dravoice 0.1.0 → 0.1.2

package/src/v2/benchmark.js

@@ -5,30 +5,22 @@ import { voicePromptPackV2 } from "./prompt.js";
 import { reviewVoiceDraftV2 } from "./review.js";
 import { clampScore } from "./text-utils.js";
 
-const BENCHMARK_SCHEMA_VERSION = 2;
-const GENERATED_BY = "dravoice-v2-benchmark";
-const VOICE_EXTENSIONS = new Set([".md", ".mdx", ".txt"]);
-const SKIP_DIRS = new Set([".git", "dist", "build", "node_modules", ".venv", "venv", "__pycache__"]);
-const SOURCE_SNAPSHOT_MARKER = ".dravoice-benchmark-sources";
-const BENCHMARK_ARTIFACT_DIRS = new Set(["prompts", "drafts", "judge"]);
-
-export function prepareVoiceBenchmark({ examplesDir, topic, outDir, seed = 1, cwd = process.cwd() }) {
-  const examplesRoot = path.resolve(resolvePath(cwd, examplesDir));
-  const outputRoot = path.resolve(resolvePath(cwd, outDir));
-  const normalizedSeed = normalizeSeed(seed);
-  const sourceFiles = walkVoiceFiles(examplesRoot);
-  if (sourceFiles.length === 0) {
-    throw new Error(`No Markdown, MDX, or text examples found at ${examplesDir}`);
-  }
-
-  const voicePackDir = path.join(outputRoot, "voice-pack");
-  const promptsDir = path.join(outputRoot, "prompts");
-  const draftsDir = path.join(outputRoot, "drafts");
-  fs.mkdirSync(promptsDir, { recursive: true });
-  fs.mkdirSync(draftsDir, { recursive: true });
-
-  const profile = learnVoicePackV2({ examplesDir: examplesRoot, outDir: voicePackDir });
-  const sources = sourceFiles.map((filePath, index) => sourceArticle(filePath, examplesRoot, index));
+const BENCHMARK_SCHEMA_VERSION = 2;
+const GENERATED_BY = "dravoice-v2-benchmark";
+
+export function prepareVoiceBenchmark({ examplesDir, topic, outDir, seed = 1, cwd = process.cwd() }) {
+  const examplesRoot = path.resolve(resolvePath(cwd, examplesDir));
+  const outputRoot = path.resolve(resolvePath(cwd, outDir));
+  const normalizedSeed = normalizeSeed(seed);
+
+  const voicePackDir = path.join(outputRoot, "voice-pack");
+  const promptsDir = path.join(outputRoot, "prompts");
+  const draftsDir = path.join(outputRoot, "drafts");
+  fs.mkdirSync(promptsDir, { recursive: true });
+  fs.mkdirSync(draftsDir, { recursive: true });
+
+  const profile = learnVoicePackV2({ examplesDir: examplesRoot, outDir: voicePackDir });
+  const sources = benchmarkSourcesFromProfile(profile);
   const blind = blindMapping(normalizedSeed);
   const benchmark = {
     schemaVersion: BENCHMARK_SCHEMA_VERSION,
@@ -538,58 +530,56 @@ function assertObject(value, field) {
   }
 }
 
-function sourceArticle(filePath, rootDir, index = 0) {
-  const contents = fs.readFileSync(filePath, "utf8");
-  const extension = path.extname(filePath).toLowerCase() || ".txt";
-  return {
-    id: `source-${index + 1}`,
-    path: `source-${index + 1}${extension}`,
-    title: `Source ${index + 1}`,
-    words: (contents.match(/[a-z][a-z0-9'-]*/gi) ?? []).length,
-  };
-}
-
-function benchmarkSources(benchmark, resolved) {
-  return resolved.corpus.files.map((file, index) => ({
-    id: file.id ?? `source-${index + 1}`,
-    path: file.path,
-    title: file.title ?? `Source ${index + 1}`,
-    words: file.words,
-  }));
-}
-
-function benchmarkDrafts(benchmark, resolved) {
-  return ["A", "B"].map((label) => {
-    const draftPath = benchmark.blind.draftFiles[label];
-    const resolvedDraftPath = resolved.blind.draftFiles[label];
-    return {
-      label,
-      path: draftPath,
-      words: fs.existsSync(resolvedDraftPath)
-        ? (fs.readFileSync(resolvedDraftPath, "utf8").match(/[a-z][a-z0-9'-]*/gi) ?? []).length
-        : 0,
-    };
-  });
-}
-
-function blindMapping(seed) {
-  const baselineIsA = seededRandom(seed) < 0.5;
-  const labels = {
-    baseline: baselineIsA ? "A" : "B",
-    voiceAssisted: baselineIsA ? "B" : "A",
-  };
-  return {
-    labels,
-    drafts: {
-      A: baselineIsA ? "baseline" : "voice-assisted",
-      B: baselineIsA ? "voice-assisted" : "baseline",
-    },
-    draftFiles: {
-      A: baselineIsA ? "drafts/baseline.md" : "drafts/voice-assisted.md",
-      B: baselineIsA ? "drafts/voice-assisted.md" : "drafts/baseline.md",
-    },
-  };
-}
+function benchmarkSourcesFromProfile(profile) {
+  return profile.source.files.map((file, index) => ({
+    id: file.id ?? `source-${index + 1}`,
+    path: `source-${index + 1}${file.extension || ".txt"}`,
+    title: `Source ${index + 1}`,
+    words: file.wordCount,
+  }));
+}
+
+function benchmarkSources(benchmark, resolved) {
+  return resolved.corpus.files.map((file, index) => ({
+    id: file.id ?? `source-${index + 1}`,
+    path: file.path,
+    title: file.title ?? `Source ${index + 1}`,
+    words: file.words,
+  }));
+}
+
+function benchmarkDrafts(benchmark, resolved) {
+  return ["A", "B"].map((label) => {
+    const draftPath = benchmark.blind.draftFiles[label];
+    const resolvedDraftPath = resolved.blind.draftFiles[label];
+    return {
+      label,
+      path: draftPath,
+      words: fs.existsSync(resolvedDraftPath)
+        ? (fs.readFileSync(resolvedDraftPath, "utf8").match(/[a-z][a-z0-9'-]*/gi) ?? []).length
+        : 0,
+    };
+  });
+}
+
+function blindMapping(seed) {
+  const baselineIsA = seededRandom(seed) < 0.5;
+  const labels = {
+    baseline: baselineIsA ? "A" : "B",
+    voiceAssisted: baselineIsA ? "B" : "A",
+  };
+  return {
+    labels,
+    drafts: {
+      A: baselineIsA ? "baseline" : "voice-assisted",
+      B: baselineIsA ? "voice-assisted" : "baseline",
+    },
+    draftFiles: {
+      A: baselineIsA ? "drafts/baseline.md" : "drafts/voice-assisted.md",
+      B: baselineIsA ? "drafts/voice-assisted.md" : "drafts/baseline.md",
+    },
+  };
+}
 
 function seededRandom(seed) {
   const next = (normalizeSeed(seed) * 1664525 + 1013904223) >>> 0;
@@ -608,66 +598,7 @@ function normalizeSeed(seed) {
   return Number(parsed);
 }
 
-function walkVoiceFiles(rootDir) {
-  if (!fs.existsSync(rootDir)) {
-    return [];
-  }
-  const result = [];
-  for (const entry of fs.readdirSync(rootDir, { withFileTypes: true }).sort((a, b) => a.name.localeCompare(b.name))) {
-    const fullPath = path.join(rootDir, entry.name);
-    if (entry.isDirectory()) {
-      if (!isSkippedVoiceDir(fullPath, entry.name)) {
-        result.push(...walkVoiceFiles(fullPath));
-      }
-    } else if (VOICE_EXTENSIONS.has(path.extname(entry.name).toLowerCase()) && !isGeneratedBenchmarkMarkdown(fullPath)) {
-      result.push(fullPath);
-    }
-  }
-  return result.sort((a, b) => toPosix(path.relative(rootDir, a)).localeCompare(toPosix(path.relative(rootDir, b))));
-}
-
-function isSkippedVoiceDir(dirPath, name) {
-  if (SKIP_DIRS.has(name)) {
-    return true;
-  }
-  if (name === "voice-pack" || name === "dravoice-voice" || name === ".dravoice") {
-    return true;
-  }
-  if (fs.existsSync(path.join(dirPath, SOURCE_SNAPSHOT_MARKER))) {
-    return true;
-  }
-  if (isBenchmarkRunDir(dirPath)) {
-    return true;
-  }
-  if (BENCHMARK_ARTIFACT_DIRS.has(name) && isBenchmarkRunDir(path.dirname(dirPath))) {
-    return true;
-  }
-  return fs.existsSync(path.join(dirPath, "profile.json"));
-}
-
-function isBenchmarkRunDir(dirPath) {
-  const benchmarkPath = path.join(dirPath, "benchmark.json");
-  if (!fs.existsSync(benchmarkPath)) {
-    return false;
-  }
-  try {
-    const benchmark = JSON.parse(fs.readFileSync(benchmarkPath, "utf8"));
-    return benchmark?.generatedBy === GENERATED_BY;
-  } catch {
-    return false;
-  }
-}
-
-function isGeneratedBenchmarkMarkdown(filePath) {
-  const basename = path.basename(filePath).toLowerCase();
-  if (!["baseline-writer.md", "voice-writer.md", "judge.md", "report.md"].includes(basename)) {
-    return false;
-  }
-  const contents = fs.readFileSync(filePath, "utf8");
-  return /^# (?:Baseline Writer Prompt|Voice-Assisted Writer Prompt|Blind V2 Voice Benchmark Judge Prompt|Dravoice V2 Voice Benchmark Report)/m.test(contents);
-}
-
-function writeIfMissing(filePath, contents) {
+function writeIfMissing(filePath, contents) {
   if (!fs.existsSync(filePath)) {
     fs.writeFileSync(filePath, contents, "utf8");
   }

package/src/v2/brief.js

@@ -1,146 +1,146 @@
-import fs from "node:fs";
-import path from "node:path";
-import { evidenceTypes } from "./analyzers/evidence.js";
-import { parseDocument } from "./document-model.js";
-import { loadVoicePackV2 } from "./profile.js";
-
-export function voiceArticleBriefV2({ voice, topic, evidence, cwd = process.cwd() }) {
-  const profile = typeof voice === "string" ? loadVoicePackV2(resolvePath(cwd, voice)) : voice;
-  const evidenceResult = evidence ? evidenceAnchorsFromFile({ evidence, cwd }) : {
-    source: null,
-    anchors: [],
-  };
-
-  return {
-    schemaVersion: 2,
-    generatedBy: "dravoice-v2-brief",
-    topic,
-    voice: {
-      corpusConfidence: profile.source.confidence,
-      sourceFileCount: profile.source.documentCount,
-      sourceWordCount: profile.source.wordCount,
-      primaryRegister: profile.families.register.features.primary.value,
-      evidenceSentenceRate: profile.families.evidence.features.evidenceSentenceRate,
-      featureFamilies: Object.keys(profile.families),
-      draftingRules: profile.guidance.draftingRules.slice(0, 5),
-    },
-    workingThesis: `Draft a grounded article about ${topic}. Let the supplied evidence set the size of each claim before broadening the lesson.`,
-    evidence: evidenceResult,
-    missingEvidence: missingEvidenceFor({ topic, evidenceAnchors: evidenceResult.anchors }),
-    outline: outlineFor(profile),
-    voiceCautions: [
-      ...profile.guidance.avoid,
-      "Mark unsupported claims as [specific evidence needed] instead of inventing proof.",
-    ].slice(0, 5),
-  };
-}
-
-export function renderVoiceBriefV2(brief) {
-  const lines = [
-    `# Article Brief: ${brief.topic}`,
-    "",
-    "## Voice Source",
-    "",
-    `- Corpus confidence: ${capitalize(brief.voice.corpusConfidence.band)} - ${brief.voice.corpusConfidence.message}`,
-    `- Source files: ${brief.voice.sourceFileCount}`,
-    `- Primary register: ${brief.voice.primaryRegister}`,
-    `- Evidence sentence rate: ${brief.voice.evidenceSentenceRate}`,
-    "",
-    "## Working Thesis",
-    "",
-    `- ${brief.workingThesis}`,
-    "",
-    "## Evidence Anchors",
-    "",
-  ];
-
-  if (brief.evidence.anchors.length) {
-    for (const item of brief.evidence.anchors) {
-      const typeList = item.types.length ? ` (${item.types.join(", ")})` : "";
-      lines.push(`- ${brief.evidence.source}:${item.line}${typeList} - ${item.text}`);
-    }
-  } else {
-    lines.push("- [specific evidence needed] Add notes, dates, quotes, examples, or source links before drafting broad claims.");
-  }
-
-  lines.push("", "## Missing Evidence", "");
-  lines.push(...brief.missingEvidence.map((item) => `- ${item}`));
-
-  lines.push("", "## Outline", "");
-  lines.push(...brief.outline.map((item, index) => `${index + 1}. ${item}`));
-
-  lines.push("", "## Voice Cautions", "");
-  lines.push(...brief.voiceCautions.map((item) => `- ${item}`));
-
-  lines.push(
-    "",
-    "## Drafting Prompt",
-    "",
-    `Write the article about ${brief.topic} using the evidence anchors above. Keep claims close to concrete support, follow the voice cautions, and write [specific evidence needed] anywhere the brief does not supply enough ground.`,
-    "",
-  );
-
-  return lines.join("\n");
-}
-
-function evidenceAnchorsFromFile({ evidence, cwd }) {
-  const evidencePath = resolvePath(cwd, evidence);
-  const contents = fs.readFileSync(evidencePath, "utf8");
-  const document = parseDocument({
-    filePath: evidencePath,
-    rootDir: cwd,
-    contents,
-  });
-
-  return {
-    source: displayPath(evidencePath, cwd),
-    anchors: document.sentences
-      .map((sentence) => ({
-        line: sentence.line,
-        text: sentence.text,
-        types: evidenceTypes(sentence.text),
-      }))
-      .filter((sentence) => sentence.types.length > 0)
-      .slice(0, 8),
-  };
-}
-
-function missingEvidenceFor({ topic, evidenceAnchors }) {
-  const items = [
-    `Add [specific evidence needed] for the central claim about ${topic}.`,
-    "Add [specific evidence needed] for any number, date, quote, source, or example the article depends on.",
-  ];
-  if (evidenceAnchors.length === 0) {
-    items.unshift("No evidence anchors were detected; collect concrete notes before asking for a full draft.");
-  }
-  return items;
-}
-
-function outlineFor(profile) {
-  const opening = profile.families.rhetoricalShape.features.openingMoves.slice(0, 3).join(" -> ");
-  const sentenceMedian = profile.families.rhythm.features.sentenceWords.median;
-  return [
-    opening
-      ? `Start from a concrete artifact or observation, keeping the opening shape compatible with: ${opening}.`
-      : "Start from a concrete artifact or observation before making the larger claim.",
-    "Name the pressure, question, or practical stakes that make the evidence matter.",
-    `Develop the article in the learned register with sentence pacing near the ${sentenceMedian}-word median where it fits.`,
-    "Close by returning to the evidence and leaving the reader with a practical handle, not a generic conclusion.",
-  ];
-}
-
-function resolvePath(cwd, value) {
-  return path.isAbsolute(value) ? value : path.join(cwd, value);
-}
-
-function displayPath(filePath, rootDir) {
-  const relative = path.relative(rootDir, filePath);
-  if (relative && !relative.startsWith("..") && !path.isAbsolute(relative)) {
-    return relative.split(path.sep).join("/");
-  }
-  return filePath.split(path.sep).join("/");
-}
-
-function capitalize(value) {
-  return String(value ?? "").charAt(0).toUpperCase() + String(value ?? "").slice(1);
-}
+import fs from "node:fs";
+import path from "node:path";
+import { evidenceTypes } from "./analyzers/evidence.js";
+import { parseDocument } from "./document-model.js";
+import { loadVoicePackV2 } from "./profile.js";
+
+export function voiceArticleBriefV2({ voice, topic, evidence, cwd = process.cwd() }) {
+  const profile = typeof voice === "string" ? loadVoicePackV2(resolvePath(cwd, voice)) : voice;
+  const evidenceResult = evidence ? evidenceAnchorsFromFile({ evidence, cwd }) : {
+    source: null,
+    anchors: [],
+  };
+
+  return {
+    schemaVersion: 2,
+    generatedBy: "dravoice-v2-brief",
+    topic,
+    voice: {
+      corpusConfidence: profile.source.confidence,
+      sourceFileCount: profile.source.documentCount,
+      sourceWordCount: profile.source.wordCount,
+      primaryRegister: profile.families.register.features.primary.value,
+      evidenceSentenceRate: profile.families.evidence.features.evidenceSentenceRate,
+      featureFamilies: Object.keys(profile.families),
+      draftingRules: profile.guidance.draftingRules.slice(0, 5),
+    },
+    workingThesis: `Draft a grounded article about ${topic}. Let the supplied evidence set the size of each claim before broadening the lesson.`,
+    evidence: evidenceResult,
+    missingEvidence: missingEvidenceFor({ topic, evidenceAnchors: evidenceResult.anchors }),
+    outline: outlineFor(profile),
+    voiceCautions: [
+      ...profile.guidance.avoid,
+      "Mark unsupported claims as [specific evidence needed] instead of inventing proof.",
+    ].slice(0, 5),
+  };
+}
+
+export function renderVoiceBriefV2(brief) {
+  const lines = [
+    `# Article Brief: ${brief.topic}`,
+    "",
+    "## Voice Source",
+    "",
+    `- Corpus confidence: ${capitalize(brief.voice.corpusConfidence.band)} - ${brief.voice.corpusConfidence.message}`,
+    `- Source files: ${brief.voice.sourceFileCount}`,
+    `- Primary register: ${brief.voice.primaryRegister}`,
+    `- Evidence sentence rate: ${brief.voice.evidenceSentenceRate}`,
+    "",
+    "## Working Thesis",
+    "",
+    `- ${brief.workingThesis}`,
+    "",
+    "## Evidence Anchors",
+    "",
+  ];
+
+  if (brief.evidence.anchors.length) {
+    for (const item of brief.evidence.anchors) {
+      const typeList = item.types.length ? ` (${item.types.join(", ")})` : "";
+      lines.push(`- ${brief.evidence.source}:${item.line}${typeList} - ${item.text}`);
+    }
+  } else {
+    lines.push("- [specific evidence needed] Add notes, dates, quotes, examples, or source links before drafting broad claims.");
+  }
+
+  lines.push("", "## Missing Evidence", "");
+  lines.push(...brief.missingEvidence.map((item) => `- ${item}`));
+
+  lines.push("", "## Outline", "");
+  lines.push(...brief.outline.map((item, index) => `${index + 1}. ${item}`));
+
+  lines.push("", "## Voice Cautions", "");
+  lines.push(...brief.voiceCautions.map((item) => `- ${item}`));
+
+  lines.push(
+    "",
+    "## Drafting Prompt",
+    "",
+    `Write the article about ${brief.topic} using the evidence anchors above. Keep claims close to concrete support, follow the voice cautions, and write [specific evidence needed] anywhere the brief does not supply enough ground.`,
+    "",
+  );
+
+  return lines.join("\n");
+}
+
+function evidenceAnchorsFromFile({ evidence, cwd }) {
+  const evidencePath = resolvePath(cwd, evidence);
+  const contents = fs.readFileSync(evidencePath, "utf8");
+  const document = parseDocument({
+    filePath: evidencePath,
+    rootDir: cwd,
+    contents,
+  });
+
+  return {
+    source: displayPath(evidencePath, cwd),
+    anchors: document.sentences
+      .map((sentence) => ({
+        line: sentence.line,
+        text: sentence.text,
+        types: evidenceTypes(sentence.text),
+      }))
+      .filter((sentence) => sentence.types.length > 0)
+      .slice(0, 8),
+  };
+}
+
+function missingEvidenceFor({ topic, evidenceAnchors }) {
+  const items = [
+    `Add [specific evidence needed] for the central claim about ${topic}.`,
+    "Add [specific evidence needed] for any number, date, quote, source, or example the article depends on.",
+  ];
+  if (evidenceAnchors.length === 0) {
+    items.unshift("No evidence anchors were detected; collect concrete notes before asking for a full draft.");
+  }
+  return items;
+}
+
+function outlineFor(profile) {
+  const opening = profile.families.rhetoricalShape.features.openingMoves.slice(0, 3).join(" -> ");
+  const sentenceMedian = profile.families.rhythm.features.sentenceWords.median;
+  return [
+    opening
+      ? `Start from a concrete artifact or observation, keeping the opening shape compatible with: ${opening}.`
+      : "Start from a concrete artifact or observation before making the larger claim.",
+    "Name the pressure, question, or practical stakes that make the evidence matter.",
+    `Develop the article in the learned register with sentence pacing near the ${sentenceMedian}-word median where it fits.`,
+    "Close by returning to the evidence and leaving the reader with a practical handle, not a generic conclusion.",
+  ];
+}
+
+function resolvePath(cwd, value) {
+  return path.isAbsolute(value) ? value : path.join(cwd, value);
+}
+
+function displayPath(filePath, rootDir) {
+  const relative = path.relative(rootDir, filePath);
+  if (relative && !relative.startsWith("..") && !path.isAbsolute(relative)) {
+    return relative.split(path.sep).join("/");
+  }
+  return filePath.split(path.sep).join("/");
+}
+
+function capitalize(value) {
+  return String(value ?? "").charAt(0).toUpperCase() + String(value ?? "").slice(1);
+}

package/src/v2/inspect.js

@@ -46,7 +46,7 @@ function featureSummary(name, features) {
     return `sentenceWords.median=${features.sentenceWords.median}; paragraphWords.median=${features.paragraphWords.median}; listDensity=${features.listDensity}; quoteDensity=${features.quoteDensity}`;
   }
   if (name === "lexical") {
-    return `wordCount=${features.wordCount}; contentTypeTokenRatio=${features.vocabularyRichness.contentTypeTokenRatio}; wordLength.median=${features.wordLength.median}`;
+    return `wordCount=${features.wordCount}; contentTypeTokenRatio=${features.vocabularyRichness.contentTypeTokenRatio}; wordLength.median=${features.wordLength.median}; maskedCharacterFourgrams=${features.maskedCharacterFourgrams?.length ?? 0}; functionWordBigrams=${features.functionWordBigrams?.length ?? 0}`;
   }
   if (name === "register") {
     return `primary=${features.primary.value} (${features.primary.score}); alternates=${features.scores.slice(1, 4).map((score) => `${score.value}:${score.score}`).join(", ")}`;
@@ -55,10 +55,10 @@ function featureSummary(name, features) {
     return `transitionRates=${Object.entries(features.transitionRates).map(([key, value]) => `${key}:${value}`).join(", ")}; sentenceCallbacks=${features.sentenceCallbacks}`;
   }
   if (name === "rhetoricalShape") {
-    return `openingMoves=${features.openingMoves.slice(0, 5).join(" -> ") || "none"}; commonSequences=${features.commonSequences.slice(0, 3).map((item) => `${item.value}:${item.count}`).join(", ") || "none"}`;
+    return `openingMoves=${features.openingMoves.slice(0, 5).join(" -> ") || "none"}; moveTrigrams=${features.moveTrigrams?.slice(0, 2).map((item) => `${item.value}:${item.count}`).join(", ") || "none"}; commonSequences=${features.commonSequences.slice(0, 3).map((item) => `${item.value}:${item.count}`).join(", ") || "none"}`;
   }
   if (name === "evidence") {
-    return `evidenceSentenceRate=${features.evidenceSentenceRate}; claimSentenceRate=${features.claimSentenceRate}; unsupportedClaimRate=${features.unsupportedClaimRate}; evidenceTypes=${features.evidenceTypes.map((item) => `${item.value}:${item.count}`).join(", ") || "none"}`;
+    return `evidenceSentenceRate=${features.evidenceSentenceRate}; claimSentenceRate=${features.claimSentenceRate}; supportedClaimRate=${features.supportedClaimRate}; unsupportedClaimRate=${features.unsupportedClaimRate}; evidenceTypes=${features.evidenceTypes.map((item) => `${item.value}:${item.count}`).join(", ") || "none"}`;
   }
   if (name === "structure") {
     return `sectionWords.median=${features.sectionWords.median}; headingCount.median=${features.headingCount.median}; listDocumentRate=${features.listDocumentRate}; quoteDocumentRate=${features.quoteDocumentRate}`;

package/src/v2/profile.js

@@ -8,6 +8,13 @@ import { analyzeRhetoricalShape } from "./analyzers/rhetorical-shape.js";
 import { analyzeRhythm } from "./analyzers/rhythm.js";
 import { analyzeStructure } from "./analyzers/structure.js";
 import { loadDocuments } from "./document-model.js";
+import {
+  STYLOMETRIC_REFERENCES,
+  defaultStyleThresholds,
+  distanceByFamily,
+  percentile,
+  stabilityFromDistances,
+} from "./stylometry.js";
 
 export function learnVoicePackV2({ examplesDir, outDir }) {
   const documents = loadDocuments({ examplesDir });
@@ -32,20 +39,12 @@ export function loadVoicePackV2(voiceDir) {
 
 export function buildVoiceProfileV2({ documents }) {
   const source = sourceSummary(documents);
-  const families = {
-    rhythm: analyzeRhythm(documents),
-    lexical: analyzeLexical(documents),
-    register: analyzeRegister(documents),
-    discourse: analyzeDiscourse(documents),
-    rhetoricalShape: analyzeRhetoricalShape(documents),
-    evidence: analyzeEvidence(documents),
-    structure: analyzeStructure(documents),
-  };
+  const families = analyzeFeatureFamilies(documents);
 
   return {
     schemaVersion: 2,
     generatedBy: "dravoice-v2",
-    tool: { name: "Dravoice", cli: "drav" },
+    tool: { name: "Dravoice", cli: "drav" },
     source,
     families,
     guidance: guidanceFor({ source, families }),
@@ -55,6 +54,7 @@ export function buildVoiceProfileV2({ documents }) {
         rhythmMedianWords: toleranceFor(source.confidence.band, 5, 8, 12),
         evidenceRate: toleranceFor(source.confidence.band, 0.12, 0.18, 0.25),
       },
+      styleThresholds: styleThresholdsFor(documents, families),
       minimumDraftSize: {
         words: source.confidence.band === "weak" ? 25 : 35,
         sentences: source.confidence.band === "weak" ? 3 : 4,
@@ -63,6 +63,56 @@ export function buildVoiceProfileV2({ documents }) {
   };
 }
 
+function analyzeFeatureFamilies(documents) {
+  return {
+    rhythm: analyzeRhythm(documents),
+    lexical: analyzeLexical(documents),
+    register: analyzeRegister(documents),
+    discourse: analyzeDiscourse(documents),
+    rhetoricalShape: analyzeRhetoricalShape(documents),
+    evidence: analyzeEvidence(documents),
+    structure: analyzeStructure(documents),
+  };
+}
+
+function styleThresholdsFor(documents, fallbackFamilies) {
+  const fallbackThresholds = defaultStyleThresholds();
+  const distancesByFamily = Object.fromEntries(Object.keys(fallbackFamilies).map((family) => [family, []]));
+
+  if (documents.length >= 2) {
+    for (let index = 0; index < documents.length; index += 1) {
+      const referenceDocuments = documents.filter((_, candidateIndex) => candidateIndex !== index);
+      const referenceFamilies = analyzeFeatureFamilies(referenceDocuments);
+      const heldoutFamilies = analyzeFeatureFamilies([documents[index]]);
+      for (const family of Object.keys(fallbackFamilies)) {
+        distancesByFamily[family].push(distanceByFamily(
+          family,
+          referenceFamilies[family].features,
+          heldoutFamilies[family].features,
+        ));
+      }
+    }
+  }
+
+  const families = {};
+  for (const family of Object.keys(fallbackFamilies)) {
+    const observations = distancesByFamily[family];
+    const observedThreshold = observations.length > 0 ? percentile(observations, 0.9) : 0;
+    const fallback = fallbackThresholds[family].threshold;
+    families[family] = {
+      threshold: Math.max(0.01, Math.min(0.95, observedThreshold || fallback)),
+      observations: observations.length,
+      stability: stabilityFromDistances(observations),
+    };
+  }
+
+  return {
+    method: "leave-one-out-cosine-delta",
+    references: STYLOMETRIC_REFERENCES,
+    families,
+  };
+}
+
 function writeVoicePackV2(outDir, profile) {
   fs.mkdirSync(outDir, { recursive: true });
   fs.writeFileSync(path.join(outDir, "profile.json"), `${JSON.stringify(profile, null, 2)}\n`, "utf8");