@eidentic/bench 0.1.0 → 0.1.1

package/dist/index.cjs

@@ -5,6 +5,9 @@ var __getOwnPropDesc = Object.getOwnPropertyDescriptor;
 var __getOwnPropNames = Object.getOwnPropertyNames;
 var __getProtoOf = Object.getPrototypeOf;
 var __hasOwnProp = Object.prototype.hasOwnProperty;
+var __esm = (fn, res) => function __init() {
+  return fn && (res = (0, fn[__getOwnPropNames(fn)[0]])(fn = 0)), res;
+};
 var __export = (target, all) => {
   for (var name in all)
     __defProp(target, name, { get: all[name], enumerable: true });
@@ -27,17 +30,128 @@ var __toESM = (mod, isNodeMode, target) => (target = mod != null ? __create(__ge
 ));
 var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: true }), mod);
 
+// src/locomo-loader.ts
+var locomo_loader_exports = {};
+__export(locomo_loader_exports, {
+  LOCOMO_SOURCE_SHA: () => LOCOMO_SOURCE_SHA,
+  loadLoCoMo: () => loadLoCoMo2
+});
+async function assertFileSize2(filePath, maxBytes) {
+  let fileSize;
+  try {
+    const s = await (0, import_promises2.stat)(filePath);
+    fileSize = s.size;
+  } catch (err) {
+    throw new Error(
+      `bench loader: cannot stat file "${filePath}": ${err.message}`
+    );
+  }
+  if (fileSize > maxBytes) {
+    const mb = (fileSize / (1024 * 1024)).toFixed(1);
+    const capMb = (maxBytes / (1024 * 1024)).toFixed(0);
+    throw new Error(
+      `bench loader: file "${filePath}" is ${mb} MiB, which exceeds the ${capMb} MiB cap. Pass a larger maxBytes option if this is intentional.`
+    );
+  }
+}
+function parseLocomoDateTime(raw) {
+  const cleaned = raw.replace(/\s*on\s+/, " ").trim();
+  const ms = Date.parse(cleaned);
+  if (Number.isNaN(ms)) {
+    const match = /^(\d{1,2}:\d{2}\s*(?:am|pm))\s+(.+)$/i.exec(cleaned);
+    if (match) {
+      const ms2 = Date.parse(`${match[2]} ${match[1]}`);
+      return Number.isNaN(ms2) ? 0 : ms2;
+    }
+    return 0;
+  }
+  return ms;
+}
+function parseSessions(conv) {
+  const indices = [];
+  for (const key of Object.keys(conv)) {
+    const m = /^session_(\d+)$/.exec(key);
+    if (m) indices.push(parseInt(m[1], 10));
+  }
+  indices.sort((a, b) => a - b);
+  const sessions = [];
+  for (const idx of indices) {
+    const turnsRaw = conv[`session_${idx}`];
+    if (!Array.isArray(turnsRaw)) continue;
+    const dateTimeRaw = typeof conv[`session_${idx}_date_time`] === "string" ? conv[`session_${idx}_date_time`] : "";
+    const turns = turnsRaw.map((t) => ({
+      speaker: t.speaker ?? "",
+      diaId: t.dia_id ?? "",
+      text: t.text ?? ""
+    }));
+    sessions.push({
+      index: idx,
+      dateTime: dateTimeRaw,
+      dateTimeMs: dateTimeRaw ? parseLocomoDateTime(dateTimeRaw) : 0,
+      turns
+    });
+  }
+  return sessions;
+}
+async function loadLoCoMo2(jsonPath, opts) {
+  await assertFileSize2(jsonPath, opts?.maxBytes ?? DEFAULT_MAX_BYTES2);
+  const raw = JSON.parse(await (0, import_promises2.readFile)(jsonPath, "utf-8"));
+  if (!Array.isArray(raw)) {
+    throw new Error(
+      `bench loader: expected the LoCoMo JSON root to be an array, but got ${typeof raw}. Did you pass the correct file?`
+    );
+  }
+  const rawSamples = raw;
+  const samples = rawSamples.map((s, i) => {
+    const sampleId = s.sample_id ?? String(i);
+    const conv = s.conversation ?? {};
+    const sessions = parseSessions(conv);
+    const qa = (s.qa ?? []).map((q) => {
+      const answer = q.answer !== void 0 ? String(q.answer) : void 0;
+      return {
+        question: q.question ?? "",
+        answer,
+        category: q.category,
+        evidence: Array.isArray(q.evidence) ? q.evidence : [],
+        adversarialAnswer: q.adversarial_answer
+      };
+    });
+    return {
+      sampleId,
+      speakerA: String(conv.speaker_a ?? ""),
+      speakerB: String(conv.speaker_b ?? ""),
+      sessions,
+      qa
+    };
+  });
+  return { samples };
+}
+var import_promises2, LOCOMO_SOURCE_SHA, DEFAULT_MAX_BYTES2;
+var init_locomo_loader = __esm({
+  "src/locomo-loader.ts"() {
+    "use strict";
+    import_promises2 = require("node:fs/promises");
+    LOCOMO_SOURCE_SHA = "3eb6f2c585f5e1699204e3c3bdf7adc5c28cb376";
+    DEFAULT_MAX_BYTES2 = 256 * 1024 * 1024;
+  }
+});
+
 // src/index.ts
 var index_exports = {};
 __export(index_exports, {
   CONTRADICTION_FIXTURES: () => CONTRADICTION_FIXTURES,
   JUNK_STREAM_FIXTURES: () => JUNK_STREAM_FIXTURES,
+  LOCOMO_SOURCE_SHA: () => LOCOMO_SOURCE_SHA,
   factRecall: () => factRecall,
-  loadLoCoMo: () => loadLoCoMo,
+  loadLoCoMo: () => loadLoCoMo2,
+  loadLoCoMoLegacy: () => loadLoCoMo,
   loadLongMemEval: () => loadLongMemEval,
   normalizeText: () => normalizeText,
   normalizedIncludes: () => normalizedIncludes,
   recallAtK: () => recallAtK,
+  renderLocomoReportMarkdown: () => renderLocomoReportMarkdown,
+  resolveEvidence: () => resolveEvidence,
+  runLocomoBench: () => runLocomoBench,
   runMemoryBench: () => runMemoryBench,
   runTemporalBench: () => runTemporalBench,
   runWriteQualityBench: () => runWriteQualityBench,
@@ -445,6 +559,541 @@ async function loadLoCoMo(jsonPath, opts) {
   return { name: "LoCoMo", cases: benchCases };
 }
 
+// src/index.ts
+init_locomo_loader();
+
+// src/locomo-run.ts
+var import_promises3 = require("node:fs/promises");
+var import_node_fs = require("node:fs");
+init_locomo_loader();
+function makeRng(seed) {
+  let s = seed >>> 0;
+  if (s === 0) s = 1;
+  return () => {
+    s ^= s << 13;
+    s ^= s >>> 17;
+    s ^= s << 5;
+    s = s >>> 0;
+    return s / 4294967296;
+  };
+}
+function seededShuffle(arr, rng) {
+  const out = [...arr];
+  for (let i = out.length - 1; i > 0; i--) {
+    const j = Math.floor(rng() * (i + 1));
+    [out[i], out[j]] = [out[j], out[i]];
+  }
+  return out;
+}
+var ANSWER_INSTRUCTION = "Answer the question concisely and specifically based only on the information provided. If the provided context contains no answer to the question, reply exactly: No information available";
+function buildMemoryAnswerPrompt(snippets, question) {
+  const context = snippets.length > 0 ? snippets.map((s, i) => `[${i + 1}] ${s}`).join("\n\n") : "(no relevant context retrieved)";
+  return `Context from conversation:
+${context}
+
+Question: ${question}
+
+${ANSWER_INSTRUCTION}`;
+}
+function buildFullContextAnswerPrompt(conversationText, question) {
+  return `Conversation:
+${conversationText}
+
+Question: ${question}
+
+${ANSWER_INSTRUCTION}`;
+}
+function buildJudgePrompt(question, goldAnswer, modelAnswer, adversarialAnswer, isCategory5) {
+  if (isCategory5) {
+    const trapNote = adversarialAnswer ? `
+WARNING \u2014 trap answer to watch for: "${adversarialAnswer}". If the model answer closely matches this trap, mark incorrect.` : "";
+    return `You are a strict judge. For this unanswerable question, a CORRECT response means the model refused to answer or said something equivalent to "No information available" or "I don't know".${trapNote}
+
+Question: ${question}
+Model answer: ${modelAnswer}
+
+Respond with JSON only: {"correct": true} if the model declined (correct), {"correct": false} if it made a claim.`;
+  }
+  return `You are a strict judge evaluating whether a model answer is correct.
+CORRECT: the model answer contains the specific information in the gold answer (exact match or clear paraphrase).
+WRONG: the model answer is vague, only topically related, or contradicts the gold answer.
+For date/time questions: equivalent date expressions referring to the same date are correct.
+
+Question: ${question}
+Gold answer: ${goldAnswer ?? "(none)"}
+Model answer: ${modelAnswer}
+
+Respond with JSON only: {"correct": true} or {"correct": false}`;
+}
+function renderConversation(sample) {
+  const lines = [];
+  for (const sess of sample.sessions) {
+    const dateLabel = sess.dateTime ? ` \u2014 ${sess.dateTime}` : "";
+    lines.push(`Session ${sess.index}${dateLabel}`);
+    for (const turn of sess.turns) {
+      lines.push(`[${turn.speaker}]: ${turn.text}`);
+    }
+    lines.push("");
+  }
+  return lines.join("\n").trim();
+}
+async function ingestSampleIntoMemory(sample, memory, scope) {
+  const events = [];
+  for (const sess of sample.sessions) {
+    const dateLabel = sess.dateTime ? ` \u2014 ${sess.dateTime}` : "";
+    const header = `Session ${sess.index}${dateLabel}`;
+    for (const turn of sess.turns) {
+      events.push({
+        id: `${sample.sampleId}:turn:${turn.diaId}`,
+        scope,
+        text: `[${header}] [${turn.speaker}]: ${turn.text}`,
+        metadata: {
+          diaId: turn.diaId,
+          sessionIndex: sess.index,
+          ingestedAt: sess.dateTimeMs || void 0
+        }
+      });
+    }
+    const sessionText = [
+      header,
+      ...sess.turns.map((t) => `[${t.speaker}]: ${t.text}`)
+    ].join("\n");
+    events.push({
+      id: `${sample.sampleId}:sess:${sess.index}:chunk`,
+      scope,
+      text: sessionText,
+      metadata: { sessionIndex: sess.index, ingestedAt: sess.dateTimeMs || void 0 }
+    });
+  }
+  await memory.ingest(events);
+}
+async function callJudge(judgeModel, prompt) {
+  const response = await judgeModel.complete({
+    messages: [{ role: "user", content: prompt }],
+    tools: [],
+    outputSchema: {
+      type: "object",
+      properties: { correct: { type: "boolean" } },
+      required: ["correct"],
+      // OpenAI strict structured-output mode requires this to be explicit.
+      additionalProperties: false
+    }
+  });
+  let correct = false;
+  if (response.object && typeof response.object.correct === "boolean") {
+    correct = response.object.correct;
+  } else {
+    const text = response.content.filter((b) => b.type === "text").map((b) => b.text ?? "").join("").toLowerCase().trim();
+    if (/"correct"\s*:\s*true/i.test(text)) correct = true;
+    else if (/"correct"\s*:\s*false/i.test(text)) correct = false;
+    else correct = text.includes("true");
+  }
+  return {
+    correct,
+    inputTokens: response.usage?.inputTokens ?? 0,
+    outputTokens: response.usage?.outputTokens ?? 0
+  };
+}
+async function loadCheckpoint(path) {
+  const done = /* @__PURE__ */ new Set();
+  if (!(0, import_node_fs.existsSync)(path)) return done;
+  const raw = await (0, import_promises3.readFile)(path, "utf-8");
+  for (const line of raw.split("\n")) {
+    const trimmed = line.trim();
+    if (!trimmed) continue;
+    try {
+      const row = JSON.parse(trimmed);
+      if (row.sampleId && row.questionIndex !== void 0) {
+        done.add(`${row.sampleId}:${row.questionIndex}`);
+      }
+    } catch {
+    }
+  }
+  return done;
+}
+async function appendCheckpointRow(path, row) {
+  await (0, import_promises3.appendFile)(path, JSON.stringify(row) + "\n", "utf-8");
+}
+var DECLINE_PATTERNS = /\bno information available\b|\bi (don'?t|do not) know\b|\bcannot (find|answer|determine)\b|\bunavailable\b/i;
+function appearsToDecline(answer) {
+  return DECLINE_PATTERNS.test(answer);
+}
+async function runLocomoBench(opts) {
+  const {
+    answerModel,
+    judgeModel,
+    mode,
+    categories = [1, 2, 3, 4, 5],
+    sampleLimit,
+    questionLimit,
+    seed = 42,
+    topK: rawTopK = 10,
+    concurrency = 1,
+    onProgress,
+    checkpointPath
+  } = opts;
+  const topK = Math.min(rawTopK, 10);
+  const rng = makeRng(seed);
+  let dataset;
+  if (opts.dataset) {
+    dataset = opts.dataset;
+  } else {
+    const { loadLoCoMo: loadLoCoMo3 } = await Promise.resolve().then(() => (init_locomo_loader(), locomo_loader_exports));
+    dataset = await loadLoCoMo3(opts.dataPath);
+  }
+  let samples = [...dataset.samples];
+  if (sampleLimit !== void 0 && sampleLimit < samples.length) {
+    samples = seededShuffle(samples, makeRng(seed)).slice(0, sampleLimit);
+  }
+  const queue = [];
+  for (const sample of samples) {
+    let qaList = sample.qa.filter((q) => categories.includes(q.category));
+    if (questionLimit !== void 0 && questionLimit < qaList.length) {
+      qaList = seededShuffle(qaList, makeRng(seed + 1)).slice(0, questionLimit);
+    }
+    for (const qa of qaList) {
+      queue.push({ sample, qaIndex: sample.qa.indexOf(qa), qa });
+    }
+  }
+  const checkpoint = checkpointPath ? await loadCheckpoint(checkpointPath) : /* @__PURE__ */ new Set();
+  const conversationText = /* @__PURE__ */ new Map();
+  if (mode === "full-context") {
+    for (const sample of samples) {
+      conversationText.set(sample.sampleId, renderConversation(sample));
+    }
+  }
+  const memories = /* @__PURE__ */ new Map();
+  let ingestInputTokens = 0;
+  let ingestOutputTokens = 0;
+  if (mode === "memory") {
+    if (!opts.memoryFactory) {
+      throw new Error("runLocomoBench: memoryFactory is required when mode='memory'");
+    }
+    for (const sample of samples) {
+      const memory = await opts.memoryFactory(sample.sampleId);
+      const scope = { kind: "agent", agentId: `locomo:${sample.sampleId}` };
+      await ingestSampleIntoMemory(sample, memory, scope);
+      memories.set(sample.sampleId, memory);
+    }
+  }
+  const results = [];
+  let totalAnswerInputTokens = 0;
+  let totalAnswerOutputTokens = 0;
+  let totalJudgeInputTokens = 0;
+  let totalJudgeOutputTokens = 0;
+  let errorCount = 0;
+  let done = 0;
+  const total = queue.length - checkpoint.size;
+  const startTime = Date.now();
+  const processItem = async (item) => {
+    const key = `${item.sample.sampleId}:${item.qaIndex}`;
+    if (checkpoint.has(key)) return;
+    let modelAnswer = "";
+    let answerIn = 0;
+    let answerOut = 0;
+    let judgeIn = 0;
+    let judgeOut = 0;
+    let correct = false;
+    let trapTriggered;
+    let errorMsg;
+    try {
+      if (mode === "memory") {
+        const memory = memories.get(item.sample.sampleId);
+        const scope = { kind: "agent", agentId: `locomo:${item.sample.sampleId}` };
+        const retrieved = await memory.retrieve({ text: item.qa.question, scope, topK });
+        const snippets = retrieved.snippets.map((s) => s.text);
+        const prompt = buildMemoryAnswerPrompt(snippets, item.qa.question);
+        const resp = await answerModel.complete({
+          messages: [{ role: "user", content: prompt }],
+          tools: []
+        });
+        const textBlocks = resp.content.filter((b) => b.type === "text");
+        modelAnswer = textBlocks.map((b) => b.text).join("").trim();
+        answerIn = resp.usage?.inputTokens ?? 0;
+        answerOut = resp.usage?.outputTokens ?? 0;
+      } else {
+        const convText = conversationText.get(item.sample.sampleId) ?? "";
+        const prompt = buildFullContextAnswerPrompt(convText, item.qa.question);
+        const resp = await answerModel.complete({
+          messages: [{ role: "user", content: prompt }],
+          tools: []
+        });
+        const textBlocks = resp.content.filter((b) => b.type === "text");
+        modelAnswer = textBlocks.map((b) => b.text).join("").trim();
+        answerIn = resp.usage?.inputTokens ?? 0;
+        answerOut = resp.usage?.outputTokens ?? 0;
+      }
+      const isCategory5 = item.qa.category === 5;
+      const judgePrompt = buildJudgePrompt(
+        item.qa.question,
+        item.qa.answer,
+        modelAnswer,
+        item.qa.adversarialAnswer,
+        isCategory5
+      );
+      const judgeResult = await callJudge(judgeModel, judgePrompt);
+      correct = judgeResult.correct;
+      judgeIn = judgeResult.inputTokens;
+      judgeOut = judgeResult.outputTokens;
+      if (isCategory5 && item.qa.adversarialAnswer) {
+        const adversarialLower = item.qa.adversarialAnswer.toLowerCase();
+        const answerLower = modelAnswer.toLowerCase();
+        trapTriggered = answerLower.includes(adversarialLower) || adversarialLower.includes(answerLower.slice(0, Math.min(answerLower.length, 30)));
+        if (trapTriggered && correct) correct = false;
+        if (!correct && appearsToDecline(modelAnswer)) {
+          correct = true;
+          trapTriggered = false;
+        }
+      }
+    } catch (err) {
+      errorMsg = err.message;
+      errorCount++;
+      correct = false;
+    }
+    const row = {
+      sampleId: item.sample.sampleId,
+      questionIndex: item.qaIndex,
+      question: item.qa.question,
+      goldAnswer: item.qa.answer,
+      category: item.qa.category,
+      modelAnswer,
+      correct,
+      ...trapTriggered !== void 0 ? { trapTriggered } : {},
+      ...errorMsg !== void 0 ? { error: errorMsg } : {},
+      answerInputTokens: answerIn,
+      answerOutputTokens: answerOut,
+      judgeInputTokens: judgeIn,
+      judgeOutputTokens: judgeOut
+    };
+    results.push(row);
+    totalAnswerInputTokens += answerIn;
+    totalAnswerOutputTokens += answerOut;
+    totalJudgeInputTokens += judgeIn;
+    totalJudgeOutputTokens += judgeOut;
+    if (checkpointPath) {
+      await appendCheckpointRow(checkpointPath, row);
+    }
+    done++;
+    if (onProgress) onProgress(done, total);
+  };
+  const concurrencyLimit = Math.max(1, concurrency);
+  const pending = [];
+  for (const item of queue) {
+    const p = processItem(item);
+    pending.push(p);
+    if (pending.length >= concurrencyLimit) {
+      await Promise.all(pending.splice(0, concurrencyLimit));
+    }
+  }
+  if (pending.length > 0) await Promise.all(pending);
+  if (checkpointPath && checkpoint.size > 0) {
+    const raw = await (0, import_promises3.readFile)(checkpointPath, "utf-8").catch(() => "");
+    for (const line of raw.split("\n")) {
+      const trimmed = line.trim();
+      if (!trimmed) continue;
+      try {
+        const row = JSON.parse(trimmed);
+        const key = `${row.sampleId}:${row.questionIndex}`;
+        if (checkpoint.has(key)) {
+          results.push(row);
+          totalAnswerInputTokens += row.answerInputTokens ?? 0;
+          totalAnswerOutputTokens += row.answerOutputTokens ?? 0;
+          totalJudgeInputTokens += row.judgeInputTokens ?? 0;
+          totalJudgeOutputTokens += row.judgeOutputTokens ?? 0;
+        }
+      } catch {
+      }
+    }
+  }
+  const byCategoryMap = {};
+  for (const row of results) {
+    const cat = String(row.category);
+    if (!byCategoryMap[cat]) byCategoryMap[cat] = { correct: 0, total: 0 };
+    byCategoryMap[cat].total++;
+    if (row.correct) byCategoryMap[cat].correct++;
+  }
+  const byCategory = {};
+  for (const [cat, stats] of Object.entries(byCategoryMap)) {
+    byCategory[cat] = {
+      ...stats,
+      accuracy: stats.total > 0 ? stats.correct / stats.total : 0
+    };
+  }
+  let j14Correct = 0;
+  let j14Total = 0;
+  for (const [cat, stats] of Object.entries(byCategory)) {
+    const n = parseInt(cat);
+    if (n >= 1 && n <= 4) {
+      j14Correct += stats.correct;
+      j14Total += stats.total;
+    }
+  }
+  const cat5Stats = byCategory["5"];
+  const cat5RefusalRate = cat5Stats ? { correct: cat5Stats.correct, total: cat5Stats.total, rate: cat5Stats.accuracy } : void 0;
+  const wallClockMs = Date.now() - startTime;
+  return {
+    config: {
+      mode,
+      topK,
+      answerModelId: answerModel.modelId ?? "(unknown)",
+      judgeModelId: judgeModel.modelId ?? "(unknown)",
+      datasetSha: LOCOMO_SOURCE_SHA,
+      seed,
+      categories: [...categories].sort((a, b) => a - b),
+      samplesRun: samples.length,
+      questionsRun: results.length
+    },
+    overallJ14: {
+      correct: j14Correct,
+      total: j14Total,
+      accuracy: j14Total > 0 ? j14Correct / j14Total : 0
+    },
+    byCategory,
+    cat5RefusalRate,
+    tokens: {
+      ingestInputTokens,
+      ingestOutputTokens,
+      answerInputTokens: totalAnswerInputTokens,
+      answerOutputTokens: totalAnswerOutputTokens,
+      judgeInputTokens: totalJudgeInputTokens,
+      judgeOutputTokens: totalJudgeOutputTokens,
+      totalInputTokens: ingestInputTokens + totalAnswerInputTokens + totalJudgeInputTokens,
+      totalOutputTokens: ingestOutputTokens + totalAnswerOutputTokens + totalJudgeOutputTokens
+    },
+    wallClockMs,
+    questions: results,
+    errorCount
+  };
+}
+
+// src/locomo-render.ts
+function pct(n) {
+  return (n * 100).toFixed(1) + "%";
+}
+function fmtNum(n) {
+  return n.toLocaleString("en-US");
+}
+function estimateCost(tokens, prices) {
+  if (!prices) return "\u2014";
+  const cost = tokens.totalInputTokens / 1e6 * prices.inputPer1M + tokens.totalOutputTokens / 1e6 * prices.outputPer1M;
+  return `$${cost.toFixed(4)}`;
+}
+function renderLocomoReportMarkdown(reports, prices) {
+  const lines = [];
+  lines.push("# LoCoMo Benchmark Results");
+  lines.push("");
+  lines.push("Dataset: [LoCoMo](https://github.com/snap-research/locomo) (Snap Research) \xB7 CC BY-NC 4.0");
+  lines.push("Raw data is not redistributed. Only aggregate results are published here.");
+  lines.push("");
+  if (reports.length === 0) {
+    lines.push("_No results yet._");
+    return lines.join("\n");
+  }
+  const headers = [
+    "System / Mode",
+    "Cat1 (multi-hop)",
+    "Cat2 (temporal)",
+    "Cat3 (open-domain)",
+    "Cat4 (single-hop)",
+    "J(1\u20134) overall",
+    "Cat5 refusal rate",
+    "Tokens/query",
+    "Est. cost/run",
+    "Answer model",
+    "Judge model",
+    "topK",
+    "n-Q",
+    "Seed",
+    "Dataset SHA"
+  ];
+  lines.push("## Results");
+  lines.push("");
+  lines.push("| " + headers.join(" | ") + " |");
+  lines.push("| " + headers.map(() => "---").join(" | ") + " |");
+  for (const r of reports) {
+    const c = r.config;
+    const cat = (n) => {
+      const s = r.byCategory[String(n)];
+      if (!s) return "\u2014";
+      return `${pct(s.accuracy)} (${s.correct}/${s.total})`;
+    };
+    const cat5 = r.cat5RefusalRate ? `${pct(r.cat5RefusalRate.rate)} (${r.cat5RefusalRate.correct}/${r.cat5RefusalRate.total})` : "\u2014";
+    const totalQ = r.questions.length;
+    const tokensPerQuery = totalQ > 0 ? Math.round((r.tokens.totalInputTokens + r.tokens.totalOutputTokens) / totalQ) : 0;
+    const row = [
+      `${c.answerModelId} / ${c.mode}`,
+      cat(1),
+      cat(2),
+      cat(3),
+      cat(4),
+      `${pct(r.overallJ14.accuracy)} (${r.overallJ14.correct}/${r.overallJ14.total})`,
+      cat5,
+      fmtNum(tokensPerQuery),
+      estimateCost(r.tokens, prices),
+      c.answerModelId,
+      c.judgeModelId,
+      String(c.topK),
+      fmtNum(r.config.questionsRun),
+      String(c.seed),
+      c.datasetSha.slice(0, 8)
+    ];
+    lines.push("| " + row.join(" | ") + " |");
+  }
+  lines.push("");
+  lines.push("## Run Configuration");
+  lines.push("");
+  for (const r of reports) {
+    const c = r.config;
+    lines.push(`### ${c.answerModelId} / ${c.mode}`);
+    lines.push("");
+    lines.push(`- **Mode**: ${c.mode}`);
+    lines.push(`- **Answer model**: ${c.answerModelId}`);
+    lines.push(`- **Judge model**: ${c.judgeModelId}`);
+    lines.push(`- **topK**: ${c.topK}`);
+    lines.push(`- **Dataset SHA**: \`${c.datasetSha}\``);
+    lines.push(`- **Seed**: ${c.seed}`);
+    lines.push(`- **Categories**: ${c.categories.join(", ")}`);
+    lines.push(`- **Samples run**: ${c.samplesRun}`);
+    lines.push(`- **Questions run**: ${c.questionsRun}`);
+    lines.push(`- **Wall-clock**: ${(r.wallClockMs / 1e3).toFixed(1)}s`);
+    lines.push(`- **Errors**: ${r.errorCount}`);
+    lines.push(`- **Tokens** (in/out): ${fmtNum(r.tokens.totalInputTokens)} / ${fmtNum(r.tokens.totalOutputTokens)}`);
+    lines.push("");
+  }
+  lines.push("## Methodology Notes");
+  lines.push("");
+  lines.push("These results were produced using the Eidentic LoCoMo fair-run harness. The following rules apply:");
+  lines.push("");
+  lines.push("1. **Both speakers are treated as humans.** Turns are ingested as `[SpeakerName]: text` \u2014 never mapped to user/assistant roles.");
+  lines.push("2. **Timestamps are structural.** Each session is prefixed with a header line `Session N \u2014 <date>` and `ingestedAt` metadata carries the epoch-ms.");
+  lines.push("3. **topK \u2264 10 in memory mode.** Larger topK values trivialise retrieval quality and are not permitted.");
+  lines.push("4. **Full-context baseline is required** alongside any memory-mode result.");
+  lines.push("5. **Judge is strict**: a model answer is correct only when it contains the gold answer's specific information. Vague/topical-only answers are wrong.");
+  lines.push("6. **Category 5 (adversarial)**: correct = model declined; adversarial-trap match = wrong.");
+  lines.push("7. **Primary metric J(1\u20134)**: denominator is the number of cat 1\u20134 questions actually run (max 1540 on full dataset).");
+  lines.push("8. **Dataset license**: CC BY-NC 4.0 \u2014 raw data is not redistributed; only aggregate results are published.");
+  lines.push("");
+  lines.push("> Category mapping in locomo10.json: 1=multi-hop (282), 2=temporal (321), 3=open-domain (96), 4=single-hop (841), 5=adversarial (446).");
+  lines.push("");
+  return lines.join("\n");
+}
+
+// src/locomo-types.ts
+function resolveEvidence(sample, diaIds) {
+  const turnMap = /* @__PURE__ */ new Map();
+  for (const sess of sample.sessions) {
+    for (const turn of sess.turns) {
+      turnMap.set(turn.diaId, turn);
+    }
+  }
+  const results = [];
+  for (const id of diaIds) {
+    const t = turnMap.get(id);
+    if (t) results.push(t.text);
+  }
+  return results;
+}
+
 // src/write-quality.ts
 var CONTRADICTION_FIXTURES = [
   {
@@ -822,7 +1471,7 @@ async function runTemporalBench(memory, dataset, opts = {}) {
 }
 
 // src/datasets/temporal.ts
-function makeRng(seed) {
+function makeRng2(seed) {
   let s = seed >>> 0;
   if (s === 0) s = 1;
   return () => {
@@ -905,7 +1554,7 @@ function syntheticTemporalDataset(opts = {}) {
   const entityCount = opts.entityCount ?? 4;
   const seed = opts.seed ?? 42;
   const changesPerProperty = opts.changesPerProperty ?? 3;
-  const rng = makeRng(seed);
+  const rng = makeRng2(seed);
   const entities = [];
   const asserts = [];
   const questions = [];
@@ -992,12 +1641,17 @@ function syntheticTemporalDataset(opts = {}) {
 0 && (module.exports = {
   CONTRADICTION_FIXTURES,
   JUNK_STREAM_FIXTURES,
+  LOCOMO_SOURCE_SHA,
   factRecall,
   loadLoCoMo,
+  loadLoCoMoLegacy,
   loadLongMemEval,
   normalizeText,
   normalizedIncludes,
   recallAtK,
+  renderLocomoReportMarkdown,
+  resolveEvidence,
+  runLocomoBench,
   runMemoryBench,
   runTemporalBench,
   runWriteQualityBench,