselineId);
-  return [...traces];
-}
-function checkCorpus(input, thresholds, metrics, issues) {
-  if (thresholds.requireCorpus && !input.dataset && (input.scenarios?.length ?? 0) === 0) {
-    issues.push({
-      axis: "corpus",
-      severity: "critical",
-      code: "missing_corpus",
-      detail: "No Dataset manifest or scenarios supplied."
-    });
-  }
-  if (metrics.scenarioCount < thresholds.minScenarioCount) {
-    issues.push({
-      axis: "corpus",
-      severity: "critical",
-      code: "few_scenarios",
-      detail: `${metrics.scenarioCount} scenario(s) < min ${thresholds.minScenarioCount}.`
-    });
-  }
-  if (thresholds.requireHoldout && metrics.splitCounts.holdout === 0) {
-    issues.push({
-      axis: "corpus",
-      severity: "critical",
-      code: "missing_holdout_split",
-      detail: "Corpus has no holdout scenarios."
-    });
-  }
-}
-function checkQuality(thresholds, metrics, issues) {
-  if (metrics.searchRuns < thresholds.minSearchRuns) {
-    issues.push({
-      axis: "quality",
-      severity: "critical",
-      code: "few_search_runs",
-      detail: `${metrics.searchRuns} search run(s) < min ${thresholds.minSearchRuns}.`
-    });
-  }
-  if (metrics.passRate < thresholds.minPassRate) {
-    issues.push({
-      axis: "quality",
-      severity: "critical",
-      code: "low_pass_rate",
-      detail: `passRate ${fmt(metrics.passRate)} < ${fmt(thresholds.minPassRate)}.`
-    });
-  }
-  if (metrics.meanScore < thresholds.minMeanScore) {
-    issues.push({
-      axis: "quality",
-      severity: "critical",
-      code: "low_mean_score",
-      detail: `meanScore ${fmt(metrics.meanScore)} < ${fmt(thresholds.minMeanScore)}.`
-    });
-  }
-}
-function checkGeneralization(gateDecision, thresholds, metrics, issues) {
-  if (thresholds.requireHoldout && metrics.holdoutRuns < thresholds.minHoldoutRuns) {
-    issues.push({
-      axis: "generalization",
-      severity: "critical",
-      code: "few_holdout_runs",
-      detail: `${metrics.holdoutRuns} holdout run(s) < min ${thresholds.minHoldoutRuns}.`
-    });
-  }
-  if (Number.isFinite(metrics.overfitGap) && metrics.overfitGap > thresholds.maxOverfitGap) {
-    issues.push({
-      axis: "generalization",
-      severity: "critical",
-      code: "overfit_gap",
-      detail: `search-holdout gap ${fmt(metrics.overfitGap)} > ${fmt(thresholds.maxOverfitGap)}.`
-    });
-  }
-  if (gateDecision && !gateDecision.promote) {
-    issues.push({
-      axis: "generalization",
-      severity: "critical",
-      code: `gate_${gateDecision.rejectionCode ?? "reject"}`,
-      detail: gateDecision.reason
-    });
-  }
-}
-function checkDiagnostics(thresholds, metrics, issues) {
-  if (!thresholds.requireAsiForFailures) return;
-  if (metrics.failedRows > metrics.failuresWithAsi) {
-    issues.push({
-      axis: "diagnostics",
-      severity: "critical",
-      code: "missing_failure_asi",
-      detail: `${metrics.failedRows - metrics.failuresWithAsi} failed row(s) have no actionable side information.`
-    });
-  }
-}
-function checkEfficiency(thresholds, metrics, issues) {
-  if (metrics.meanCostUsd > thresholds.maxMeanCostUsd) {
-    issues.push({
-      axis: "efficiency",
-      severity: "critical",
-      code: "cost_budget",
-      detail: `meanCostUsd ${fmt(metrics.meanCostUsd)} > ${fmt(thresholds.maxMeanCostUsd)}.`
-    });
-  }
-  if (metrics.p95WallMs > thresholds.maxP95WallMs) {
-    issues.push({
-      axis: "efficiency",
-      severity: "critical",
-      code: "latency_budget",
-      detail: `p95WallMs ${fmt(metrics.p95WallMs)} > ${fmt(thresholds.maxP95WallMs)}.`
-    });
-  }
-}
-function buildAxes(metrics, thresholds, gateDecision, issues) {
-  return [
-    axis(
-      "corpus",
-      issues,
-      bounded(metrics.scenarioCount / Math.max(1, thresholds.minScenarioCount)),
-      `${metrics.scenarioCount} scenarios; holdout=${metrics.splitCounts.holdout}`
-    ),
-    axis(
-      "quality",
-      issues,
-      Math.min(metrics.passRate, metrics.meanScore),
-      `passRate=${fmt(metrics.passRate)} meanScore=${fmt(metrics.meanScore)}`
-    ),
-    axis(
-      "generalization",
-      issues,
-      gateDecision && !gateDecision.promote ? 0 : gapScore(metrics.overfitGap, thresholds.maxOverfitGap),
-      `holdoutRuns=${metrics.holdoutRuns} overfitGap=${fmt(metrics.overfitGap)}`
-    ),
-    axis(
-      "diagnostics",
-      issues,
-      metrics.failedRows === 0 ? 1 : metrics.failuresWithAsi / metrics.failedRows,
-      `failuresWithAsi=${metrics.failuresWithAsi}/${metrics.failedRows}`
-    ),
-    axis(
-      "efficiency",
-      issues,
-      efficiencyScore(metrics, thresholds),
-      `meanCostUsd=${fmt(metrics.meanCostUsd)} p95WallMs=${fmt(metrics.p95WallMs)}`
-    )
-  ];
-}
-function axis(name, issues, score, detail) {
-  const own = issues.filter((i) => i.axis === name);
-  const status = own.some((i) => i.severity === "critical") ? "fail" : own.length > 0 ? "warn" : "pass";
-  return { name, status, score: bounded(score), detail };
-}
-function countScenarioSplits(scenarios) {
-  const counts = { train: 0, dev: 0, test: 0, holdout: 0 };
-  for (const scenario of scenarios) counts[scenario.split ?? "train"]++;
-  return counts;
-}
-function countDomains(scenarios) {
-  const out = {};
-  for (const scenario of scenarios) {
-    const domain = scenario.tags?.domain ?? scenario.tags?.category ?? "uncategorized";
-    out[domain] = (out[domain] ?? 0) + 1;
-  }
-  return out;
-}
-function countFailureModes(runs, traces, threshold) {
-  const out = {};
-  for (const run of runs) {
-    const score = run.outcome.holdoutScore ?? run.outcome.searchScore;
-    if (run.failureMode || score !== void 0 && score < threshold) {
-      const mode = run.failureMode ?? "low_score";
-      out[mode] = (out[mode] ?? 0) + 1;
-    }
-  }
-  for (const trace of traces) {
-    if (trace.failureMode || trace.ok === false || trace.score !== void 0 && trace.score < threshold) {
-      const mode = trace.failureMode ?? (trace.ok === false ? "not_ok" : "low_score");
-      out[mode] = (out[mode] ?? 0) + 1;
-    }
-  }
-  return out;
-}
-function countResponsibleSurfaces(traces) {
-  const out = {};
-  for (const trace of traces) {
-    for (const asi of trace.asi ?? []) {
-      const surface = asi.responsibleSurface ?? "unknown";
-      out[surface] = (out[surface] ?? 0) + 1;
-    }
-  }
-  return out;
-}
-function failedRows(runs, traces, threshold) {
-  const out = [];
-  for (const run of runs) {
-    const score = run.outcome.holdoutScore ?? run.outcome.searchScore;
-    if (run.failureMode || score !== void 0 && score < threshold) {
-      const asiMetric = run.outcome.raw.asi;
-      out.push({ hasAsi: typeof asiMetric === "number" && asiMetric > 0 });
-    }
-  }
-  for (const trace of traces) {
-    if (trace.failureMode || trace.ok === false || trace.score !== void 0 && trace.score < threshold) {
-      out.push({ hasAsi: (trace.asi?.length ?? 0) > 0 });
-    }
-  }
-  return out;
-}
-function passRate(runs, traces, threshold) {
-  const outcomes = [
-    ...runs.map((run) => {
-      const score = run.outcome.holdoutScore ?? run.outcome.searchScore;
-      return !run.failureMode && score !== void 0 && score >= threshold;
-    }),
-    ...traces.map(
-      (trace) => trace.ok !== false && (trace.score === void 0 || trace.score >= threshold)
-    )
-  ];
-  if (outcomes.length === 0) return 0;
-  return outcomes.filter(Boolean).length / outcomes.length;
-}
-function scoresFor(runs, split) {
-  return runs.filter((run) => run.splitTag === split).map((run) => split === "holdout" ? run.outcome.holdoutScore : run.outcome.searchScore).filter(isFiniteNumber);
-}
-function mean(xs) {
-  if (xs.length === 0) return Number.NaN;
-  return xs.reduce((sum, x) => sum + x, 0) / xs.length;
-}
-function percentile(xs, p) {
-  if (xs.length === 0) return Number.NaN;
-  const sorted = [...xs].sort((a, b) => a - b);
-  return sorted[Math.min(sorted.length - 1, Math.max(0, Math.ceil(p * sorted.length) - 1))];
-}
-function isFiniteNumber(value) {
-  return typeof value === "number" && Number.isFinite(value);
-}
-function safeDiff(a, b) {
-  if (!Number.isFinite(a) || !Number.isFinite(b)) return Number.NaN;
-  return a - b;
-}
-function gapScore(gap, maxGap) {
-  if (!Number.isFinite(gap)) return 0;
-  if (maxGap <= 0) return gap <= 0 ? 1 : 0;
-  return bounded(1 - Math.max(0, gap) / maxGap);
-}
-function efficiencyScore(metrics, thresholds) {
-  const cost = Number.isFinite(thresholds.maxMeanCostUsd) && Number.isFinite(metrics.meanCostUsd) ? bounded(thresholds.maxMeanCostUsd / Math.max(metrics.meanCostUsd, 1e-12)) : 1;
-  const latency = Number.isFinite(thresholds.maxP95WallMs) && Number.isFinite(metrics.p95WallMs) ? bounded(thresholds.maxP95WallMs / Math.max(metrics.p95WallMs, 1e-12)) : 1;
-  return Math.min(cost, latency);
-}
-function bounded(x) {
-  if (!Number.isFinite(x)) return 0;
-  return Math.max(0, Math.min(1, x));
-}
-function renderSummary(target, status, metrics, issues) {
-  const prefix = `release confidence ${status}: ${target}`;
-  const metricText = `scenarios=${metrics.scenarioCount} searchRuns=${metrics.searchRuns} holdoutRuns=${metrics.holdoutRuns} passRate=${fmt(metrics.passRate)} meanScore=${fmt(metrics.meanScore)}`;
-  if (issues.length === 0) return `${prefix}; ${metricText}`;
-  return `${prefix}; ${metricText}; issues=${issues.map((i) => i.code).join(",")}`;
-}
-function fmt(x) {
-  if (!Number.isFinite(x)) return String(x);
-  return x.toFixed(4);
-}
-
-// src/promotion-gate.ts
-function bootstrapCi(baseline, candidate, options = {}) {
-  const alpha = options.alpha ?? 0.05;
-  const iterations = options.iterations ?? 1e3;
-  const minTotal = options.minTotalSamples ?? 6;
-  const rng = mulberry32(options.seed ?? hashSeed(baseline, candidate));
-  const baselineMean = mean2(baseline);
-  const candidateMean = mean2(candidate);
-  const delta = candidateMean - baselineMean;
-  if (baseline.length + candidate.length < minTotal || baseline.length === 0 || candidate.length === 0) {
-    return {
-      baselineMean,
-      candidateMean,
-      delta,
-      ciLower: -Infinity,
-      ciUpper: Infinity,
-      iterations: 0,
-      alpha,
-      verdict: "INCONCLUSIVE"
-    };
-  }
-  const deltas = new Array(iterations);
-  for (let i = 0; i < iterations; i++) {
-    const bResample = resample(baseline, rng);
-    const cResample = resample(candidate, rng);
-    deltas[i] = mean2(cResample) - mean2(bResample);
-  }
-  deltas.sort((a, b) => a - b);
-  const lowerIdx = Math.floor(alpha / 2 * iterations);
-  const upperIdx = Math.floor((1 - alpha / 2) * iterations) - 1;
-  const ciLower = deltas[Math.max(0, lowerIdx)];
-  const ciUpper = deltas[Math.min(iterations - 1, upperIdx)];
-  let verdict;
-  if (ciLower > 0) verdict = "ADVANCE";
-  else if (ciUpper < 0) verdict = "REVERT";
-  else if (delta >= 0) verdict = "KEEP";
-  else verdict = "INCONCLUSIVE";
-  return {
-    baselineMean,
-    candidateMean,
-    delta,
-    ciLower,
-    ciUpper,
-    iterations,
-    alpha,
-    verdict
-  };
-}
-function mean2(xs) {
-  if (xs.length === 0) return 0;
-  let s = 0;
-  for (const x of xs) s += x;
-  return s / xs.length;
-}
-function resample(xs, rng) {
-  const out = new Array(xs.length);
-  for (let i = 0; i < xs.length; i++) out[i] = xs[Math.floor(rng() * xs.length)];
-  return out;
-}
-function mulberry32(seed) {
-  let t = seed >>> 0;
-  return () => {
-    t += 1831565813;
-    let r = t;
-    r = Math.imul(r ^ r >>> 15, r | 1);
-    r ^= r + Math.imul(r ^ r >>> 7, r | 61);
-    return ((r ^ r >>> 14) >>> 0) / 4294967296;
-  };
-}
-function hashSeed(a, b) {
-  let h = 2166136261;
-  for (const x of [...a, ...b]) {
-    const view = new Float64Array([x]);
-    const bytes = new Uint8Array(view.buffer);
-    for (const byte of bytes) {
-      h ^= byte;
-      h = Math.imul(h, 16777619);
-    }
-  }
-  return h >>> 0;
-}
-async function judgeReplayGate(args) {
-  const concurrency = args.judgeConcurrency ?? 4;
-  const baselineScores = await scoreAll(args.baselineOutputs, args.judge, concurrency);
-  const candidateScores = await scoreAll(args.candidateOutputs, args.judge, concurrency);
-  const ci = bootstrapCi(baselineScores, candidateScores, {
-    ...args.alpha !== void 0 ? { alpha: args.alpha } : {},
-    ...args.iterations !== void 0 ? { iterations: args.iterations } : {},
-    ...args.seed !== void 0 ? { seed: args.seed } : {}
-  });
-  return {
-    ...ci,
-    baselineSamples: baselineScores.length,
-    candidateSamples: candidateScores.length
-  };
-}
-async function scoreAll(outputs, judge, concurrency) {
-  const results = new Array(outputs.length);
-  let next = 0;
-  async function worker() {
-    while (true) {
-      const i = next++;
-      if (i >= outputs.length) return;
-      const v = await judge(outputs[i]);
-      results[i] = Number.isFinite(v) ? v : 0;
-    }
-  }
-  await Promise.all(Array.from({ length: Math.max(1, concurrency) }, () => worker()));
-  return results;
-}
-
-// src/release-report.ts
-function renderReleaseReport(scorecard, options = {}) {
-  const title = options.title ?? `Release Report: ${scorecard.target}`;
-  const lines = [];
-  lines.push(`# ${title}`);
-  lines.push("");
-  lines.push(`Status: **${scorecard.status.toUpperCase()}**`);
-  lines.push(`Promote: **${scorecard.promote ? "yes" : "no"}**`);
-  if (scorecard.candidateId) lines.push(`Candidate: \`${scorecard.candidateId}\``);
-  if (scorecard.baselineId) lines.push(`Baseline: \`${scorecard.baselineId}\``);
-  lines.push("");
-  lines.push(scorecard.summary);
-  lines.push("");
-  lines.push("## Metrics");
-  lines.push("");
-  lines.push("| Metric | Value |");
-  lines.push("|---|---:|");
-  lines.push(`| Scenarios | ${scorecard.metrics.scenarioCount} |`);
-  lines.push(`| Search runs | ${scorecard.metrics.searchRuns} |`);
-  lines.push(`| Holdout runs | ${scorecard.metrics.holdoutRuns} |`);
-  lines.push(`| Pass rate | ${pct(scorecard.metrics.passRate)} |`);
-  lines.push(`| Mean score | ${num(scorecard.metrics.meanScore)} |`);
-  lines.push(`| Search mean | ${num(scorecard.metrics.searchMeanScore)} |`);
-  lines.push(`| Holdout mean | ${num(scorecard.metrics.holdoutMeanScore)} |`);
-  lines.push(`| Overfit gap | ${num(scorecard.metrics.overfitGap)} |`);
-  lines.push(`| Mean cost | $${num(scorecard.metrics.meanCostUsd)} |`);
-  lines.push(`| p95 wall time | ${Math.round(scorecard.metrics.p95WallMs)} ms |`);
-  lines.push("");
-  if (scorecard.issues.length > 0) {
-    lines.push("## Issues");
-    lines.push("");
-    for (const issue of scorecard.issues) {
-      lines.push(`- **${issue.severity}** \`${issue.code}\` (${issue.axis}): ${issue.detail}`);
-    }
-    lines.push("");
-  }
-  const surfaces = entries(scorecard.metrics.responsibleSurfaceCounts);
-  if (surfaces.length > 0) {
-    lines.push("## Responsible Surfaces");
-    lines.push("");
-    for (const [surface, count] of surfaces) lines.push(`- ${surface}: ${count}`);
-    lines.push("");
-  }
-  const failures = entries(scorecard.metrics.failureModeCounts);
-  if (failures.length > 0) {
-    lines.push("## Failure Modes");
-    lines.push("");
-    for (const [mode, count] of failures) lines.push(`- ${mode}: ${count}`);
-    lines.push("");
-  }
-  if (options.runs && options.runs.length > 0) {
-    lines.push("## Run Summary");
-    lines.push("");
-    lines.push(
-      summaryTable([...options.runs], {
-        comparator: options.comparator ?? scorecard.baselineId ?? void 0,
-        split: "holdout"
-      }).markdown
-    );
-    lines.push("");
-  }
-  if (options.traceAnalystFindings && options.traceAnalystFindings.length > 0) {
-    lines.push("## TraceAnalyst Findings");
-    lines.push("");
-    for (const finding of options.traceAnalystFindings) lines.push(`- ${finding}`);
-    lines.push("");
-  }
-  const nextActions = options.nextActions ?? defaultNextActions(scorecard);
-  if (nextActions.length > 0) {
-    lines.push("## Next Actions");
-    lines.push("");
-    for (const action of nextActions) lines.push(`- ${action}`);
-    lines.push("");
-  }
-  return `${lines.join("\n").trimEnd()}
-`;
-}
-function defaultNextActions(scorecard) {
-  if (scorecard.promote) return ["Promote the candidate and keep canaries enabled."];
-  return scorecard.issues.filter((issue) => issue.severity === "critical").map((issue) => `Resolve ${issue.code}: ${issue.detail}`);
-}
-function entries(values) {
-  return Object.entries(values).filter(([, count]) => count > 0).sort((a, b) => b[1] - a[1] || a[0].localeCompare(b[0]));
-}
-function pct(value) {
-  return Number.isFinite(value) ? `${(value * 100).toFixed(1)}%` : "n/a";
-}
-function num(value) {
-  return Number.isFinite(value) ? value.toFixed(3) : "n/a";
-}
-
-export {
-  evaluateReleaseConfidence,
-  assertReleaseConfidence,
-  bootstrapCi,
-  judgeReplayGate,
-  renderReleaseReport
-};
-//# sourceMappingURL=chunk-MOXWMGPC.js.map