@tangle-network/agent-eval 0.126.7 → 0.127.0

package/dist/chunk-EJGRPCO3.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/rollout/exporters.ts","../src/rollout/ledger.ts","../src/rollout/release/card.ts","../src/rollout/release/scrub.ts","../src/rollout/release/hf-dataset.ts"],"sourcesContent":["/**\n * Pure exporters over `tangle.rollout.v1` lines → the training-data shapes\n * the improvement loops feed:\n *   - SFT chat JSONL           (clean trainable successes, {messages, metadata})\n *   - reward rows              (every scored line, success or failure, with steps)\n *   - Prime Intellect verifiers RolloutOutput (prompt/completion split + reward)\n *   - OpenAI RFT items         (prompt turns + verdict reference fields)\n *\n * All exporters are pure functions of the lines — filtering (never train on\n * holdout, reward thresholds, the realness gate) happens HERE, on inline\n * labels, no joins.\n */\n\nimport type { ChatMessage, RolloutLine, RolloutSplit, RolloutStep, ToolDef } from './schema'\n\n// ---------------------------------------------------------------------------\n// (a) SFT chat JSONL\n// ---------------------------------------------------------------------------\n\nexport interface TrainingExportOptions {\n  /** Include held-out evaluation data in training output. Default false. */\n  allowHeldOutTrainingData?: boolean\n  /** Require reward to be strictly greater than this value. Default 0. */\n  minimumQualityExclusive?: number\n}\n\nexport type SftExportOptions = TrainingExportOptions\n\nexport interface SftRow {\n  messages: ChatMessage[]\n  metadata: {\n    rollout_id: string\n    run_id: string\n    candidate_id: string | null\n    instance_id: string\n    reward: number\n  }\n}\n\n/**\n * Supervised fine-tune rows: the completed conversation of each qualifying\n * line. Fail-closed filters: trainable split only (never holdout/canary),\n * positive reward, realness-gated lines never qualify, gap lines carry\n * no trainable content.\n */\nexport function toSftRows(lines: RolloutLine[], options: SftExportOptions = {}): SftRow[] {\n  return lines\n    .filter((line) => isTrainingLineEligible(line, options) && line.messages.length > 0)\n    .map((line) => ({\n      messages: line.messages,\n      metadata: {\n        rollout_id: line.rollout_id,\n        run_id: line.run_id,\n        candidate_id: line.candidate_id ?? null,\n        instance_id: line.task.instance_id,\n        reward: line.outcome.reward as number,\n      },\n    }))\n}\n\n// ---------------------------------------------------------------------------\n// (b) Reward rows\n// ---------------------------------------------------------------------------\n\nexport interface RewardRow {\n  /** First user turn — the task prompt. */\n  prompt: string\n  steps: RolloutStep[]\n  reward: number\n  metadata: {\n    rollout_id: string\n    run_id: string\n    candidate_id: string | null\n    instance_id: string\n    split: RolloutSplit\n  }\n}\n\n/**\n * Reward-labeled rows for completed, positive-quality training runs.\n */\nexport function toRewardRows(\n  lines: RolloutLine[],\n  options: TrainingExportOptions = {},\n): RewardRow[] {\n  return lines\n    .filter(\n      (line) =>\n        isTrainingLineEligible(line, options) &&\n        line.messages.some(\n          (message) =>\n            message.role === 'user' &&\n            typeof message.content === 'string' &&\n            message.content.length > 0,\n        ),\n    )\n    .map((line) => ({\n      prompt: line.messages.find((m) => m.role === 'user')?.content ?? '',\n      steps: line.steps ?? [],\n      reward: line.outcome.reward as number,\n      metadata: {\n        rollout_id: line.rollout_id,\n        run_id: line.run_id,\n        candidate_id: line.candidate_id ?? null,\n        instance_id: line.task.instance_id,\n        split: line.task.split,\n      },\n    }))\n}\n\n// ---------------------------------------------------------------------------\n// (c) Prime Intellect verifiers RolloutOutput\n// ---------------------------------------------------------------------------\n\nexport interface VerifiersTokenUsage {\n  input_tokens: number | null\n  output_tokens: number | null\n  reasoning_tokens: number | null\n  cache_read_tokens: number | null\n  cache_write_tokens: number | null\n}\n\nexport interface VerifiersRolloutOutput {\n  /** Messages through the last turn BEFORE the first assistant turn. */\n  prompt: ChatMessage[]\n  /** The first assistant turn onward — what the policy produced. */\n  completion: ChatMessage[]\n  reward: number | null\n  metrics: Record<string, unknown>\n  tool_defs: ToolDef[]\n  token_usage: VerifiersTokenUsage\n  info: {\n    task: RolloutLine['task']\n    policy: RolloutLine['policy']\n    rollout_id: string\n    run_id: string\n    experiment_id: string | null\n    candidate_id: string | null\n    generation: number | null\n    candidate_index: number | null\n    role: RolloutLine['role']\n  }\n}\n\n/** Index of the first assistant turn; messages.length when none exists. */\nfunction firstAssistantIndex(messages: ChatMessage[]): number {\n  const index = messages.findIndex((m) => m.role === 'assistant')\n  return index === -1 ? messages.length : index\n}\n\nexport function toVerifiersRolloutOutput(line: RolloutLine): VerifiersRolloutOutput {\n  const split = firstAssistantIndex(line.messages)\n  return {\n    prompt: line.messages.slice(0, split),\n    completion: line.messages.slice(split),\n    reward: line.outcome.reward,\n    metrics: line.outcome.metrics,\n    tool_defs: line.tool_defs,\n    token_usage: {\n      input_tokens: line.cost.tokens_in,\n      output_tokens: line.cost.tokens_out,\n      reasoning_tokens: line.cost.tokens_reasoning,\n      cache_read_tokens: line.cost.cache_read,\n      cache_write_tokens: line.cost.cache_write,\n    },\n    info: {\n      task: line.task,\n      policy: line.policy,\n      rollout_id: line.rollout_id,\n      run_id: line.run_id,\n      experiment_id: line.experiment_id ?? null,\n      candidate_id: line.candidate_id ?? null,\n      generation: line.generation,\n      candidate_index: line.candidate_index,\n      role: line.role,\n    },\n  }\n}\n\nexport function toVerifiersRolloutOutputs(\n  lines: RolloutLine[],\n  options: TrainingExportOptions = {},\n): VerifiersRolloutOutput[] {\n  return lines\n    .filter(\n      (line) =>\n        isTrainingLineEligible(line, options) &&\n        firstAssistantIndex(line.messages) > 0 &&\n        firstAssistantIndex(line.messages) < line.messages.length,\n    )\n    .map(toVerifiersRolloutOutput)\n}\n\n// ---------------------------------------------------------------------------\n// (d) OpenAI RFT items\n// ---------------------------------------------------------------------------\n\nexport interface RftItem {\n  /** Prompt turns only — the graded completion is re-sampled during RFT. */\n  messages: ChatMessage[]\n  /** Verdict/label fields the grader references as item.reference.* */\n  reference: {\n    reward: number | null\n    reward_source: string | null\n    verdict: unknown\n    instance_id: string\n    suite: string\n    split: RolloutSplit\n    rollout_id: string\n  }\n}\n\nexport function toRftItem(line: RolloutLine): RftItem {\n  const split = firstAssistantIndex(line.messages)\n  return {\n    messages: line.messages.slice(0, split),\n    reference: {\n      reward: line.outcome.reward,\n      reward_source: line.outcome.reward_source,\n      verdict: line.outcome.verdict,\n      instance_id: line.task.instance_id,\n      suite: line.task.suite,\n      split: line.task.split,\n      rollout_id: line.rollout_id,\n    },\n  }\n}\n\n/** RFT needs a real prompt: lines whose transcript starts with prompt turns. */\nexport function toRftItems(lines: RolloutLine[], options: TrainingExportOptions = {}): RftItem[] {\n  return lines\n    .filter(\n      (line) =>\n        isTrainingLineEligible(line, options) &&\n        line.messages.length > 0 &&\n        firstAssistantIndex(line.messages) > 0,\n    )\n    .map(toRftItem)\n}\n\n// ---------------------------------------------------------------------------\n// Serialization — one JSON object per line, the interchange format for\n// every export. `tangle.rollout.v1` lines and export rows alike.\n// ---------------------------------------------------------------------------\n\nexport function toJsonl(rows: ReadonlyArray<unknown>): string {\n  return rows.map((r) => JSON.stringify(r)).join('\\n') + (rows.length ? '\\n' : '')\n}\n\nfunction isTrainingLineEligible(\n  line: RolloutLine,\n  options: TrainingExportOptions,\n): line is RolloutLine & { outcome: RolloutLine['outcome'] & { reward: number } } {\n  const minimumQualityExclusive = options.minimumQualityExclusive ?? 0\n  if (!Number.isFinite(minimumQualityExclusive)) {\n    throw new Error('minimumQualityExclusive must be finite')\n  }\n\n  const reward = line.outcome.reward\n  if (reward === null) return false\n  if (!Number.isFinite(reward)) {\n    throw new Error(`training reward for rollout \"${line.rollout_id}\" must be finite`)\n  }\n  if (reward <= minimumQualityExclusive) return false\n  if (!line.outcome.is_completed || line.outcome.is_truncated || line.outcome.error !== null) {\n    return false\n  }\n  if (line.outcome.realness_gated === true) return false\n  if (line.task.split === 'search') return true\n  return line.task.split === 'holdout' && options.allowHeldOutTrainingData === true\n}\n","/**\n * Rollout-ledger file API — append-only JSONL of validated `tangle.rollout.v1`\n * lines. Writes validate BEFORE touching disk (a bad line never lands);\n * reads validate line-by-line and fail loud with the line number, because a\n * silently-skipped rollout is a corrupted dataset.\n */\n\nimport { appendFile, mkdir, readFile, writeFile } from 'node:fs/promises'\nimport { dirname } from 'node:path'\nimport { assertRolloutLine, type RolloutLine } from './schema'\n\nfunction serialize(lines: RolloutLine[]): string {\n  for (const [i, line] of lines.entries()) assertRolloutLine(line, `rollout line [${i}]`)\n  return lines.map((line) => JSON.stringify(line)).join('\\n') + (lines.length > 0 ? '\\n' : '')\n}\n\n/** Replace the ledger file with exactly `lines`. */\nexport async function writeRolloutLedger(path: string, lines: RolloutLine[]): Promise<void> {\n  const payload = serialize(lines)\n  await mkdir(dirname(path), { recursive: true })\n  await writeFile(path, payload)\n}\n\n/** Append `lines` to the ledger file (created if absent). */\nexport async function appendRolloutLines(path: string, lines: RolloutLine[]): Promise<void> {\n  if (lines.length === 0) return\n  const payload = serialize(lines)\n  await mkdir(dirname(path), { recursive: true })\n  await appendFile(path, payload)\n}\n\n/**\n * Read and validate every line. Throws on the first malformed/invalid line\n * (with its 1-based line number) — fail-closed, never a silent drop.\n */\nexport async function readRolloutLedger(path: string): Promise<RolloutLine[]> {\n  const raw = await readFile(path, 'utf8')\n  const lines: RolloutLine[] = []\n  const rawLines = raw.split('\\n')\n  for (let i = 0; i < rawLines.length; i++) {\n    const text = rawLines[i]\n    if (!text?.trim()) continue\n    let parsed: unknown\n    try {\n      parsed = JSON.parse(text)\n    } catch (error) {\n      throw new Error(\n        `${path}:${i + 1}: malformed JSON — ${error instanceof Error ? error.message : String(error)}`,\n      )\n    }\n    assertRolloutLine(parsed, `${path}:${i + 1}`)\n    lines.push(parsed)\n  }\n  return lines\n}\n","/**\n * HuggingFace dataset-card (README.md) generation for a rollout-ledger release.\n *\n * The card is a pure function of the SCRUBBED lines plus the release options —\n * no timestamps, no environment reads — so rebuilding from the same ledger\n * yields byte-identical output. It documents the schema, provenance (run ids,\n * generations, the official judge), per-role reward semantics including the\n * inherited/contribution caveat, and a role × reward counts table.\n */\n\nimport { ROLLOUT_ROLES, ROLLOUT_SCHEMA, type RolloutLine, type RolloutRole } from '../schema'\nimport type { ScrubCounts } from './scrub'\n\nexport const RELEASE_FORMATS = ['sft', 'verifiers', 'rft', 'raw'] as const\nexport type ReleaseFormat = (typeof RELEASE_FORMATS)[number]\n\n/** Format → data file path inside the dataset dir (train split only). */\nexport const FORMAT_FILES: Record<ReleaseFormat, string> = {\n  sft: 'sft/train.jsonl',\n  verifiers: 'verifiers/train.jsonl',\n  rft: 'rft/train.jsonl',\n  raw: 'raw/train.jsonl',\n}\n\nconst FORMAT_DESCRIPTIONS: Record<ReleaseFormat, string> = {\n  sft: 'Successful trainable-split transcripts (`reward >= 1`, never realness-gated) as `{messages, metadata}` chat JSONL.',\n  verifiers:\n    'Prime Intellect verifiers `RolloutOutput`: prompt/completion split at the first assistant turn, plus reward, metrics, tool defs, and token usage.',\n  rft: 'OpenAI RFT items: prompt turns plus `reference.*` verdict fields for a grader (completions are re-sampled during RFT).',\n  raw: `Full \\`${ROLLOUT_SCHEMA}\\` ledger lines (scrubbed), one per agent invocation.`,\n}\n\nexport interface DatasetCardInputs {\n  /** Scrubbed, release-filtered lines (what actually ships). */\n  lines: RolloutLine[]\n  formats: ReleaseFormat[]\n  includeProposers: boolean\n  /** Source ledger basenames, for provenance. */\n  sourceFiles: string[]\n  scrubTotals: ScrubCounts\n  excluded: { proposers: number; nonTrain: number }\n  formatCounts: Partial<Record<ReleaseFormat, number>>\n}\n\nfunction unique(values: Array<string | null>): string[] {\n  return [...new Set(values.filter((v): v is string => v !== null))].sort()\n}\n\nfunction formatReward(reward: number | null): string {\n  if (reward === null) return 'null'\n  return Number.isInteger(reward) ? String(reward) : reward.toFixed(4)\n}\n\nfunction markdownTable(header: string[], rows: string[][]): string {\n  return [\n    `| ${header.join(' | ')} |`,\n    `| ${header.map(() => '---').join(' | ')} |`,\n    ...rows.map((row) => `| ${row.join(' | ')} |`),\n  ].join('\\n')\n}\n\nfunction roleRewardRows(lines: RolloutLine[]): string[][] {\n  const counts = new Map<RolloutRole, Map<string, number>>()\n  for (const line of lines) {\n    const byReward = counts.get(line.role) ?? new Map<string, number>()\n    const key = formatReward(line.outcome.reward)\n    byReward.set(key, (byReward.get(key) ?? 0) + 1)\n    counts.set(line.role, byReward)\n  }\n  const rows: string[][] = []\n  for (const role of ROLLOUT_ROLES) {\n    const byReward = counts.get(role)\n    if (!byReward) continue\n    const keys = [...byReward.keys()].sort((a, b) => {\n      if (a === 'null') return 1\n      if (b === 'null') return -1\n      return Number(b) - Number(a)\n    })\n    for (const key of keys) rows.push([role, key, String(byReward.get(key))])\n  }\n  return rows\n}\n\nexport function buildDatasetCard(inputs: DatasetCardInputs): string {\n  const { lines, formats, includeProposers, sourceFiles, scrubTotals, excluded, formatCounts } =\n    inputs\n\n  const runIds = unique(lines.map((line) => line.run_id))\n  const generations = [...new Set(lines.map((line) => line.generation))]\n    .filter((g): g is number => g !== null)\n    .sort((a, b) => a - b)\n  const models = unique(lines.map((line) => line.policy.model))\n  const harnesses = unique(lines.map((line) => line.policy.harness))\n  const captures = unique(lines.map((line) => line.provenance.capture))\n  const rewardSources = unique(lines.map((line) => line.outcome.reward_source))\n  const gapLines = lines.filter((line) => line.messages.length === 0).length\n\n  const configs = formats\n    .map((format) =>\n      [\n        `  - config_name: ${format}`,\n        '    data_files:',\n        '      - split: train',\n        `        path: ${FORMAT_FILES[format]}`,\n      ].join('\\n'),\n    )\n    .join('\\n')\n\n  const frontmatter = [\n    '---',\n    'license: unknown',\n    'pretty_name: Tangle rollout ledger — agent trajectories',\n    'configs:',\n    configs,\n    '---',\n  ].join('\\n')\n\n  const formatsTable = markdownTable(\n    ['config', 'path', 'rows', 'contents'],\n    formats.map((format) => [\n      format,\n      `\\`${FORMAT_FILES[format]}\\``,\n      String(formatCounts[format] ?? 0),\n      FORMAT_DESCRIPTIONS[format],\n    ]),\n  )\n\n  const countsTable = markdownTable(['role', 'reward', 'lines'], roleRewardRows(lines))\n\n  const scrubTable = markdownTable(\n    ['rule', 'rewrites'],\n    Object.entries(scrubTotals).map(([rule, count]) => [rule, String(count)]),\n  )\n\n  const proposerNote = includeProposers\n    ? 'Proposer sessions are INCLUDED (`--include-proposers`); their transcripts contain improvement-loop harness source.'\n    : `Proposer sessions are excluded by default (${excluded.proposers} lines dropped); they contain improvement-loop harness source. Rebuild with \\`--include-proposers\\` to keep them.`\n\n  return `${frontmatter}\n\n# Tangle rollout ledger — agent trajectories\n\nOne line per agent invocation (supervisor episode, worker session, proposer shot, judge call, analyst pass) captured by the \\`${ROLLOUT_SCHEMA}\\` rollout ledger, labeled with improvement-loop coordinates and the official-judge reward, with the full message transcript inline.\n\nThis release contains the **trainable split only** (\\`search\\`). Holdout, dev, and canary splits are structurally excluded at build time, and the build additionally drops any non-trainable line as a fail-closed filter (${excluded.nonTrain} dropped here).\n\n## Formats\n\n${formatsTable}\n\n## Schema (\\`${ROLLOUT_SCHEMA}\\`)\n\nEach raw line carries:\n\n- \\`rollout_id\\` / \\`parent_rollout_id\\` — invocation identity; workers point at their spawning supervisor episode.\n- \\`run_id\\`, \\`experiment_id\\`, \\`candidate_id\\` — run/experiment/candidate identity from the producing RunRecord, when present.\n- \\`generation\\`, \\`candidate_index\\` — improvement-loop coordinates (\\`-1\\` = baseline campaign; \\`null\\` = not an improvement loop).\n- \\`role\\` — one of ${ROLLOUT_ROLES.map((role) => `\\`${role}\\``).join(', ')}.\n- \\`task\\` — suite, instance id, split, seed, replicate index.\n- \\`policy\\` — harness, model, provider, profile commit, prompt/config hashes, sampling params.\n- \\`messages\\` / \\`tool_defs\\` — full transcript in canonical OpenAI chat-with-tools form (including \\`reasoning_content\\`). An empty \\`messages\\` array is a labeled gap line; \\`provenance.gap\\` says why the transcript could not be recovered (${gapLines} gap lines in this release).\n- \\`outcome\\` — \\`reward\\` (the single scalar), \\`reward_source\\`, the verbatim judge \\`verdict\\`, non-scalar \\`metrics\\`, and \\`realness_gated\\` (anti-Goodhart flag: reward forced to 0, never SFT-eligible).\n- \\`cost\\` — usd, token counts, wall time.\n- \\`artifacts\\` / \\`provenance\\` — patch/run-dir/transcript pointers (scrubbed) and capture metadata.\n\n## Provenance\n\n- Source ledgers: ${sourceFiles.map((file) => `\\`${file}\\``).join(', ')}\n- Run ids: ${runIds.map((id) => `\\`${id}\\``).join(', ')}\n- Generations: ${generations.join(', ')} (\\`-1\\` = baseline campaign)\n- Models: ${models.map((m) => `\\`${m}\\``).join(', ')}\n- Harnesses: ${harnesses.map((h) => `\\`${h}\\``).join(', ')}\n- Capture modes: ${captures.join(', ')}\n- Every reward traces to a named source (reward sources in this release: ${rewardSources.map((s) => `\\`${s}\\``).join(', ')}).\n\n${proposerNote}\n\n## Reward semantics per role\n\n- **agent** — the producing RunRecord's holdout/search score, with the realness gate forcing gamed successes to 0.\n- **supervisor** — the official-judge verdict on the episode's delivered artifact (1 = resolved, 0 = not).\n- **worker** — INHERITED from the parent supervisor episode (\\`…/inherited\\`). Caveat: reward 1 does not establish this worker's individual contribution (sibling workers in the same episode share the episode outcome), and reward 0 does not prove this worker failed.\n- **proposer** — the fraction of improvement-set instances the proposed candidate resolved (\\`…/candidate-resolved-fraction\\`); a scalar in [0, 1], not a binary verdict.\n- **judge / analyst** — carry the episode verdict where one applies; otherwise \\`reward: null\\` (a labeled gap, never 0).\n\n## Counts\n\n${countsTable}\n\nTotal lines: ${lines.length}\n\n## Scrubbing\n\nAbsolute home paths were rewritten to \\`$WORK\\`, credential-shaped strings were replaced with \\`[REDACTED:<kind>]\\` markers, internal hostnames were normalized to \\`*.internal.example\\`, and username-bearing incidentals (\\`ls -l\\` owner columns, per-user pytest tmpdirs) were normalized to \\`user\\` / \\`$USER\\`. Rewrite counts for this release (full per-file breakdown in \\`scrub-report.json\\`):\n\n${scrubTable}\n\n## License\n\n\\`license: unknown\\` is a placeholder — the releasing operator must set the real SPDX license id in the frontmatter above before publishing.\n\n## Citation\n\n\\`\\`\\`bibtex\n@misc{tangle_rollout_ledger,\n  title = {Tangle rollout ledger — agent trajectories},\n  author = {{Tangle Network}},\n  howpublished = {HuggingFace Datasets},\n  note = {Operator: fill in the repository URL, authors, and year before publishing}\n}\n\\`\\`\\`\n`\n}\n","/**\n * Deterministic scrubbing pass over rollout-ledger lines before public release.\n *\n * Every rule is a pure regex rewrite applied to every string value in a line\n * (messages, artifacts, run ids, tool arguments — everywhere), so the scrubbed\n * line is still a valid `tangle.rollout.v1` line. Rules are idempotent:\n * scrub(scrub(x)) === scrub(x), and a second pass counts zero hits — that is\n * the property the release pipeline relies on to prove nothing half-scrubbed\n * ships. Rule order matters: whole `KEY=value` env pairs are redacted before\n * the bare-key rule so one secret is never counted twice.\n */\n\nimport type { RolloutLine } from '../schema'\n\nexport interface ScrubRule {\n  name: string\n  pattern: RegExp\n  /** Rewrite for one match; `g1` is the first capture group when present. */\n  rewrite: (match: string, g1?: string) => string\n}\n\nexport const SCRUB_RULES: readonly ScrubRule[] = [\n  {\n    // /home/<user> and /Users/<user> prefixes → $WORK ($HOME-derived paths).\n    name: 'home-path',\n    pattern: /\\/(?:home|Users)\\/[A-Za-z0-9._-]+/g,\n    rewrite: () => '$WORK',\n  },\n  {\n    // Harness stores encode cwd as a dashed path segment (-home-drew-…);\n    // strip the leading -home-<user> so the username never ships.\n    name: 'home-path-encoded',\n    pattern: /(?<=\\/)-(?:home|Users)-[A-Za-z0-9_.]+/g,\n    rewrite: () => '$WORK',\n  },\n  {\n    // pytest names its tmpdir after the invoking user (/tmp/pytest-of-drew).\n    name: 'tmp-user-dir',\n    pattern: /\\/tmp\\/pytest-of-[A-Za-z0-9._-]+/g,\n    rewrite: () => '/tmp/pytest-of-$USER',\n  },\n  {\n    // `ls -l` owner/group columns leak the username in captured tool output.\n    // The `(?!user user…)` guard keeps a second pass at zero rewrites.\n    name: 'ls-owner',\n    pattern:\n      /([-bcdlps][-rwxsStT]{9}[.+@]?\\s+\\d+\\s+)(?!user user(?=\\s))[A-Za-z0-9._-]+\\s+[A-Za-z0-9._-]+(?=\\s)/g,\n    rewrite: (_match, prefix) => `${prefix}user user`,\n  },\n  {\n    // Env-var-shaped secrets: NAME=value where NAME looks credential-bearing.\n    // The (?!\\[REDACTED:) guard keeps the rule idempotent on its own output.\n    name: 'env-secret',\n    pattern:\n      /\\b([A-Z][A-Z0-9_]*(?:KEY|TOKEN|SECRET|PASSWORD|PASSWD|CREDENTIALS?))=(?!\\[REDACTED:)(\"[^\"]*\"|'[^']*'|[^\\s\"']+)/g,\n    rewrite: (_match, name) => `${name}=[REDACTED:env]`,\n  },\n  {\n    name: 'bearer-token',\n    pattern: /\\b(Bearer|Basic)\\s+(?!\\[REDACTED:)[A-Za-z0-9\\-._~+/=]{8,}/g,\n    rewrite: (_match, scheme) => `${scheme} [REDACTED:bearer]`,\n  },\n  {\n    // Bare provider-prefixed keys (OpenAI/Anthropic sk-, GitHub ghp_/gho_/…,\n    // fine-grained PATs, HuggingFace hf_, Slack xox*, AWS AKIA).\n    name: 'api-key',\n    pattern:\n      /\\b(?:sk-[A-Za-z0-9_-]{16,}|gh[pousr]_[A-Za-z0-9]{16,}|github_pat_[A-Za-z0-9_]{20,}|hf_[A-Za-z0-9]{16,}|xox[baprs]-[A-Za-z0-9-]{10,}|AKIA[0-9A-Z]{16})\\b/g,\n    rewrite: () => '[REDACTED:api-key]',\n  },\n  {\n    // Our infra hostnames → placeholder domain, subdomain preserved\n    // (router.tangle.tools → router.internal.example).\n    name: 'infra-host',\n    pattern: /(?<![A-Za-z0-9.-])((?:[A-Za-z0-9-]+\\.)*)tangle\\.(?:tools|network)(?![A-Za-z0-9-])/g,\n    rewrite: (_match, prefix) => `${prefix ?? ''}internal.example`,\n  },\n  {\n    // Known workstation hostnames leak through email Message-IDs and fqdn\n    // lookups inside captured test output; extend the list as machines join.\n    name: 'machine-host',\n    pattern: /\\b[A-Za-z0-9]+-GTR-Pro\\b/g,\n    rewrite: () => 'workstation',\n  },\n]\n\n/** Rule name → number of matches rewritten. Always carries every rule (0 is data). */\nexport type ScrubCounts = Record<string, number>\n\nexport function emptyScrubCounts(): ScrubCounts {\n  return Object.fromEntries(SCRUB_RULES.map((rule) => [rule.name, 0]))\n}\n\nexport function addScrubCounts(into: ScrubCounts, from: ScrubCounts): ScrubCounts {\n  for (const [name, count] of Object.entries(from)) into[name] = (into[name] ?? 0) + count\n  return into\n}\n\nexport function scrubText(text: string, counts: ScrubCounts): string {\n  let out = text\n  for (const rule of SCRUB_RULES) {\n    out = out.replace(rule.pattern, (match: string, ...rest: unknown[]) => {\n      counts[rule.name] = (counts[rule.name] ?? 0) + 1\n      return rule.rewrite(match, typeof rest[0] === 'string' ? rest[0] : undefined)\n    })\n  }\n  return out\n}\n\nfunction scrubValue(value: unknown, counts: ScrubCounts): unknown {\n  if (typeof value === 'string') return scrubText(value, counts)\n  if (Array.isArray(value)) return value.map((item) => scrubValue(item, counts))\n  if (value !== null && typeof value === 'object') {\n    const out: Record<string, unknown> = {}\n    for (const [key, item] of Object.entries(value)) out[key] = scrubValue(item, counts)\n    return out\n  }\n  return value\n}\n\n/** Scrub every string value in a line; structure and key order are preserved. */\nexport function scrubRolloutLine(line: RolloutLine, counts: ScrubCounts): RolloutLine {\n  return scrubValue(line, counts) as RolloutLine\n}\n\nexport function scrubLines(lines: RolloutLine[]): { lines: RolloutLine[]; counts: ScrubCounts } {\n  const counts = emptyScrubCounts()\n  return { lines: lines.map((line) => scrubRolloutLine(line, counts)), counts }\n}\n\n/**\n * A `RolloutScrubber` (text → text) applying the full rule set — the\n * default hook to pass to `mintRolloutRows({ scrub })` so lines are\n * scrubbed at mint time, before they ever reach a ledger file. Release\n * builds re-run `scrubLines` regardless (idempotent), so double-scrubbing\n * is safe and counted as zero.\n */\nexport function defaultRolloutScrubber(text: string): string {\n  return scrubText(text, emptyScrubCounts())\n}\n","/**\n * One-command HuggingFace dataset release from rollout ledgers:\n *\n *   agent-eval rollout-release <ledger.jsonl...> --out <dir> \\\n *     [--formats sft,verifiers,rft,raw] [--include-proposers] [--push <org/name>]\n *\n * Pipeline per input ledger: read + validate → fail-closed filters\n * (trainable split only; proposer sessions dropped unless\n * --include-proposers, they contain improvement-loop harness source) →\n * deterministic scrub → export the requested formats + scrub-report.json +\n * auto-generated README.md card. Deterministic: same inputs and flags →\n * byte-identical output dir.\n *\n * --push uploads the built dir with `huggingface-cli upload` only when the\n * CLI exists on PATH and HF_TOKEN is present in the env; the token is\n * never printed. Everything else runs fully offline.\n */\n\nimport { spawnSync } from 'node:child_process'\nimport { mkdir, writeFile } from 'node:fs/promises'\nimport { basename, dirname, join } from 'node:path'\nimport { toJsonl, toRftItems, toSftRows, toVerifiersRolloutOutputs } from '../exporters'\nimport { readRolloutLedger, writeRolloutLedger } from '../ledger'\nimport { isTrainableSplit, type RolloutLine } from '../schema'\nimport { buildDatasetCard, FORMAT_FILES, RELEASE_FORMATS, type ReleaseFormat } from './card'\nimport { addScrubCounts, emptyScrubCounts, type ScrubCounts, scrubLines } from './scrub'\n\nexport interface BuildOptions {\n  out: string\n  formats: ReleaseFormat[]\n  includeProposers: boolean\n}\n\nexport interface ScrubReport {\n  /** Input ledger path → rule → rewrite count (only shipped lines are scrubbed). */\n  files: Record<string, ScrubCounts>\n  totals: ScrubCounts\n  excluded: { proposers: number; nonTrain: number }\n}\n\nexport interface BuildSummary {\n  inputs: string[]\n  read: number\n  kept: number\n  scrub: ScrubReport\n  formatCounts: Partial<Record<ReleaseFormat, number>>\n  files: string[]\n}\n\nexport async function buildHfDataset(\n  inputs: string[],\n  options: BuildOptions,\n): Promise<BuildSummary> {\n  if (inputs.length === 0) throw new Error('no input ledgers given')\n  if (options.formats.length === 0) throw new Error('no formats selected')\n\n  const report: ScrubReport = {\n    files: {},\n    totals: emptyScrubCounts(),\n    excluded: { proposers: 0, nonTrain: 0 },\n  }\n  const kept: RolloutLine[] = []\n  let read = 0\n\n  for (const input of inputs) {\n    const lines = await readRolloutLedger(input)\n    read += lines.length\n    const shippable = lines.filter((line) => {\n      if (!isTrainableSplit(line.task.split)) {\n        report.excluded.nonTrain += 1\n        return false\n      }\n      if (!options.includeProposers && line.role === 'proposer') {\n        report.excluded.proposers += 1\n        return false\n      }\n      return true\n    })\n    const scrubbed = scrubLines(shippable)\n    report.files[input] = scrubbed.counts\n    addScrubCounts(report.totals, scrubbed.counts)\n    kept.push(...scrubbed.lines)\n  }\n\n  const formatCounts: Partial<Record<ReleaseFormat, number>> = {}\n  const files: string[] = []\n\n  for (const format of options.formats) {\n    const path = join(options.out, FORMAT_FILES[format])\n    await mkdir(dirname(path), { recursive: true })\n    if (format === 'raw') {\n      // writeRolloutLedger re-validates every scrubbed line before it lands.\n      await writeRolloutLedger(path, kept)\n      formatCounts.raw = kept.length\n    } else if (format === 'sft') {\n      const rows = toSftRows(kept)\n      await writeFile(path, toJsonl(rows))\n      formatCounts.sft = rows.length\n    } else if (format === 'verifiers') {\n      const outputs = toVerifiersRolloutOutputs(kept)\n      await writeFile(path, toJsonl(outputs))\n      formatCounts.verifiers = outputs.length\n    } else {\n      const items = toRftItems(kept)\n      await writeFile(path, toJsonl(items))\n      formatCounts.rft = items.length\n    }\n    files.push(path)\n  }\n\n  const reportPath = join(options.out, 'scrub-report.json')\n  await writeFile(reportPath, `${JSON.stringify(report, null, 2)}\\n`)\n  files.push(reportPath)\n\n  const cardPath = join(options.out, 'README.md')\n  await writeFile(\n    cardPath,\n    buildDatasetCard({\n      lines: kept,\n      formats: options.formats,\n      includeProposers: options.includeProposers,\n      sourceFiles: inputs.map((input) => basename(input)),\n      scrubTotals: report.totals,\n      excluded: report.excluded,\n      formatCounts,\n    }),\n  )\n  files.push(cardPath)\n\n  return { inputs, read, kept: kept.length, scrub: report, formatCounts, files }\n}\n\nexport function planPushCommand(repo: string, outDir: string): string[] {\n  return ['huggingface-cli', 'upload', repo, outDir, '.', '--repo-type', 'dataset']\n}\n\nexport function pushDataset(repo: string, outDir: string): void {\n  const found = spawnSync('which', ['huggingface-cli'], { stdio: 'ignore' })\n  if (found.status !== 0) {\n    throw new Error(\n      'huggingface-cli not found on PATH — install huggingface_hub[cli] before --push',\n    )\n  }\n  if (!process.env.HF_TOKEN) {\n    throw new Error('HF_TOKEN not present in env — refusing to push')\n  }\n  const [command, ...args] = planPushCommand(repo, outDir) as [string, ...string[]]\n  // Token stays in the inherited env; it is never echoed or interpolated.\n  const run = spawnSync(command, args, { stdio: 'inherit' })\n  if (run.status !== 0) throw new Error(`huggingface-cli upload exited ${String(run.status)}`)\n}\n\nexport interface RolloutReleaseCliArgs extends BuildOptions {\n  inputs: string[]\n  push: string | null\n}\n\nexport const ROLLOUT_RELEASE_USAGE =\n  'usage: agent-eval rollout-release <ledger.jsonl...> --out <dir> [--formats sft,verifiers,rft,raw] [--include-proposers] [--push <org/name>]'\n\nexport function parseRolloutReleaseArgs(argv: string[]): RolloutReleaseCliArgs {\n  const args: RolloutReleaseCliArgs = {\n    inputs: [],\n    out: '',\n    formats: [...RELEASE_FORMATS],\n    includeProposers: false,\n    push: null,\n  }\n  for (let i = 0; i < argv.length; i++) {\n    const arg = argv[i]!\n    if (arg === '--out') {\n      args.out = argv[++i] ?? ''\n    } else if (arg === '--formats') {\n      const raw = (argv[++i] ?? '').split(',').filter(Boolean)\n      for (const format of raw) {\n        if (!RELEASE_FORMATS.includes(format as ReleaseFormat)) {\n          throw new Error(\n            `unknown format \"${format}\" — expected one of ${RELEASE_FORMATS.join(',')}`,\n          )\n        }\n      }\n      args.formats = raw as ReleaseFormat[]\n    } else if (arg === '--include-proposers') {\n      args.includeProposers = true\n    } else if (arg === '--push') {\n      args.push = argv[++i] ?? null\n    } else if (arg.startsWith('--')) {\n      throw new Error(`unknown flag \"${arg}\"`)\n    } else {\n      args.inputs.push(arg)\n    }\n  }\n  if (args.inputs.length === 0 || !args.out) {\n    throw new Error(ROLLOUT_RELEASE_USAGE)\n  }\n  if (args.push !== null && !/^[\\w.-]+\\/[\\w.-]+$/.test(args.push)) {\n    throw new Error(`--push expects <org/name>, got \"${args.push}\"`)\n  }\n  return args\n}\n\n/** CLI driver for `agent-eval rollout-release`. Returns the process exit code. */\nexport async function runRolloutReleaseCli(argv: string[]): Promise<number> {\n  let args: RolloutReleaseCliArgs\n  try {\n    args = parseRolloutReleaseArgs(argv)\n  } catch (error) {\n    process.stderr.write(`${error instanceof Error ? error.message : String(error)}\\n`)\n    return 2\n  }\n  const summary = await buildHfDataset(args.inputs, args)\n  process.stdout.write(\n    `${JSON.stringify(\n      {\n        read: summary.read,\n        kept: summary.kept,\n        formatCounts: summary.formatCounts,\n        scrub: summary.scrub,\n      },\n      null,\n      2,\n    )}\\n`,\n  )\n  process.stdout.write(`dataset → ${args.out} (${summary.files.length} files)\\n`)\n  if (args.push !== null) {\n    pushDataset(args.push, args.out)\n    process.stdout.write(`pushed → ${args.push}\\n`)\n  }\n  return 0\n}\n"],"mappings":";;;;;;;;AA6CO,SAAS,UAAU,OAAsB,UAA4B,CAAC,GAAa;AACxF,SAAO,MACJ,OAAO,CAAC,SAAS,uBAAuB,MAAM,OAAO,KAAK,KAAK,SAAS,SAAS,CAAC,EAClF,IAAI,CAAC,UAAU;AAAA,IACd,UAAU,KAAK;AAAA,IACf,UAAU;AAAA,MACR,YAAY,KAAK;AAAA,MACjB,QAAQ,KAAK;AAAA,MACb,cAAc,KAAK,gBAAgB;AAAA,MACnC,aAAa,KAAK,KAAK;AAAA,MACvB,QAAQ,KAAK,QAAQ;AAAA,IACvB;AAAA,EACF,EAAE;AACN;AAuBO,SAAS,aACd,OACA,UAAiC,CAAC,GACrB;AACb,SAAO,MACJ;AAAA,IACC,CAAC,SACC,uBAAuB,MAAM,OAAO,KACpC,KAAK,SAAS;AAAA,MACZ,CAAC,YACC,QAAQ,SAAS,UACjB,OAAO,QAAQ,YAAY,YAC3B,QAAQ,QAAQ,SAAS;AAAA,IAC7B;AAAA,EACJ,EACC,IAAI,CAAC,UAAU;AAAA,IACd,QAAQ,KAAK,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,MAAM,GAAG,WAAW;AAAA,IACjE,OAAO,KAAK,SAAS,CAAC;AAAA,IACtB,QAAQ,KAAK,QAAQ;AAAA,IACrB,UAAU;AAAA,MACR,YAAY,KAAK;AAAA,MACjB,QAAQ,KAAK;AAAA,MACb,cAAc,KAAK,gBAAgB;AAAA,MACnC,aAAa,KAAK,KAAK;AAAA,MACvB,OAAO,KAAK,KAAK;AAAA,IACnB;AAAA,EACF,EAAE;AACN;AAqCA,SAAS,oBAAoB,UAAiC;AAC5D,QAAM,QAAQ,SAAS,UAAU,CAAC,MAAM,EAAE,SAAS,WAAW;AAC9D,SAAO,UAAU,KAAK,SAAS,SAAS;AAC1C;AAEO,SAAS,yBAAyB,MAA2C;AAClF,QAAM,QAAQ,oBAAoB,KAAK,QAAQ;AAC/C,SAAO;AAAA,IACL,QAAQ,KAAK,SAAS,MAAM,GAAG,KAAK;AAAA,IACpC,YAAY,KAAK,SAAS,MAAM,KAAK;AAAA,IACrC,QAAQ,KAAK,QAAQ;AAAA,IACrB,SAAS,KAAK,QAAQ;AAAA,IACtB,WAAW,KAAK;AAAA,IAChB,aAAa;AAAA,MACX,cAAc,KAAK,KAAK;AAAA,MACxB,eAAe,KAAK,KAAK;AAAA,MACzB,kBAAkB,KAAK,KAAK;AAAA,MAC5B,mBAAmB,KAAK,KAAK;AAAA,MAC7B,oBAAoB,KAAK,KAAK;AAAA,IAChC;AAAA,IACA,MAAM;AAAA,MACJ,MAAM,KAAK;AAAA,MACX,QAAQ,KAAK;AAAA,MACb,YAAY,KAAK;AAAA,MACjB,QAAQ,KAAK;AAAA,MACb,eAAe,KAAK,iBAAiB;AAAA,MACrC,cAAc,KAAK,gBAAgB;AAAA,MACnC,YAAY,KAAK;AAAA,MACjB,iBAAiB,KAAK;AAAA,MACtB,MAAM,KAAK;AAAA,IACb;AAAA,EACF;AACF;AAEO,SAAS,0BACd,OACA,UAAiC,CAAC,GACR;AAC1B,SAAO,MACJ;AAAA,IACC,CAAC,SACC,uBAAuB,MAAM,OAAO,KACpC,oBAAoB,KAAK,QAAQ,IAAI,KACrC,oBAAoB,KAAK,QAAQ,IAAI,KAAK,SAAS;AAAA,EACvD,EACC,IAAI,wBAAwB;AACjC;AAqBO,SAAS,UAAU,MAA4B;AACpD,QAAM,QAAQ,oBAAoB,KAAK,QAAQ;AAC/C,SAAO;AAAA,IACL,UAAU,KAAK,SAAS,MAAM,GAAG,KAAK;AAAA,IACtC,WAAW;AAAA,MACT,QAAQ,KAAK,QAAQ;AAAA,MACrB,eAAe,KAAK,QAAQ;AAAA,MAC5B,SAAS,KAAK,QAAQ;AAAA,MACtB,aAAa,KAAK,KAAK;AAAA,MACvB,OAAO,KAAK,KAAK;AAAA,MACjB,OAAO,KAAK,KAAK;AAAA,MACjB,YAAY,KAAK;AAAA,IACnB;AAAA,EACF;AACF;AAGO,SAAS,WAAW,OAAsB,UAAiC,CAAC,GAAc;AAC/F,SAAO,MACJ;AAAA,IACC,CAAC,SACC,uBAAuB,MAAM,OAAO,KACpC,KAAK,SAAS,SAAS,KACvB,oBAAoB,KAAK,QAAQ,IAAI;AAAA,EACzC,EACC,IAAI,SAAS;AAClB;AAOO,SAAS,QAAQ,MAAsC;AAC5D,SAAO,KAAK,IAAI,CAAC,MAAM,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,KAAK,KAAK,SAAS,OAAO;AAC/E;AAEA,SAAS,uBACP,MACA,SACgF;AAChF,QAAM,0BAA0B,QAAQ,2BAA2B;AACnE,MAAI,CAAC,OAAO,SAAS,uBAAuB,GAAG;AAC7C,UAAM,IAAI,MAAM,wCAAwC;AAAA,EAC1D;AAEA,QAAM,SAAS,KAAK,QAAQ;AAC5B,MAAI,WAAW,KAAM,QAAO;AAC5B,MAAI,CAAC,OAAO,SAAS,MAAM,GAAG;AAC5B,UAAM,IAAI,MAAM,gCAAgC,KAAK,UAAU,kBAAkB;AAAA,EACnF;AACA,MAAI,UAAU,wBAAyB,QAAO;AAC9C,MAAI,CAAC,KAAK,QAAQ,gBAAgB,KAAK,QAAQ,gBAAgB,KAAK,QAAQ,UAAU,MAAM;AAC1F,WAAO;AAAA,EACT;AACA,MAAI,KAAK,QAAQ,mBAAmB,KAAM,QAAO;AACjD,MAAI,KAAK,KAAK,UAAU,SAAU,QAAO;AACzC,SAAO,KAAK,KAAK,UAAU,aAAa,QAAQ,6BAA6B;AAC/E;;;ACvQA,SAAS,YAAY,OAAO,UAAU,iBAAiB;AACvD,SAAS,eAAe;AAGxB,SAAS,UAAU,OAA8B;AAC/C,aAAW,CAAC,GAAG,IAAI,KAAK,MAAM,QAAQ,EAAG,mBAAkB,MAAM,iBAAiB,CAAC,GAAG;AACtF,SAAO,MAAM,IAAI,CAAC,SAAS,KAAK,UAAU,IAAI,CAAC,EAAE,KAAK,IAAI,KAAK,MAAM,SAAS,IAAI,OAAO;AAC3F;AAGA,eAAsB,mBAAmB,MAAc,OAAqC;AAC1F,QAAM,UAAU,UAAU,KAAK;AAC/B,QAAM,MAAM,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;AAC9C,QAAM,UAAU,MAAM,OAAO;AAC/B;AAGA,eAAsB,mBAAmB,MAAc,OAAqC;AAC1F,MAAI,MAAM,WAAW,EAAG;AACxB,QAAM,UAAU,UAAU,KAAK;AAC/B,QAAM,MAAM,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;AAC9C,QAAM,WAAW,MAAM,OAAO;AAChC;AAMA,eAAsB,kBAAkB,MAAsC;AAC5E,QAAM,MAAM,MAAM,SAAS,MAAM,MAAM;AACvC,QAAM,QAAuB,CAAC;AAC9B,QAAM,WAAW,IAAI,MAAM,IAAI;AAC/B,WAAS,IAAI,GAAG,IAAI,SAAS,QAAQ,KAAK;AACxC,UAAM,OAAO,SAAS,CAAC;AACvB,QAAI,CAAC,MAAM,KAAK,EAAG;AACnB,QAAI;AACJ,QAAI;AACF,eAAS,KAAK,MAAM,IAAI;AAAA,IAC1B,SAAS,OAAO;AACd,YAAM,IAAI;AAAA,QACR,GAAG,IAAI,IAAI,IAAI,CAAC,2BAAsB,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK,CAAC;AAAA,MAC9F;AAAA,IACF;AACA,sBAAkB,QAAQ,GAAG,IAAI,IAAI,IAAI,CAAC,EAAE;AAC5C,UAAM,KAAK,MAAM;AAAA,EACnB;AACA,SAAO;AACT;;;ACzCO,IAAM,kBAAkB,CAAC,OAAO,aAAa,OAAO,KAAK;AAIzD,IAAM,eAA8C;AAAA,EACzD,KAAK;AAAA,EACL,WAAW;AAAA,EACX,KAAK;AAAA,EACL,KAAK;AACP;AAEA,IAAM,sBAAqD;AAAA,EACzD,KAAK;AAAA,EACL,WACE;AAAA,EACF,KAAK;AAAA,EACL,KAAK,UAAU,cAAc;AAC/B;AAcA,SAAS,OAAO,QAAwC;AACtD,SAAO,CAAC,GAAG,IAAI,IAAI,OAAO,OAAO,CAAC,MAAmB,MAAM,IAAI,CAAC,CAAC,EAAE,KAAK;AAC1E;AAEA,SAAS,aAAa,QAA+B;AACnD,MAAI,WAAW,KAAM,QAAO;AAC5B,SAAO,OAAO,UAAU,MAAM,IAAI,OAAO,MAAM,IAAI,OAAO,QAAQ,CAAC;AACrE;AAEA,SAAS,cAAc,QAAkB,MAA0B;AACjE,SAAO;AAAA,IACL,KAAK,OAAO,KAAK,KAAK,CAAC;AAAA,IACvB,KAAK,OAAO,IAAI,MAAM,KAAK,EAAE,KAAK,KAAK,CAAC;AAAA,IACxC,GAAG,KAAK,IAAI,CAAC,QAAQ,KAAK,IAAI,KAAK,KAAK,CAAC,IAAI;AAAA,EAC/C,EAAE,KAAK,IAAI;AACb;AAEA,SAAS,eAAe,OAAkC;AACxD,QAAM,SAAS,oBAAI,IAAsC;AACzD,aAAW,QAAQ,OAAO;AACxB,UAAM,WAAW,OAAO,IAAI,KAAK,IAAI,KAAK,oBAAI,IAAoB;AAClE,UAAM,MAAM,aAAa,KAAK,QAAQ,MAAM;AAC5C,aAAS,IAAI,MAAM,SAAS,IAAI,GAAG,KAAK,KAAK,CAAC;AAC9C,WAAO,IAAI,KAAK,MAAM,QAAQ;AAAA,EAChC;AACA,QAAM,OAAmB,CAAC;AAC1B,aAAW,QAAQ,eAAe;AAChC,UAAM,WAAW,OAAO,IAAI,IAAI;AAChC,QAAI,CAAC,SAAU;AACf,UAAM,OAAO,CAAC,GAAG,SAAS,KAAK,CAAC,EAAE,KAAK,CAAC,GAAG,MAAM;AAC/C,UAAI,MAAM,OAAQ,QAAO;AACzB,UAAI,MAAM,OAAQ,QAAO;AACzB,aAAO,OAAO,CAAC,IAAI,OAAO,CAAC;AAAA,IAC7B,CAAC;AACD,eAAW,OAAO,KAAM,MAAK,KAAK,CAAC,MAAM,KAAK,OAAO,SAAS,IAAI,GAAG,CAAC,CAAC,CAAC;AAAA,EAC1E;AACA,SAAO;AACT;AAEO,SAAS,iBAAiB,QAAmC;AAClE,QAAM,EAAE,OAAO,SAAS,kBAAkB,aAAa,aAAa,UAAU,aAAa,IACzF;AAEF,QAAM,SAAS,OAAO,MAAM,IAAI,CAAC,SAAS,KAAK,MAAM,CAAC;AACtD,QAAM,cAAc,CAAC,GAAG,IAAI,IAAI,MAAM,IAAI,CAAC,SAAS,KAAK,UAAU,CAAC,CAAC,EAClE,OAAO,CAAC,MAAmB,MAAM,IAAI,EACrC,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AACvB,QAAM,SAAS,OAAO,MAAM,IAAI,CAAC,SAAS,KAAK,OAAO,KAAK,CAAC;AAC5D,QAAM,YAAY,OAAO,MAAM,IAAI,CAAC,SAAS,KAAK,OAAO,OAAO,CAAC;AACjE,QAAM,WAAW,OAAO,MAAM,IAAI,CAAC,SAAS,KAAK,WAAW,OAAO,CAAC;AACpE,QAAM,gBAAgB,OAAO,MAAM,IAAI,CAAC,SAAS,KAAK,QAAQ,aAAa,CAAC;AAC5E,QAAM,WAAW,MAAM,OAAO,CAAC,SAAS,KAAK,SAAS,WAAW,CAAC,EAAE;AAEpE,QAAM,UAAU,QACb;AAAA,IAAI,CAAC,WACJ;AAAA,MACE,oBAAoB,MAAM;AAAA,MAC1B;AAAA,MACA;AAAA,MACA,iBAAiB,aAAa,MAAM,CAAC;AAAA,IACvC,EAAE,KAAK,IAAI;AAAA,EACb,EACC,KAAK,IAAI;AAEZ,QAAM,cAAc;AAAA,IAClB;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,EAAE,KAAK,IAAI;AAEX,QAAM,eAAe;AAAA,IACnB,CAAC,UAAU,QAAQ,QAAQ,UAAU;AAAA,IACrC,QAAQ,IAAI,CAAC,WAAW;AAAA,MACtB;AAAA,MACA,KAAK,aAAa,MAAM,CAAC;AAAA,MACzB,OAAO,aAAa,MAAM,KAAK,CAAC;AAAA,MAChC,oBAAoB,MAAM;AAAA,IAC5B,CAAC;AAAA,EACH;AAEA,QAAM,cAAc,cAAc,CAAC,QAAQ,UAAU,OAAO,GAAG,eAAe,KAAK,CAAC;AAEpF,QAAM,aAAa;AAAA,IACjB,CAAC,QAAQ,UAAU;AAAA,IACnB,OAAO,QAAQ,WAAW,EAAE,IAAI,CAAC,CAAC,MAAM,KAAK,MAAM,CAAC,MAAM,OAAO,KAAK,CAAC,CAAC;AAAA,EAC1E;AAEA,QAAM,eAAe,mBACjB,uHACA,8CAA8C,SAAS,SAAS;AAEpE,SAAO,GAAG,WAAW;AAAA;AAAA;AAAA;AAAA,gIAIyG,cAAc;AAAA;AAAA,6NAE+E,SAAS,QAAQ;AAAA;AAAA;AAAA;AAAA,EAI5O,YAAY;AAAA;AAAA,eAEC,cAAc;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,2BAOP,cAAc,IAAI,CAAC,SAAS,KAAK,IAAI,IAAI,EAAE,KAAK,IAAI,CAAC;AAAA;AAAA;AAAA,0PAG0K,QAAQ;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,oBAOzO,YAAY,IAAI,CAAC,SAAS,KAAK,IAAI,IAAI,EAAE,KAAK,IAAI,CAAC;AAAA,aAC1D,OAAO,IAAI,CAAC,OAAO,KAAK,EAAE,IAAI,EAAE,KAAK,IAAI,CAAC;AAAA,iBACtC,YAAY,KAAK,IAAI,CAAC;AAAA,YAC3B,OAAO,IAAI,CAAC,MAAM,KAAK,CAAC,IAAI,EAAE,KAAK,IAAI,CAAC;AAAA,eACrC,UAAU,IAAI,CAAC,MAAM,KAAK,CAAC,IAAI,EAAE,KAAK,IAAI,CAAC;AAAA,mBACvC,SAAS,KAAK,IAAI,CAAC;AAAA,2EACqC,cAAc,IAAI,CAAC,MAAM,KAAK,CAAC,IAAI,EAAE,KAAK,IAAI,CAAC;AAAA;AAAA,EAExH,YAAY;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAYZ,WAAW;AAAA;AAAA,eAEE,MAAM,MAAM;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAMzB,UAAU;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAiBZ;;;AC/LO,IAAM,cAAoC;AAAA,EAC/C;AAAA;AAAA,IAEE,MAAM;AAAA,IACN,SAAS;AAAA,IACT,SAAS,MAAM;AAAA,EACjB;AAAA,EACA;AAAA;AAAA;AAAA,IAGE,MAAM;AAAA,IACN,SAAS;AAAA,IACT,SAAS,MAAM;AAAA,EACjB;AAAA,EACA;AAAA;AAAA,IAEE,MAAM;AAAA,IACN,SAAS;AAAA,IACT,SAAS,MAAM;AAAA,EACjB;AAAA,EACA;AAAA;AAAA;AAAA,IAGE,MAAM;AAAA,IACN,SACE;AAAA,IACF,SAAS,CAAC,QAAQ,WAAW,GAAG,MAAM;AAAA,EACxC;AAAA,EACA;AAAA;AAAA;AAAA,IAGE,MAAM;AAAA,IACN,SACE;AAAA,IACF,SAAS,CAAC,QAAQ,SAAS,GAAG,IAAI;AAAA,EACpC;AAAA,EACA;AAAA,IACE,MAAM;AAAA,IACN,SAAS;AAAA,IACT,SAAS,CAAC,QAAQ,WAAW,GAAG,MAAM;AAAA,EACxC;AAAA,EACA;AAAA;AAAA;AAAA,IAGE,MAAM;AAAA,IACN,SACE;AAAA,IACF,SAAS,MAAM;AAAA,EACjB;AAAA,EACA;AAAA;AAAA;AAAA,IAGE,MAAM;AAAA,IACN,SAAS;AAAA,IACT,SAAS,CAAC,QAAQ,WAAW,GAAG,UAAU,EAAE;AAAA,EAC9C;AAAA,EACA;AAAA;AAAA;AAAA,IAGE,MAAM;AAAA,IACN,SAAS;AAAA,IACT,SAAS,MAAM;AAAA,EACjB;AACF;AAKO,SAAS,mBAAgC;AAC9C,SAAO,OAAO,YAAY,YAAY,IAAI,CAAC,SAAS,CAAC,KAAK,MAAM,CAAC,CAAC,CAAC;AACrE;AAEO,SAAS,eAAe,MAAmB,MAAgC;AAChF,aAAW,CAAC,MAAM,KAAK,KAAK,OAAO,QAAQ,IAAI,EAAG,MAAK,IAAI,KAAK,KAAK,IAAI,KAAK,KAAK;AACnF,SAAO;AACT;AAEO,SAAS,UAAU,MAAc,QAA6B;AACnE,MAAI,MAAM;AACV,aAAW,QAAQ,aAAa;AAC9B,UAAM,IAAI,QAAQ,KAAK,SAAS,CAAC,UAAkB,SAAoB;AACrE,aAAO,KAAK,IAAI,KAAK,OAAO,KAAK,IAAI,KAAK,KAAK;AAC/C,aAAO,KAAK,QAAQ,OAAO,OAAO,KAAK,CAAC,MAAM,WAAW,KAAK,CAAC,IAAI,MAAS;AAAA,IAC9E,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAEA,SAAS,WAAW,OAAgB,QAA8B;AAChE,MAAI,OAAO,UAAU,SAAU,QAAO,UAAU,OAAO,MAAM;AAC7D,MAAI,MAAM,QAAQ,KAAK,EAAG,QAAO,MAAM,IAAI,CAAC,SAAS,WAAW,MAAM,MAAM,CAAC;AAC7E,MAAI,UAAU,QAAQ,OAAO,UAAU,UAAU;AAC/C,UAAM,MAA+B,CAAC;AACtC,eAAW,CAAC,KAAK,IAAI,KAAK,OAAO,QAAQ,KAAK,EAAG,KAAI,GAAG,IAAI,WAAW,MAAM,MAAM;AACnF,WAAO;AAAA,EACT;AACA,SAAO;AACT;AAGO,SAAS,iBAAiB,MAAmB,QAAkC;AACpF,SAAO,WAAW,MAAM,MAAM;AAChC;AAEO,SAAS,WAAW,OAAqE;AAC9F,QAAM,SAAS,iBAAiB;AAChC,SAAO,EAAE,OAAO,MAAM,IAAI,CAAC,SAAS,iBAAiB,MAAM,MAAM,CAAC,GAAG,OAAO;AAC9E;AASO,SAAS,uBAAuB,MAAsB;AAC3D,SAAO,UAAU,MAAM,iBAAiB,CAAC;AAC3C;;;ACzHA,SAAS,iBAAiB;AAC1B,SAAS,SAAAA,QAAO,aAAAC,kBAAiB;AACjC,SAAS,UAAU,WAAAC,UAAS,YAAY;AA6BxC,eAAsB,eACpB,QACA,SACuB;AACvB,MAAI,OAAO,WAAW,EAAG,OAAM,IAAI,MAAM,wBAAwB;AACjE,MAAI,QAAQ,QAAQ,WAAW,EAAG,OAAM,IAAI,MAAM,qBAAqB;AAEvE,QAAM,SAAsB;AAAA,IAC1B,OAAO,CAAC;AAAA,IACR,QAAQ,iBAAiB;AAAA,IACzB,UAAU,EAAE,WAAW,GAAG,UAAU,EAAE;AAAA,EACxC;AACA,QAAM,OAAsB,CAAC;AAC7B,MAAI,OAAO;AAEX,aAAW,SAAS,QAAQ;AAC1B,UAAM,QAAQ,MAAM,kBAAkB,KAAK;AAC3C,YAAQ,MAAM;AACd,UAAM,YAAY,MAAM,OAAO,CAAC,SAAS;AACvC,UAAI,CAAC,iBAAiB,KAAK,KAAK,KAAK,GAAG;AACtC,eAAO,SAAS,YAAY;AAC5B,eAAO;AAAA,MACT;AACA,UAAI,CAAC,QAAQ,oBAAoB,KAAK,SAAS,YAAY;AACzD,eAAO,SAAS,aAAa;AAC7B,eAAO;AAAA,MACT;AACA,aAAO;AAAA,IACT,CAAC;AACD,UAAM,WAAW,WAAW,SAAS;AACrC,WAAO,MAAM,KAAK,IAAI,SAAS;AAC/B,mBAAe,OAAO,QAAQ,SAAS,MAAM;AAC7C,SAAK,KAAK,GAAG,SAAS,KAAK;AAAA,EAC7B;AAEA,QAAM,eAAuD,CAAC;AAC9D,QAAM,QAAkB,CAAC;AAEzB,aAAW,UAAU,QAAQ,SAAS;AACpC,UAAM,OAAO,KAAK,QAAQ,KAAK,aAAa,MAAM,CAAC;AACnD,UAAMC,OAAMC,SAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;AAC9C,QAAI,WAAW,OAAO;AAEpB,YAAM,mBAAmB,MAAM,IAAI;AACnC,mBAAa,MAAM,KAAK;AAAA,IAC1B,WAAW,WAAW,OAAO;AAC3B,YAAM,OAAO,UAAU,IAAI;AAC3B,YAAMC,WAAU,MAAM,QAAQ,IAAI,CAAC;AACnC,mBAAa,MAAM,KAAK;AAAA,IAC1B,WAAW,WAAW,aAAa;AACjC,YAAM,UAAU,0BAA0B,IAAI;AAC9C,YAAMA,WAAU,MAAM,QAAQ,OAAO,CAAC;AACtC,mBAAa,YAAY,QAAQ;AAAA,IACnC,OAAO;AACL,YAAM,QAAQ,WAAW,IAAI;AAC7B,YAAMA,WAAU,MAAM,QAAQ,KAAK,CAAC;AACpC,mBAAa,MAAM,MAAM;AAAA,IAC3B;AACA,UAAM,KAAK,IAAI;AAAA,EACjB;AAEA,QAAM,aAAa,KAAK,QAAQ,KAAK,mBAAmB;AACxD,QAAMA,WAAU,YAAY,GAAG,KAAK,UAAU,QAAQ,MAAM,CAAC,CAAC;AAAA,CAAI;AAClE,QAAM,KAAK,UAAU;AAErB,QAAM,WAAW,KAAK,QAAQ,KAAK,WAAW;AAC9C,QAAMA;AAAA,IACJ;AAAA,IACA,iBAAiB;AAAA,MACf,OAAO;AAAA,MACP,SAAS,QAAQ;AAAA,MACjB,kBAAkB,QAAQ;AAAA,MAC1B,aAAa,OAAO,IAAI,CAAC,UAAU,SAAS,KAAK,CAAC;AAAA,MAClD,aAAa,OAAO;AAAA,MACpB,UAAU,OAAO;AAAA,MACjB;AAAA,IACF,CAAC;AAAA,EACH;AACA,QAAM,KAAK,QAAQ;AAEnB,SAAO,EAAE,QAAQ,MAAM,MAAM,KAAK,QAAQ,OAAO,QAAQ,cAAc,MAAM;AAC/E;AAEO,SAAS,gBAAgB,MAAc,QAA0B;AACtE,SAAO,CAAC,mBAAmB,UAAU,MAAM,QAAQ,KAAK,eAAe,SAAS;AAClF;AAEO,SAAS,YAAY,MAAc,QAAsB;AAC9D,QAAM,QAAQ,UAAU,SAAS,CAAC,iBAAiB,GAAG,EAAE,OAAO,SAAS,CAAC;AACzE,MAAI,MAAM,WAAW,GAAG;AACtB,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,MAAI,CAAC,QAAQ,IAAI,UAAU;AACzB,UAAM,IAAI,MAAM,qDAAgD;AAAA,EAClE;AACA,QAAM,CAAC,SAAS,GAAG,IAAI,IAAI,gBAAgB,MAAM,MAAM;AAEvD,QAAM,MAAM,UAAU,SAAS,MAAM,EAAE,OAAO,UAAU,CAAC;AACzD,MAAI,IAAI,WAAW,EAAG,OAAM,IAAI,MAAM,iCAAiC,OAAO,IAAI,MAAM,CAAC,EAAE;AAC7F;AAOO,IAAM,wBACX;AAEK,SAAS,wBAAwB,MAAuC;AAC7E,QAAM,OAA8B;AAAA,IAClC,QAAQ,CAAC;AAAA,IACT,KAAK;AAAA,IACL,SAAS,CAAC,GAAG,eAAe;AAAA,IAC5B,kBAAkB;AAAA,IAClB,MAAM;AAAA,EACR;AACA,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,MAAM,KAAK,CAAC;AAClB,QAAI,QAAQ,SAAS;AACnB,WAAK,MAAM,KAAK,EAAE,CAAC,KAAK;AAAA,IAC1B,WAAW,QAAQ,aAAa;AAC9B,YAAM,OAAO,KAAK,EAAE,CAAC,KAAK,IAAI,MAAM,GAAG,EAAE,OAAO,OAAO;AACvD,iBAAW,UAAU,KAAK;AACxB,YAAI,CAAC,gBAAgB,SAAS,MAAuB,GAAG;AACtD,gBAAM,IAAI;AAAA,YACR,mBAAmB,MAAM,4BAAuB,gBAAgB,KAAK,GAAG,CAAC;AAAA,UAC3E;AAAA,QACF;AAAA,MACF;AACA,WAAK,UAAU;AAAA,IACjB,WAAW,QAAQ,uBAAuB;AACxC,WAAK,mBAAmB;AAAA,IAC1B,WAAW,QAAQ,UAAU;AAC3B,WAAK,OAAO,KAAK,EAAE,CAAC,KAAK;AAAA,IAC3B,WAAW,IAAI,WAAW,IAAI,GAAG;AAC/B,YAAM,IAAI,MAAM,iBAAiB,GAAG,GAAG;AAAA,IACzC,OAAO;AACL,WAAK,OAAO,KAAK,GAAG;AAAA,IACtB;AAAA,EACF;AACA,MAAI,KAAK,OAAO,WAAW,KAAK,CAAC,KAAK,KAAK;AACzC,UAAM,IAAI,MAAM,qBAAqB;AAAA,EACvC;AACA,MAAI,KAAK,SAAS,QAAQ,CAAC,qBAAqB,KAAK,KAAK,IAAI,GAAG;AAC/D,UAAM,IAAI,MAAM,mCAAmC,KAAK,IAAI,GAAG;AAAA,EACjE;AACA,SAAO;AACT;AAGA,eAAsB,qBAAqB,MAAiC;AAC1E,MAAI;AACJ,MAAI;AACF,WAAO,wBAAwB,IAAI;AAAA,EACrC,SAAS,OAAO;AACd,YAAQ,OAAO,MAAM,GAAG,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK,CAAC;AAAA,CAAI;AAClF,WAAO;AAAA,EACT;AACA,QAAM,UAAU,MAAM,eAAe,KAAK,QAAQ,IAAI;AACtD,UAAQ,OAAO;AAAA,IACb,GAAG,KAAK;AAAA,MACN;AAAA,QACE,MAAM,QAAQ;AAAA,QACd,MAAM,QAAQ;AAAA,QACd,cAAc,QAAQ;AAAA,QACtB,OAAO,QAAQ;AAAA,MACjB;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAAA;AAAA,EACH;AACA,UAAQ,OAAO,MAAM,kBAAa,KAAK,GAAG,KAAK,QAAQ,MAAM,MAAM;AAAA,CAAW;AAC9E,MAAI,KAAK,SAAS,MAAM;AACtB,gBAAY,KAAK,MAAM,KAAK,GAAG;AAC/B,YAAQ,OAAO,MAAM,iBAAY,KAAK,IAAI;AAAA,CAAI;AAAA,EAChD;AACA,SAAO;AACT;","names":["mkdir","writeFile","dirname","mkdir","dirname","writeFile"]}

package/dist/{chunk-K6IAZZ6L.js → chunk-FA3MFYNA.js}

@@ -1,25 +1,22 @@
-import {
-  summarizeBackendIntegrity
-} from "./chunk-ZVCHKKOP.js";
 import {
   paretoChart
-} from "./chunk-DPZAEKA6.js";
+} from "./chunk-IMGJZBFG.js";
+import {
+  pairRunRecords
+} from "./chunk-DPUHNQLN.js";
 import {
-  cohensD,
   continuousAgreement,
   pairedBootstrap,
+  pairedCohensDz,
   pairedMde,
   pairedTTest,
   pearsonR,
-  requiredSampleSize,
+  requiredPairedSampleSize,
   spearmanR
-} from "./chunk-PJQFMIOX.js";
+} from "./chunk-MHELPNRP.js";
 import {
   llmSpans
 } from "./chunk-ZET2UAYW.js";
-import {
-  resolveRunCostProvenance
-} from "./chunk-IILEIWGW.js";
 
 // src/contamination-guard.ts
 function checkCanaries(output, scenarios) {
@@ -146,15 +143,15 @@ async function analyzeRuns(opts) {
     runs,
     histogramBins: bins
   });
-  const knownCostRuns = runs.filter((run) => resolveRunCostProvenance(run).kind !== "uncaptured");
-  const costs = knownCostRuns.map((r) => r.costUsd).filter(Number.isFinite);
+  const knownCostRuns = runs.filter((run) => run.costProvenance.kind !== "uncaptured");
+  const costs = knownCostRuns.map((r) => r.costUsd).filter(isFiniteNumber);
   const costDist = distributionOf(costs, bins);
   const pareto = paretoChart(knownCostRuns, { split });
   const degraded = {};
   if (provenance.uncaptured.n > 0) {
     degraded.cost = diagnoseCostCoverage(runs, provenance);
   } else if (costs.length === 0 || costs.every((c) => c === 0)) {
-    degraded.cost = runs.length > 0 && runs.every((run) => run.costProvenance !== void 0) ? `all ${runs.length} explicitly observed or estimated USD values are $0` : diagnoseZeroCost(runs);
+    degraded.cost = `all ${runs.length} explicitly observed or estimated USD values are $0`;
   }
   if (pareto.points.length < 2) {
     degraded.pareto = pareto.points.length === 0 ? "no candidates \u2014 Pareto unavailable" : "single candidate \u2014 Pareto is a single point, not a frontier";
@@ -169,7 +166,7 @@ async function analyzeRuns(opts) {
   const interRater = opts.raterScores ? computeInterRater(opts.raterScores) : void 0;
   const lift = computeLift(runs, opts.baselineCandidateId, opts.candidateCandidateId, split);
   const failureClusters = opts.analyst ? await computeFailureClusters(runs, opts.analyst, split) : void 0;
-  const failureModes = computeFailureModes(runs);
+  const failureModes = computeFailureModes(runs, split);
   const contamination = opts.canaryScenarios ? computeContamination(runs, opts.canaryScenarios) : void 0;
   const outcomeCorrelation = opts.outcomeSignal ? computeOutcomeCorrelation(runs, opts.outcomeSignal, split) : void 0;
   const release = buildReleaseScorecard(composite, lift, contamination);
@@ -213,16 +210,34 @@ function computeExecutionInsight(runs, bins) {
     (row) => row.costUsd !== void 0 ? [row.costUsd] : []
   );
   const modelCounts = /* @__PURE__ */ new Map();
-  let failureRuns = 0;
-  let reportedErrorEvents = 0;
+  let executionErrorRuns = 0;
+  let executionErrorEvents = 0;
   let errorReportingRuns = 0;
+  let errorSpanEvents = 0;
+  let errorSpanReportingRuns = 0;
+  const terminalOutcomes = {
+    succeeded: 0,
+    failed: 0,
+    cancelled: 0,
+    incomplete: 0,
+    unknown: 0
+  };
+  const errorsByTerminalOutcome = {
+    succeeded: { withErrors: 0, withoutErrors: 0, unreported: 0 },
+    failed: { withErrors: 0, withoutErrors: 0, unreported: 0 },
+    cancelled: { withErrors: 0, withoutErrors: 0, unreported: 0 },
+    incomplete: { withErrors: 0, withoutErrors: 0, unreported: 0 },
+    unknown: { withErrors: 0, withoutErrors: 0, unreported: 0 }
+  };
   let modelCallRuns = 0;
   let modelCallEvents = 0;
   let modelCallReportingRuns = 0;
   for (const run of runs) {
     modelCounts.set(run.model, (modelCounts.get(run.model) ?? 0) + 1);
-    const modelCalls = run.outcome.raw.llm_span_count;
-    if (Number.isFinite(modelCalls)) {
+    const terminalOutcome = run.terminalOutcome;
+    terminalOutcomes[terminalOutcome] += 1;
+    const modelCalls = nonNegativeCountRaw(run, "llm_span_count");
+    if (modelCalls !== void 0) {
       modelCallEvents += modelCalls;
       modelCallReportingRuns += 1;
     }
@@ -230,13 +245,19 @@ function computeExecutionInsight(runs, bins) {
     if ((modelCalls ?? 0) > 0 || usage.input > 0 || usage.output > 0 || (usage.cached ?? 0) > 0 || (usage.cacheWrite ?? 0) > 0) {
       modelCallRuns += 1;
     }
-    const errorEvents = run.outcome.raw.error_span_count;
-    if (Number.isFinite(errorEvents)) {
-      reportedErrorEvents += errorEvents;
+    const errorEvents = reportedExecutionErrorEvents(run);
+    if (errorEvents !== void 0) {
+      executionErrorEvents += errorEvents;
       errorReportingRuns += 1;
-    }
-    if (run.failureClass !== void 0 && run.failureClass !== "success" || run.failureMode !== void 0 || (errorEvents ?? 0) > 0) {
-      failureRuns += 1;
+      if (errorEvents > 0) {
+        executionErrorRuns += 1;
+        errorsByTerminalOutcome[terminalOutcome].withErrors += 1;
+      } else errorsByTerminalOutcome[terminalOutcome].withoutErrors += 1;
+    } else errorsByTerminalOutcome[terminalOutcome].unreported += 1;
+    const reportedErrorSpans = nonNegativeCountRaw(run, "error_span_count");
+    if (reportedErrorSpans !== void 0) {
+      errorSpanEvents += reportedErrorSpans;
+      errorSpanReportingRuns += 1;
     }
   }
   return {
@@ -267,14 +288,28 @@ function computeExecutionInsight(runs, bins) {
       events: modelCallEvents,
       reportingRuns: modelCallReportingRuns
     },
-    failures: {
-      runs: failureRuns,
-      fraction: runs.length > 0 ? failureRuns / runs.length : 0,
-      reportedErrorEvents,
-      reportingRuns: errorReportingRuns
-    }
+    executionErrors: {
+      runs: executionErrorRuns,
+      fraction: errorReportingRuns > 0 ? executionErrorRuns / errorReportingRuns : null,
+      events: executionErrorEvents,
+      reportingRuns: errorReportingRuns,
+      errorSpanEvents,
+      errorSpanReportingRuns,
+      byTerminalOutcome: errorsByTerminalOutcome
+    },
+    terminalOutcomes
   };
 }
+function reportedExecutionErrorEvents(run) {
+  const canonical = nonNegativeCountRaw(run, "execution_error_count");
+  if (canonical !== void 0) return canonical;
+  const toolErrors = nonNegativeCountRaw(run, "tool_errors");
+  return toolErrors;
+}
+function nonNegativeCountRaw(run, key) {
+  const value = finiteRaw(run, key);
+  return value !== void 0 && Number.isInteger(value) && value >= 0 ? value : void 0;
+}
 function summarizeTokenUsage(usages, bins) {
   const reasoning = usages.flatMap(
     (usage) => usage.reasoning !== void 0 ? [usage.reasoning] : []
@@ -332,7 +367,7 @@ function summarizeCostProvenance(runs) {
     knownFraction: 0
   };
   for (const run of runs) {
-    const cost = resolveRunCostProvenance(run);
+    const cost = run.costProvenance;
     if (cost.kind === "uncaptured") {
       summary.uncaptured.n += 1;
     } else {
@@ -347,30 +382,15 @@ function summarizeCostProvenance(runs) {
 function diagnoseCostCoverage(runs, provenance) {
   const uncaptured = provenance.uncaptured.n;
   const known = provenance.observed.n + provenance.estimated.n;
-  const explicitUncaptured = runs.some((run) => run.costProvenance?.kind === "uncaptured");
-  if (uncaptured === runs.length && !explicitUncaptured) return diagnoseZeroCost(runs);
   if (uncaptured === runs.length) {
     return `USD cost uncaptured for all ${runs.length} runs \u2014 no observed or estimated USD values; token and wall-time metrics remain available.`;
   }
   return `USD cost uncaptured for ${uncaptured}/${runs.length} runs; excluded those rows from cost statistics (${known}/${runs.length} retained: ${provenance.observed.n} observed, ${provenance.estimated.n} estimated).`;
 }
-function diagnoseZeroCost(runs) {
-  const integrity = summarizeBackendIntegrity(runs);
-  const { totalRecords, stubRecords, uncostedRecords } = integrity;
-  if (totalRecords > 0 && stubRecords === totalRecords) {
-    return `no costUsd values recorded \u2014 all ${totalRecords} records are stub-mode (zero token usage). The backend never reported real LLM activity, so cost cannot be computed; verify the backend actually ran before trusting this corpus.`;
-  }
-  if (uncostedRecords > 0) {
-    return `no costUsd values recorded \u2014 ${uncostedRecords}/${totalRecords} records have token usage but $0 cost (unpriced model). Check isModelPriced(model) for the run's model id and add it to FAMILY_PRICING.`;
-  }
-  if (stubRecords > 0) {
-    return `no costUsd values recorded \u2014 ${stubRecords}/${totalRecords} records are stub-mode (zero token usage); the remainder reported neither tokens nor cost. Cost axis carries no signal.`;
-  }
-  return "no costUsd values recorded \u2014 cost axis carries no signal";
-}
-function computeFailureModes(runs) {
+function computeFailureModes(runs, split) {
   const counts = /* @__PURE__ */ new Map();
   for (const r of runs) {
+    if (!isTaskFailure(r, split)) continue;
     const key = r.failureClass ?? r.failureMode;
     if (key) counts.set(key, (counts.get(key) ?? 0) + 1);
   }
@@ -434,7 +454,10 @@ function computePriorPeriodComparison(current, baseline, split, windowLabel) {
   };
 }
 function knownCostValues(runs) {
-  return runs.filter((run) => resolveRunCostProvenance(run).kind !== "uncaptured").map((run) => run.costUsd).filter(Number.isFinite);
+  return runs.filter((run) => run.costProvenance.kind !== "uncaptured").map((run) => run.costUsd).filter(isFiniteNumber);
+}
+function isFiniteNumber(value) {
+  return typeof value === "number" && Number.isFinite(value);
 }
 function collectPerDimension(runs) {
   const out = {};
@@ -465,15 +488,15 @@ function welchCompare(baseline, current) {
   const pooledStddev = Math.sqrt(
     ((baselineN - 1) * baselineVar + (currentN - 1) * currentVar) / Math.max(1, baselineN + currentN - 2)
   );
-  const cohensD2 = pooledStddev > 0 ? delta / pooledStddev : 0;
-  const significant = pValue < 0.05 && Math.abs(cohensD2) >= 0.2;
+  const cohensD = pooledStddev > 0 ? delta / pooledStddev : 0;
+  const significant = pValue < 0.05 && Math.abs(cohensD) >= 0.2;
   return {
     current: currentMean,
     baseline: baselineMean,
     delta,
     ci95,
     pValue,
-    cohensD: cohensD2,
+    cohensD,
     baselineN,
     currentN,
     significant
@@ -504,21 +527,19 @@ function resolveSplit(runs, pref) {
   return hasHoldout ? "holdout" : "search";
 }
 function compositeOf(run, split) {
-  const primary = split === "holdout" ? run.outcome.holdoutScore : run.outcome.searchScore;
-  if (Number.isFinite(primary)) return primary;
-  const alt = split === "holdout" ? run.outcome.searchScore : run.outcome.holdoutScore;
-  return Number.isFinite(alt) ? alt : Number.NaN;
+  const score = split === "holdout" ? run.outcome.holdoutScore : run.outcome.searchScore;
+  return Number.isFinite(score) ? score : Number.NaN;
 }
 function distributionOf(values, bins, withIds) {
   if (values.length === 0) {
     return {
       n: 0,
-      mean: 0,
-      p50: 0,
-      p95: 0,
-      stddev: 0,
-      min: 0,
-      max: 0,
+      mean: null,
+      p50: null,
+      p95: null,
+      stddev: null,
+      min: null,
+      max: null,
       histogram: []
     };
   }
@@ -677,41 +698,28 @@ function computeLift(runs, baselineId, candidateId, split) {
     const ids = [...new Set(runs.map((r) => r.candidateId))];
     if (ids.length !== 2) return void 0;
     const [idA, idB] = ids;
-    const meanA = mean(runs.filter((r) => r.candidateId === idA).map((r) => compositeOf(r, split)));
-    const meanB = mean(runs.filter((r) => r.candidateId === idB).map((r) => compositeOf(r, split)));
+    const scoresA = finiteCompositeScores(
+      runs.filter((run) => run.candidateId === idA),
+      split
+    );
+    const scoresB = finiteCompositeScores(
+      runs.filter((run) => run.candidateId === idB),
+      split
+    );
+    if (scoresA.length === 0 || scoresB.length === 0) return void 0;
+    const meanA = mean(scoresA);
+    const meanB = mean(scoresB);
     bId = meanA <= meanB ? idA : idB;
     cId = meanA <= meanB ? idB : idA;
   }
   const baseline = runs.filter((r) => r.candidateId === bId);
   const candidate = runs.filter((r) => r.candidateId === cId);
   if (baseline.length === 0 || candidate.length === 0) return void 0;
-  const baselineByKey = new Map(baseline.map((r) => [pairingKey(r), r]));
-  const pairedBaseline = [];
-  const pairedCandidate = [];
-  let usedKeyPairing = false;
-  for (const cand of candidate) {
-    const b = baselineByKey.get(pairingKey(cand));
-    if (b) {
-      const bC = compositeOf(b, split);
-      const cC = compositeOf(cand, split);
-      if (Number.isFinite(bC) && Number.isFinite(cC)) {
-        pairedBaseline.push(bC);
-        pairedCandidate.push(cC);
-        usedKeyPairing = true;
-      }
-    }
-  }
-  if (!usedKeyPairing) {
-    const n = Math.min(baseline.length, candidate.length);
-    for (let i = 0; i < n; i++) {
-      const bC = compositeOf(baseline[i], split);
-      const cC = compositeOf(candidate[i], split);
-      if (Number.isFinite(bC) && Number.isFinite(cC)) {
-        pairedBaseline.push(bC);
-        pairedCandidate.push(cC);
-      }
-    }
-  }
+  const scoredBaseline = baseline.filter((run) => Number.isFinite(compositeOf(run, split)));
+  const scoredCandidate = candidate.filter((run) => Number.isFinite(compositeOf(run, split)));
+  const pairing = pairRunRecords(scoredBaseline, scoredCandidate);
+  const pairedBaseline = pairing.pairs.map((pair) => compositeOf(pair.baseline, split));
+  const pairedCandidate = pairing.pairs.map((pair) => compositeOf(pair.treatment, split));
   if (pairedBaseline.length === 0) return void 0;
   const baselineMean = mean(pairedBaseline);
   const candidateMean = mean(pairedCandidate);
@@ -722,10 +730,10 @@ function computeLift(runs, baselineId, candidateId, split) {
     statistic: "mean"
   });
   const tTest = pairedTTest(pairedBaseline, pairedCandidate);
-  const d = cohensD(pairedBaseline, pairedCandidate);
+  const d = pairedCohensDz(pairedBaseline, pairedCandidate);
   const mde = pairedMde({ nPaired: pairedBaseline.length, power: 0.8, alpha: 0.05 });
-  const requiredN = requiredSampleSize({
-    effect: Math.max(Math.abs(delta), 1e-6),
+  const requiredN = d === null || d === 0 ? null : requiredPairedSampleSize({
+    effect: Math.abs(d),
     power: 0.8,
     alpha: 0.05
   });
@@ -736,19 +744,18 @@ function computeLift(runs, baselineId, candidateId, split) {
     ci95: [bootstrap.low, bootstrap.high],
     pValue: tTest.p,
     n: pairedBaseline.length,
+    unpairedBaseline: pairing.unpairedBaseline.length,
+    unpairedCandidate: pairing.unpairedTreatment.length,
     cohensD: d,
     mde,
     requiredN
   };
 }
-function pairingKey(r) {
-  return `${r.experimentId}::${r.seed}`;
-}
 function mean(arr) {
   return arr.length === 0 ? 0 : arr.reduce((s, v) => s + v, 0) / arr.length;
 }
 async function computeFailureClusters(runs, analyst, split) {
-  const failed = runs.filter((r) => compositeOf(r, split) < 0.5 || r.failureMode !== void 0);
+  const failed = runs.filter((run) => isTaskFailure(run, split));
   if (failed.length === 0) return { clusters: [], totalFailures: 0 };
   const clusters = /* @__PURE__ */ new Map();
   for (const run of failed) {
@@ -775,6 +782,13 @@ async function computeFailureClusters(runs, analyst, split) {
   clusterList.sort((a, b) => b.share - a.share);
   return { clusters: clusterList, totalFailures: failed.length };
 }
+function finiteCompositeScores(runs, split) {
+  return runs.map((run) => compositeOf(run, split)).filter(Number.isFinite);
+}
+function isTaskFailure(run, split) {
+  const score = compositeOf(run, split);
+  return Number.isFinite(score) && score < 0.5;
+}
 function computeContamination(runs, canaries) {
   let leaks = 0;
   const details = [];
@@ -832,24 +846,30 @@ function computeOutcomeCorrelation(runs, outcome, split) {
 }
 function buildReleaseScorecard(composite, lift, contamination) {
   const axes = [];
-  const liftPass = lift === void 0 || lift.ci95[0] > 0 ? "pass" : lift.delta > 0 ? "warn" : "fail";
+  const liftPass = lift === void 0 ? "not_evaluated" : lift.ci95[0] > 0 ? "pass" : lift.delta > 0 ? "warn" : "fail";
   axes.push({
     name: "quality-lift",
     status: liftPass,
     detail: lift ? `delta=${lift.delta.toFixed(3)}, CI95=[${lift.ci95[0].toFixed(3)}, ${lift.ci95[1].toFixed(3)}], n=${lift.n}` : "no baseline/candidate pair available"
   });
-  const contamPass = contamination === void 0 || contamination.leaks === 0 ? "pass" : "fail";
+  const contamPass = contamination === void 0 ? "not_evaluated" : contamination.leaks === 0 ? "pass" : "fail";
   axes.push({
     name: "contamination",
     status: contamPass,
     detail: contamination ? `${contamination.leaks} canary leak(s)` : "no canaries supplied"
   });
-  axes.push({
-    name: "composite-distribution",
-    status: composite.mean >= 0.5 ? "pass" : composite.mean >= 0.3 ? "warn" : "fail",
-    detail: `mean=${composite.mean.toFixed(3)}, p50=${composite.p50.toFixed(3)}, p95=${composite.p95.toFixed(3)} over n=${composite.n}`
-  });
-  const status = axes.some((a) => a.status === "fail") ? "fail" : axes.some((a) => a.status === "warn") ? "warn" : "pass";
+  axes.push(
+    composite.n === 0 ? {
+      name: "composite-distribution",
+      status: "not_evaluated",
+      detail: "no task-quality scores available"
+    } : {
+      name: "composite-distribution",
+      status: composite.mean !== null && composite.mean >= 0.5 ? "pass" : composite.mean !== null && composite.mean >= 0.3 ? "warn" : "fail",
+      detail: composite.mean === null || composite.p50 === null || composite.p95 === null ? "task-quality distribution is internally incomplete" : `mean=${composite.mean.toFixed(3)}, p50=${composite.p50.toFixed(3)}, p95=${composite.p95.toFixed(3)} over n=${composite.n}`
+    }
+  );
+  const status = axes.some((a) => a.status === "fail") ? "fail" : axes.some((a) => a.status === "warn" || a.status === "not_evaluated") ? "warn" : "pass";
   return {
     status,
     axes,
@@ -884,7 +904,7 @@ function buildRecommendations(ctx) {
       });
     }
   }
-  if (ctx.composite.n > 0) {
+  if (ctx.composite.n > 0 && ctx.composite.mean !== null && ctx.composite.p50 !== null && ctx.composite.p95 !== null) {
     if (ctx.composite.mean < 0.3) {
       const tail = ctx.composite.tailRuns ?? [];
       const names = tail.slice(0, 5).map((t) => `${t.runId}=${t.score.toFixed(3)}`).join(", ");
@@ -929,6 +949,8 @@ function buildRecommendations(ctx) {
     });
   }
   if (ctx.lift) {
+    const pairedEffect = ctx.lift.cohensD === null ? "undefined (zero delta variance)" : ctx.lift.cohensD.toFixed(2);
+    const requiredRuns = ctx.lift.requiredN === null ? "not estimable" : `~${ctx.lift.requiredN} paired runs`;
     const decisive = ctx.lift.ci95[0] > ctx.threshold;
     const inconclusive = ctx.lift.ci95[0] <= ctx.threshold && ctx.lift.ci95[1] > ctx.threshold;
     if (decisive) {
@@ -936,14 +958,14 @@ function buildRecommendations(ctx) {
         priority: "critical",
         kind: "ship",
         title: `Ship \u2014 lift ${ctx.lift.delta.toFixed(3)} (95% CI ${ctx.lift.ci95[0].toFixed(3)}..${ctx.lift.ci95[1].toFixed(3)})`,
-        detail: `Holdout lift exceeds threshold ${ctx.threshold} with 95% bootstrap confidence (n=${ctx.lift.n}, p=${ctx.lift.pValue.toFixed(4)}, d=${ctx.lift.cohensD.toFixed(2)}).`,
+        detail: `Holdout lift exceeds threshold ${ctx.threshold} with 95% bootstrap confidence (n=${ctx.lift.n}, p=${ctx.lift.pValue.toFixed(4)}, paired d=${pairedEffect}).`,
         evidencePath: "lift"
       });
     } else if (inconclusive) {
       out.push({
         priority: "high",
         kind: "expand-corpus",
-        title: `Inconclusive \u2014 need ~${ctx.lift.requiredN} paired runs (have ${ctx.lift.n}) at current effect size`,
+        title: `Inconclusive \u2014 required sample is ${requiredRuns} (have ${ctx.lift.n}) at current effect size`,
         detail: `CI straddles threshold. Current MDE at 80% power is ${ctx.lift.mde.toFixed(3)}; observed delta is ${ctx.lift.delta.toFixed(3)}.`,
         evidencePath: "lift"
       });
@@ -1006,4 +1028,4 @@ export {
   summarizeExecution,
   analyzeRuns
 };
-//# sourceMappingURL=chunk-K6IAZZ6L.js.map
+//# sourceMappingURL=chunk-FA3MFYNA.js.map