causantic 0.2.0

package/dist/eval/corpus-builder.js

@@ -0,0 +1,88 @@
+/**
+ * Build a test corpus from real Claude Code sessions.
+ *
+ * Samples chunks uniformly across the code-ratio spectrum per session.
+ */
+import { readdir, stat } from 'node:fs/promises';
+import { join } from 'node:path';
+import { readSessionMessages, getSessionInfo } from '../parser/session-reader.js';
+import { assembleTurns } from '../parser/turn-assembler.js';
+import { chunkTurns, resetChunkCounter } from '../parser/chunker.js';
+/**
+ * Compute code ratio for a chunk (0 = pure NL, 1 = pure code).
+ */
+function codeRatio(chunk) {
+    const codeMarkers = (chunk.text.match(/```/g)?.length ?? 0) / 2;
+    const toolResults = chunk.metadata.toolUseCount;
+    const totalSignals = codeMarkers + toolResults + 1; // +1 to avoid div by zero
+    return (codeMarkers + toolResults * 0.5) / totalSignals;
+}
+/**
+ * Sample chunks uniformly across the code-ratio spectrum.
+ */
+function sampleUniformByCodeRatio(chunks, maxCount) {
+    if (chunks.length <= maxCount)
+        return chunks;
+    // Sort by code ratio
+    const sorted = [...chunks].sort((a, b) => codeRatio(a) - codeRatio(b));
+    // Sample evenly across the sorted array
+    const step = (sorted.length - 1) / (maxCount - 1);
+    const sampled = [];
+    for (let i = 0; i < maxCount; i++) {
+        const idx = Math.round(i * step);
+        sampled.push(sorted[idx]);
+    }
+    return sampled;
+}
+/**
+ * Build a test corpus from session files.
+ */
+export async function buildCorpus(config) {
+    const { sessionPaths, maxChunksPerSession = 30, maxTokensPerChunk = 4096, includeThinking = true, renderMode = 'full', } = config;
+    resetChunkCounter();
+    const allChunks = [];
+    const sessions = [];
+    for (const sessionPath of sessionPaths) {
+        const info = await getSessionInfo(sessionPath);
+        sessions.push(info);
+        const messages = await readSessionMessages(sessionPath);
+        const turns = assembleTurns(messages);
+        const chunks = chunkTurns(turns, {
+            maxTokens: maxTokensPerChunk,
+            sessionId: info.sessionId,
+            sessionSlug: info.slug,
+            includeThinking,
+            renderMode,
+        });
+        const sampled = sampleUniformByCodeRatio(chunks, maxChunksPerSession);
+        allChunks.push(...sampled);
+        console.log(`  ${info.slug}: ${messages.length} messages -> ${turns.length} turns -> ${chunks.length} chunks -> ${sampled.length} sampled`);
+    }
+    return {
+        chunks: allChunks,
+        sessions,
+        config,
+        builtAt: new Date().toISOString(),
+    };
+}
+/**
+ * Discover session JSONL files in a Claude Code projects directory.
+ */
+export async function discoverSessions(projectDir) {
+    const entries = await readdir(projectDir);
+    const jsonlFiles = [];
+    for (const entry of entries) {
+        if (entry.endsWith('.jsonl')) {
+            const fullPath = join(projectDir, entry);
+            const stats = await stat(fullPath);
+            if (stats.isFile() && stats.size > 1000) {
+                jsonlFiles.push(fullPath);
+            }
+        }
+    }
+    // Sort by file size descending — richest sessions first
+    const withSizes = await Promise.all(jsonlFiles.map(async (f) => ({ path: f, size: (await stat(f)).size })));
+    withSizes.sort((a, b) => b.size - a.size);
+    return withSizes.map((f) => f.path);
+}
+//# sourceMappingURL=corpus-builder.js.map

package/dist/eval/corpus-builder.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"corpus-builder.js","sourceRoot":"","sources":["../../src/eval/corpus-builder.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,MAAM,kBAAkB,CAAC;AACjD,OAAO,EAAE,IAAI,EAAY,MAAM,WAAW,CAAC;AAC3C,OAAO,EAAE,mBAAmB,EAAE,cAAc,EAAE,MAAM,6BAA6B,CAAC;AAClF,OAAO,EAAE,aAAa,EAAE,MAAM,6BAA6B,CAAC;AAC5D,OAAO,EAAE,UAAU,EAAE,iBAAiB,EAAE,MAAM,sBAAsB,CAAC;AAuBrE;;GAEG;AACH,SAAS,SAAS,CAAC,KAAY;IAC7B,MAAM,WAAW,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,EAAE,MAAM,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC;IAChE,MAAM,WAAW,GAAG,KAAK,CAAC,QAAQ,CAAC,YAAY,CAAC;IAChD,MAAM,YAAY,GAAG,WAAW,GAAG,WAAW,GAAG,CAAC,CAAC,CAAC,0BAA0B;IAC9E,OAAO,CAAC,WAAW,GAAG,WAAW,GAAG,GAAG,CAAC,GAAG,YAAY,CAAC;AAC1D,CAAC;AAED;;GAEG;AACH,SAAS,wBAAwB,CAAC,MAAe,EAAE,QAAgB;IACjE,IAAI,MAAM,CAAC,MAAM,IAAI,QAAQ;QAAE,OAAO,MAAM,CAAC;IAE7C,qBAAqB;IACrB,MAAM,MAAM,GAAG,CAAC,GAAG,MAAM,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,SAAS,CAAC,CAAC,CAAC,GAAG,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC;IAEvE,wCAAwC;IACxC,MAAM,IAAI,GAAG,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC,GAAG,CAAC,QAAQ,GAAG,CAAC,CAAC,CAAC;IAClD,MAAM,OAAO,GAAY,EAAE,CAAC;IAC5B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,EAAE,CAAC,EAAE,EAAE,CAAC;QAClC,MAAM,GAAG,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC;QACjC,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC;IAC5B,CAAC;IAED,OAAO,OAAO,CAAC;AACjB,CAAC;AAED;;GAEG;AACH,MAAM,CAAC,KAAK,UAAU,WAAW,CAAC,MAAoB;IACpD,MAAM,EACJ,YAAY,EACZ,mBAAmB,GAAG,EAAE,EACxB,iBAAiB,GAAG,IAAI,EACxB,eAAe,GAAG,IAAI,EACtB,UAAU,GAAG,MAAM,GACpB,GAAG,MAAM,CAAC;IAEX,iBAAiB,EAAE,CAAC;IAEpB,MAAM,SAAS,GAAY,EAAE,CAAC;IAC9B,MAAM,QAAQ,GAAkB,EAAE,CAAC;IAEnC,KAAK,MAAM,WAAW,IAAI,YAAY,EAAE,CAAC;QACvC,MAAM,IAAI,GAAG,MAAM,cAAc,CAAC,WAAW,CAAC,CAAC;QAC/C,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QAEpB,MAAM,QAAQ,GAAG,MAAM,mBAAmB,CAAC,WAAW,CAAC,CAAC;QACxD,MAAM,KAAK,GAAG,aAAa,CAAC,QAAQ,CAAC,CAAC;QACtC,MAAM,MAAM,GAAG,UAAU,CAAC,KAAK,EAAE;YAC/B,SAAS,EAAE,iBAAiB;YAC5B,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,WAAW,EAAE,IAAI,CAAC,IAAI;YACtB,eAAe;YACf,UAAU;SACX,CAAC,CAAC;QAEH,MAAM,OAAO,GAAG,wBAAwB,CAAC,MAAM,EAAE,mBAAmB,CAAC,CAAC;QACtE,SAAS,CAAC,IAAI,CAAC,GAAG,OAAO,CAAC,CAAC;QAE3B,OAAO,CAAC,GAAG,CACT,KAAK,IAAI,CAAC,IAAI,KAAK,QAAQ,CAAC,MAAM,gBAAgB,KAAK,CAAC,MAAM,aAAa,MAAM,CAAC,MAAM,cAAc,OAAO,CAAC,MAAM,UAAU,CAC/H,CAAC;IACJ,CAAC;IAED,OAAO;QACL,MAAM,EAAE,SAAS;QACjB,QAAQ;QACR,MAAM;QACN,OAAO,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;KAClC,CAAC;AACJ,CAAC;AAED;;GAEG;AACH,MAAM,CAAC,KAAK,UAAU,gBAAgB,CACpC,UAAkB;IAElB,MAAM,OAAO,GAAG,MAAM,OAAO,CAAC,UAAU,CAAC,CAAC;IAC1C,MAAM,UAAU,GAAa,EAAE,CAAC;IAEhC,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;QAC5B,IAAI,KAAK,CAAC,QAAQ,CAAC,QAAQ,CAAC,EAAE,CAAC;YAC7B,MAAM,QAAQ,GAAG,IAAI,CAAC,UAAU,EAAE,KAAK,CAAC,CAAC;YACzC,MAAM,KAAK,GAAG,MAAM,IAAI,CAAC,QAAQ,CAAC,CAAC;YACnC,IAAI,KAAK,CAAC,MAAM,EAAE,IAAI,KAAK,CAAC,IAAI,GAAG,IAAI,EAAE,CAAC;gBACxC,UAAU,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;YAC5B,CAAC;QACH,CAAC;IACH,CAAC;IAED,wDAAwD;IACxD,MAAM,SAAS,GAAG,MAAM,OAAO,CAAC,GAAG,CACjC,UAAU,CAAC,GAAG,CAAC,KAAK,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,IAAI,EAAE,CAAC,EAAE,IAAI,EAAE,CAAC,MAAM,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,CACvE,CAAC;IACF,SAAS,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC;IAC1C,OAAO,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC;AACtC,CAAC"}

package/dist/eval/experiments/boilerplate-filter.d.ts

@@ -0,0 +1,26 @@
+/**
+ * Experiment 3: Boilerplate Filtering
+ *
+ * Question: Does stripping "This session is being continued..." boilerplate
+ * improve embedding discrimination?
+ *
+ * Defines a boilerplate filter that strips known prefixes from chunk text
+ * before embedding, then compares ROC AUC and cluster membership vs unfiltered.
+ */
+import type { Chunk } from '../../parser/types.js';
+import type { LabeledPair } from '../annotation-schema.js';
+import { type SingleModelResult } from './single-model-run.js';
+import { type ExperimentResult } from './types.js';
+/**
+ * Strip known boilerplate text from a chunk.
+ * Returns the filtered text and whether any boilerplate was found.
+ */
+export declare function stripBoilerplate(text: string): {
+    text: string;
+    stripped: boolean;
+};
+/**
+ * Run the boilerplate filtering experiment.
+ */
+export declare function runBoilerplateExperiment(chunks: Chunk[], pairs: LabeledPair[], baselineResult?: SingleModelResult): Promise<ExperimentResult>;
+//# sourceMappingURL=boilerplate-filter.d.ts.map

package/dist/eval/experiments/boilerplate-filter.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"boilerplate-filter.d.ts","sourceRoot":"","sources":["../../../src/eval/experiments/boilerplate-filter.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,KAAK,EAAE,KAAK,EAAE,MAAM,uBAAuB,CAAC;AACnD,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,yBAAyB,CAAC;AAC3D,OAAO,EAIL,KAAK,iBAAiB,EACvB,MAAM,uBAAuB,CAAC;AAC/B,OAAO,EAAgB,KAAK,gBAAgB,EAAE,MAAM,YAAY,CAAC;AAgBjE;;;GAGG;AACH,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,MAAM,GAAG;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,QAAQ,EAAE,OAAO,CAAA;CAAE,CAclF;AAED;;GAEG;AACH,wBAAsB,wBAAwB,CAC5C,MAAM,EAAE,KAAK,EAAE,EACf,KAAK,EAAE,WAAW,EAAE,EACpB,cAAc,CAAC,EAAE,iBAAiB,GACjC,OAAO,CAAC,gBAAgB,CAAC,CA0D3B"}

package/dist/eval/experiments/boilerplate-filter.js

@@ -0,0 +1,95 @@
+/**
+ * Experiment 3: Boilerplate Filtering
+ *
+ * Question: Does stripping "This session is being continued..." boilerplate
+ * improve embedding discrimination?
+ *
+ * Defines a boilerplate filter that strips known prefixes from chunk text
+ * before embedding, then compares ROC AUC and cluster membership vs unfiltered.
+ */
+import { singleModelRun, embedTextsAndScore, toSnapshot, } from './single-model-run.js';
+import { computeDelta } from './types.js';
+const MODEL_ID = 'jina-small';
+/**
+ * Known boilerplate patterns found in Claude Code session continuations.
+ * These are matched at line boundaries within the chunk text.
+ */
+const BOILERPLATE_PATTERNS = [
+    /This session is being continued from a previous conversation[^\n]*/g,
+    /If you need specific details from before compaction[^\n]*/g,
+    /Please continue the conversation from where we left it off[^\n]*/g,
+    /Here is a summary of the conversation so far[^\n]*/g,
+    /The conversation has been compacted[^\n]*/g,
+];
+/**
+ * Strip known boilerplate text from a chunk.
+ * Returns the filtered text and whether any boilerplate was found.
+ */
+export function stripBoilerplate(text) {
+    let filtered = text;
+    let stripped = false;
+    for (const pattern of BOILERPLATE_PATTERNS) {
+        const before = filtered;
+        filtered = filtered.replace(pattern, '');
+        if (filtered !== before)
+            stripped = true;
+    }
+    // Clean up resulting empty lines
+    filtered = filtered.replace(/\n{3,}/g, '\n\n').trim();
+    return { text: filtered, stripped };
+}
+/**
+ * Run the boilerplate filtering experiment.
+ */
+export async function runBoilerplateExperiment(chunks, pairs, baselineResult) {
+    console.log('\n=== Experiment 3: Boilerplate Filtering ===');
+    // Baseline: unfiltered
+    let baseline;
+    if (baselineResult) {
+        console.log('  Reusing cached baseline embeddings');
+        baseline = baselineResult;
+    }
+    else {
+        console.log('  Running baseline (unfiltered)...');
+        baseline = await singleModelRun(MODEL_ID, chunks, pairs);
+    }
+    // Filter boilerplate from chunk texts
+    let affectedCount = 0;
+    const filteredTexts = chunks.map((c) => {
+        const { text, stripped } = stripBoilerplate(c.text);
+        if (stripped)
+            affectedCount++;
+        return { id: c.id, text };
+    });
+    console.log(`  ${affectedCount}/${chunks.length} chunks contained boilerplate ` +
+        `(${((affectedCount / chunks.length) * 100).toFixed(1)}%)`);
+    if (affectedCount === 0) {
+        console.log('  No boilerplate found — variant identical to baseline');
+        const snap = toSnapshot(baseline, chunks.length);
+        return {
+            name: 'boilerplate-filter',
+            description: 'No boilerplate found in corpus. Variant identical to baseline.',
+            baseline: snap,
+            variant: snap,
+            delta: computeDelta(snap, snap),
+        };
+    }
+    // Variant: filtered text
+    console.log('  Running variant (filtered)...');
+    const variant = await embedTextsAndScore(MODEL_ID, filteredTexts, pairs);
+    const baselineSnap = toSnapshot(baseline, chunks.length);
+    const variantSnap = toSnapshot(variant, chunks.length);
+    const delta = computeDelta(baselineSnap, variantSnap);
+    console.log(`  Baseline ROC AUC: ${baselineSnap.rocAuc.toFixed(3)}`);
+    console.log(`  Filtered ROC AUC: ${variantSnap.rocAuc.toFixed(3)} (delta: ${delta.rocAuc >= 0 ? '+' : ''}${delta.rocAuc.toFixed(3)})`);
+    console.log(`  Baseline Silhouette: ${baselineSnap.silhouetteScore.toFixed(3)}`);
+    console.log(`  Filtered Silhouette: ${variantSnap.silhouetteScore.toFixed(3)} (delta: ${delta.silhouetteScore >= 0 ? '+' : ''}${delta.silhouetteScore.toFixed(3)})`);
+    return {
+        name: 'boilerplate-filter',
+        description: `Stripped boilerplate from ${affectedCount}/${chunks.length} chunks before embedding.`,
+        baseline: baselineSnap,
+        variant: variantSnap,
+        delta,
+    };
+}
+//# sourceMappingURL=boilerplate-filter.js.map

package/dist/eval/experiments/boilerplate-filter.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"boilerplate-filter.js","sourceRoot":"","sources":["../../../src/eval/experiments/boilerplate-filter.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAIH,OAAO,EACL,cAAc,EACd,kBAAkB,EAClB,UAAU,GAEX,MAAM,uBAAuB,CAAC;AAC/B,OAAO,EAAE,YAAY,EAAyB,MAAM,YAAY,CAAC;AAEjE,MAAM,QAAQ,GAAG,YAAY,CAAC;AAE9B;;;GAGG;AACH,MAAM,oBAAoB,GAAG;IAC3B,qEAAqE;IACrE,4DAA4D;IAC5D,mEAAmE;IACnE,qDAAqD;IACrD,4CAA4C;CAC7C,CAAC;AAEF;;;GAGG;AACH,MAAM,UAAU,gBAAgB,CAAC,IAAY;IAC3C,IAAI,QAAQ,GAAG,IAAI,CAAC;IACpB,IAAI,QAAQ,GAAG,KAAK,CAAC;IAErB,KAAK,MAAM,OAAO,IAAI,oBAAoB,EAAE,CAAC;QAC3C,MAAM,MAAM,GAAG,QAAQ,CAAC;QACxB,QAAQ,GAAG,QAAQ,CAAC,OAAO,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC;QACzC,IAAI,QAAQ,KAAK,MAAM;YAAE,QAAQ,GAAG,IAAI,CAAC;IAC3C,CAAC;IAED,iCAAiC;IACjC,QAAQ,GAAG,QAAQ,CAAC,OAAO,CAAC,SAAS,EAAE,MAAM,CAAC,CAAC,IAAI,EAAE,CAAC;IAEtD,OAAO,EAAE,IAAI,EAAE,QAAQ,EAAE,QAAQ,EAAE,CAAC;AACtC,CAAC;AAED;;GAEG;AACH,MAAM,CAAC,KAAK,UAAU,wBAAwB,CAC5C,MAAe,EACf,KAAoB,EACpB,cAAkC;IAElC,OAAO,CAAC,GAAG,CAAC,+CAA+C,CAAC,CAAC;IAE7D,uBAAuB;IACvB,IAAI,QAA2B,CAAC;IAChC,IAAI,cAAc,EAAE,CAAC;QACnB,OAAO,CAAC,GAAG,CAAC,sCAAsC,CAAC,CAAC;QACpD,QAAQ,GAAG,cAAc,CAAC;IAC5B,CAAC;SAAM,CAAC;QACN,OAAO,CAAC,GAAG,CAAC,oCAAoC,CAAC,CAAC;QAClD,QAAQ,GAAG,MAAM,cAAc,CAAC,QAAQ,EAAE,MAAM,EAAE,KAAK,CAAC,CAAC;IAC3D,CAAC;IAED,sCAAsC;IACtC,IAAI,aAAa,GAAG,CAAC,CAAC;IACtB,MAAM,aAAa,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE;QACrC,MAAM,EAAE,IAAI,EAAE,QAAQ,EAAE,GAAG,gBAAgB,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC;QACpD,IAAI,QAAQ;YAAE,aAAa,EAAE,CAAC;QAC9B,OAAO,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,EAAE,IAAI,EAAE,CAAC;IAC5B,CAAC,CAAC,CAAC;IAEH,OAAO,CAAC,GAAG,CACT,KAAK,aAAa,IAAI,MAAM,CAAC,MAAM,gCAAgC;QACnE,IAAI,CAAC,CAAC,aAAa,GAAG,MAAM,CAAC,MAAM,CAAC,GAAG,GAAG,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,IAAI,CAC3D,CAAC;IAEF,IAAI,aAAa,KAAK,CAAC,EAAE,CAAC;QACxB,OAAO,CAAC,GAAG,CAAC,wDAAwD,CAAC,CAAC;QACtE,MAAM,IAAI,GAAG,UAAU,CAAC,QAAQ,EAAE,MAAM,CAAC,MAAM,CAAC,CAAC;QACjD,OAAO;YACL,IAAI,EAAE,oBAAoB;YAC1B,WAAW,EAAE,gEAAgE;YAC7E,QAAQ,EAAE,IAAI;YACd,OAAO,EAAE,IAAI;YACb,KAAK,EAAE,YAAY,CAAC,IAAI,EAAE,IAAI,CAAC;SAChC,CAAC;IACJ,CAAC;IAED,yBAAyB;IACzB,OAAO,CAAC,GAAG,CAAC,iCAAiC,CAAC,CAAC;IAC/C,MAAM,OAAO,GAAG,MAAM,kBAAkB,CAAC,QAAQ,EAAE,aAAa,EAAE,KAAK,CAAC,CAAC;IAEzE,MAAM,YAAY,GAAG,UAAU,CAAC,QAAQ,EAAE,MAAM,CAAC,MAAM,CAAC,CAAC;IACzD,MAAM,WAAW,GAAG,UAAU,CAAC,OAAO,EAAE,MAAM,CAAC,MAAM,CAAC,CAAC;IACvD,MAAM,KAAK,GAAG,YAAY,CAAC,YAAY,EAAE,WAAW,CAAC,CAAC;IAEtD,OAAO,CAAC,GAAG,CAAC,uBAAuB,YAAY,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;IACrE,OAAO,CAAC,GAAG,CAAC,uBAAuB,WAAW,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,YAAY,KAAK,CAAC,MAAM,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,GAAG,KAAK,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC;IACvI,OAAO,CAAC,GAAG,CAAC,0BAA0B,YAAY,CAAC,eAAe,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;IACjF,OAAO,CAAC,GAAG,CAAC,0BAA0B,WAAW,CAAC,eAAe,CAAC,OAAO,CAAC,CAAC,CAAC,YAAY,KAAK,CAAC,eAAe,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,GAAG,KAAK,CAAC,eAAe,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC;IAErK,OAAO;QACL,IAAI,EAAE,oBAAoB;QAC1B,WAAW,EAAE,6BAA6B,aAAa,IAAI,MAAM,CAAC,MAAM,2BAA2B;QACnG,QAAQ,EAAE,YAAY;QACtB,OAAO,EAAE,WAAW;QACpB,KAAK;KACN,CAAC;AACJ,CAAC"}

package/dist/eval/experiments/cluster-threshold/index.d.ts

@@ -0,0 +1,6 @@
+/**
+ * Cluster threshold experiment exports.
+ */
+export { runThresholdSweep } from './run-threshold-sweep.js';
+export type { ThresholdResult, ThresholdSweepResult } from './run-threshold-sweep.js';
+//# sourceMappingURL=index.d.ts.map

package/dist/eval/experiments/cluster-threshold/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../../src/eval/experiments/cluster-threshold/index.ts"],"names":[],"mappings":"AAAA;;GAEG;AAEH,OAAO,EAAE,iBAAiB,EAAE,MAAM,0BAA0B,CAAC;AAC7D,YAAY,EAAE,eAAe,EAAE,oBAAoB,EAAE,MAAM,0BAA0B,CAAC"}

package/dist/eval/experiments/cluster-threshold/index.js

@@ -0,0 +1,5 @@
+/**
+ * Cluster threshold experiment exports.
+ */
+export { runThresholdSweep } from './run-threshold-sweep.js';
+//# sourceMappingURL=index.js.map

package/dist/eval/experiments/cluster-threshold/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../../src/eval/experiments/cluster-threshold/index.ts"],"names":[],"mappings":"AAAA;;GAEG;AAEH,OAAO,EAAE,iBAAiB,EAAE,MAAM,0BAA0B,CAAC"}

package/dist/eval/experiments/cluster-threshold/run-threshold-sweep.d.ts

@@ -0,0 +1,58 @@
+/**
+ * Phase 0.1: Cluster Assignment Threshold Sweep
+ *
+ * Find optimal angular distance threshold for assigning chunks to clusters.
+ *
+ * Method:
+ * 1. Load embedding benchmark corpus
+ * 2. Run HDBSCAN to get cluster assignments
+ * 3. Compute pairwise angular distances within same-cluster vs cross-cluster
+ * 4. Sweep threshold: 0.2, 0.3, 0.4, 0.5, 0.6
+ * 5. Measure precision/recall of cluster membership prediction
+ */
+/**
+ * Result for a single threshold value.
+ */
+export interface ThresholdResult {
+    threshold: number;
+    precision: number;
+    recall: number;
+    f1: number;
+    truePositives: number;
+    falsePositives: number;
+    falseNegatives: number;
+    trueNegatives: number;
+}
+/**
+ * Complete sweep results.
+ */
+export interface ThresholdSweepResult {
+    generatedAt: string;
+    corpusSize: number;
+    numClusters: number;
+    noiseRatio: number;
+    silhouetteScore: number;
+    thresholds: ThresholdResult[];
+    recommendedThreshold: number;
+    withinClusterDistances: DistanceStats;
+    crossClusterDistances: DistanceStats;
+}
+interface DistanceStats {
+    mean: number;
+    median: number;
+    std: number;
+    min: number;
+    max: number;
+    count: number;
+}
+/**
+ * Run the cluster threshold sweep experiment.
+ */
+export declare function runThresholdSweep(corpusPath: string, options?: {
+    thresholds?: number[];
+    minClusterSize?: number;
+    embeddingModel?: string;
+    verbose?: boolean;
+}): Promise<ThresholdSweepResult>;
+export {};
+//# sourceMappingURL=run-threshold-sweep.d.ts.map

package/dist/eval/experiments/cluster-threshold/run-threshold-sweep.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"run-threshold-sweep.d.ts","sourceRoot":"","sources":["../../../../src/eval/experiments/cluster-threshold/run-threshold-sweep.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AASH;;GAEG;AACH,MAAM,WAAW,eAAe;IAC9B,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,EAAE,MAAM,CAAC;IAClB,MAAM,EAAE,MAAM,CAAC;IACf,EAAE,EAAE,MAAM,CAAC;IACX,aAAa,EAAE,MAAM,CAAC;IACtB,cAAc,EAAE,MAAM,CAAC;IACvB,cAAc,EAAE,MAAM,CAAC;IACvB,aAAa,EAAE,MAAM,CAAC;CACvB;AAED;;GAEG;AACH,MAAM,WAAW,oBAAoB;IACnC,WAAW,EAAE,MAAM,CAAC;IACpB,UAAU,EAAE,MAAM,CAAC;IACnB,WAAW,EAAE,MAAM,CAAC;IACpB,UAAU,EAAE,MAAM,CAAC;IACnB,eAAe,EAAE,MAAM,CAAC;IACxB,UAAU,EAAE,eAAe,EAAE,CAAC;IAC9B,oBAAoB,EAAE,MAAM,CAAC;IAC7B,sBAAsB,EAAE,aAAa,CAAC;IACtC,qBAAqB,EAAE,aAAa,CAAC;CACtC;AAED,UAAU,aAAa;IACrB,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,EAAE,MAAM,CAAC;IACf,GAAG,EAAE,MAAM,CAAC;IACZ,GAAG,EAAE,MAAM,CAAC;IACZ,GAAG,EAAE,MAAM,CAAC;IACZ,KAAK,EAAE,MAAM,CAAC;CACf;AAED;;GAEG;AACH,wBAAsB,iBAAiB,CACrC,UAAU,EAAE,MAAM,EAClB,OAAO,GAAE;IACP,UAAU,CAAC,EAAE,MAAM,EAAE,CAAC;IACtB,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,OAAO,CAAC,EAAE,OAAO,CAAC;CACd,GACL,OAAO,CAAC,oBAAoB,CAAC,CAsM/B"}

package/dist/eval/experiments/cluster-threshold/run-threshold-sweep.js

@@ -0,0 +1,246 @@
+/**
+ * Phase 0.1: Cluster Assignment Threshold Sweep
+ *
+ * Find optimal angular distance threshold for assigning chunks to clusters.
+ *
+ * Method:
+ * 1. Load embedding benchmark corpus
+ * 2. Run HDBSCAN to get cluster assignments
+ * 3. Compute pairwise angular distances within same-cluster vs cross-cluster
+ * 4. Sweep threshold: 0.2, 0.3, 0.4, 0.5, 0.6
+ * 5. Measure precision/recall of cluster membership prediction
+ */
+import { readFile } from 'fs/promises';
+import { Embedder } from '../../../models/embedder.js';
+import { getModel } from '../../../models/model-registry.js';
+import { clusterEmbeddings, getClusterMembership } from '../../cluster-evaluator.js';
+import { angularDistance } from '../../../utils/angular-distance.js';
+/**
+ * Run the cluster threshold sweep experiment.
+ */
+export async function runThresholdSweep(corpusPath, options = {}) {
+    const { thresholds = [0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5, 0.55, 0.6], minClusterSize = 4, embeddingModel = 'jina-small', verbose = true, } = options;
+    if (verbose) {
+        console.log('\n=== Cluster Threshold Sweep Experiment ===\n');
+        console.log(`Loading corpus from ${corpusPath}...`);
+    }
+    // Load corpus
+    const corpusJson = await readFile(corpusPath, 'utf-8');
+    const parsed = JSON.parse(corpusJson);
+    // Handle both { chunks: [...] } and plain array formats
+    const chunks = Array.isArray(parsed) ? parsed : parsed.chunks;
+    if (verbose) {
+        console.log(`Loaded ${chunks.length} chunks`);
+        console.log(`\nEmbedding with ${embeddingModel}...`);
+    }
+    // Embed all chunks
+    const embedder = new Embedder();
+    await embedder.load(getModel(embeddingModel));
+    const embeddings = [];
+    for (let i = 0; i < chunks.length; i++) {
+        if (verbose && i % 50 === 0) {
+            process.stdout.write(`\r  Embedding: ${i}/${chunks.length}`);
+        }
+        const result = await embedder.embed(chunks[i].text, false);
+        embeddings.push(result.embedding);
+    }
+    await embedder.dispose();
+    if (verbose) {
+        console.log(`\r  Embedding: ${chunks.length}/${chunks.length} complete`);
+        console.log(`\nRunning HDBSCAN (minClusterSize=${minClusterSize})...`);
+    }
+    // Run HDBSCAN
+    const clusterResult = clusterEmbeddings(embeddings, { minClusterSize });
+    const membership = getClusterMembership(clusterResult.labels);
+    if (verbose) {
+        console.log(`  Clusters found: ${clusterResult.numClusters}`);
+        console.log(`  Noise ratio: ${(clusterResult.noiseRatio * 100).toFixed(1)}%`);
+    }
+    // Compute pairwise distances
+    if (verbose) {
+        console.log(`\nComputing pairwise distances...`);
+    }
+    const withinClusterDists = [];
+    const crossClusterDists = [];
+    // Build cluster assignments (excluding noise)
+    const clusterAssignments = new Map(); // chunkIndex -> clusterId
+    for (const [clusterId, indices] of membership) {
+        if (clusterId < 0)
+            continue; // Skip noise
+        for (const idx of indices) {
+            clusterAssignments.set(idx, clusterId);
+        }
+    }
+    // Sample pairs to avoid O(n²) explosion
+    const MAX_PAIRS = 50000;
+    const assignedIndices = [...clusterAssignments.keys()];
+    // Within-cluster pairs
+    for (const [clusterId, indices] of membership) {
+        if (clusterId < 0)
+            continue;
+        for (let i = 0; i < indices.length; i++) {
+            for (let j = i + 1; j < indices.length; j++) {
+                const dist = angularDistance(embeddings[indices[i]], embeddings[indices[j]]);
+                withinClusterDists.push(dist);
+                if (withinClusterDists.length >= MAX_PAIRS)
+                    break;
+            }
+            if (withinClusterDists.length >= MAX_PAIRS)
+                break;
+        }
+        if (withinClusterDists.length >= MAX_PAIRS)
+            break;
+    }
+    // Cross-cluster pairs (sample)
+    const numCrossSamples = Math.min(MAX_PAIRS, assignedIndices.length * 10);
+    for (let s = 0; s < numCrossSamples; s++) {
+        const i = assignedIndices[Math.floor(Math.random() * assignedIndices.length)];
+        const j = assignedIndices[Math.floor(Math.random() * assignedIndices.length)];
+        if (i === j)
+            continue;
+        if (clusterAssignments.get(i) === clusterAssignments.get(j))
+            continue;
+        const dist = angularDistance(embeddings[i], embeddings[j]);
+        crossClusterDists.push(dist);
+    }
+    if (verbose) {
+        console.log(`  Within-cluster pairs: ${withinClusterDists.length}`);
+        console.log(`  Cross-cluster pairs: ${crossClusterDists.length}`);
+    }
+    // Compute distance statistics
+    const withinStats = computeStats(withinClusterDists);
+    const crossStats = computeStats(crossClusterDists);
+    if (verbose) {
+        console.log(`\n  Within-cluster distance: mean=${withinStats.mean.toFixed(3)}, median=${withinStats.median.toFixed(3)}`);
+        console.log(`  Cross-cluster distance: mean=${crossStats.mean.toFixed(3)}, median=${crossStats.median.toFixed(3)}`);
+    }
+    // Compute silhouette score (simplified)
+    const silhouette = computeSimplifiedSilhouette(embeddings, clusterResult.labels);
+    if (verbose) {
+        console.log(`  Silhouette score: ${silhouette.toFixed(3)}`);
+    }
+    // Sweep thresholds
+    if (verbose) {
+        console.log(`\n--- Threshold Sweep ---`);
+        console.log('Threshold | Precision | Recall | F1     | TP     | FP     | FN');
+        console.log('-'.repeat(70));
+    }
+    const thresholdResults = [];
+    for (const threshold of thresholds) {
+        // For each pair, predict "same cluster" if distance < threshold
+        // True positive: predicted same, actually same
+        // False positive: predicted same, actually different
+        // False negative: predicted different, actually same
+        // True negative: predicted different, actually different
+        let tp = 0, fp = 0, fn = 0, tn = 0;
+        // Within-cluster (should predict same)
+        for (const dist of withinClusterDists) {
+            if (dist < threshold) {
+                tp++;
+            }
+            else {
+                fn++;
+            }
+        }
+        // Cross-cluster (should predict different)
+        for (const dist of crossClusterDists) {
+            if (dist < threshold) {
+                fp++;
+            }
+            else {
+                tn++;
+            }
+        }
+        const precision = tp > 0 ? tp / (tp + fp) : 0;
+        const recall = tp > 0 ? tp / (tp + fn) : 0;
+        const f1 = precision + recall > 0 ? 2 * precision * recall / (precision + recall) : 0;
+        thresholdResults.push({
+            threshold,
+            precision,
+            recall,
+            f1,
+            truePositives: tp,
+            falsePositives: fp,
+            falseNegatives: fn,
+            trueNegatives: tn,
+        });
+        if (verbose) {
+            console.log(`${threshold.toFixed(2).padEnd(9)} | ${precision.toFixed(3).padEnd(9)} | ${recall.toFixed(3).padEnd(6)} | ${f1.toFixed(3).padEnd(6)} | ${tp.toString().padEnd(6)} | ${fp.toString().padEnd(6)} | ${fn}`);
+        }
+    }
+    // Find optimal threshold (highest F1)
+    const best = thresholdResults.reduce((a, b) => a.f1 > b.f1 ? a : b);
+    if (verbose) {
+        console.log('-'.repeat(70));
+        console.log(`\nRecommended threshold: ${best.threshold} (F1=${best.f1.toFixed(3)})`);
+    }
+    return {
+        generatedAt: new Date().toISOString(),
+        corpusSize: chunks.length,
+        numClusters: clusterResult.numClusters,
+        noiseRatio: clusterResult.noiseRatio,
+        silhouetteScore: silhouette,
+        thresholds: thresholdResults,
+        recommendedThreshold: best.threshold,
+        withinClusterDistances: withinStats,
+        crossClusterDistances: crossStats,
+    };
+}
+/**
+ * Compute basic statistics for an array of numbers.
+ */
+function computeStats(values) {
+    if (values.length === 0) {
+        return { mean: 0, median: 0, std: 0, min: 0, max: 0, count: 0 };
+    }
+    const sorted = [...values].sort((a, b) => a - b);
+    const mean = values.reduce((a, b) => a + b, 0) / values.length;
+    const median = sorted[Math.floor(sorted.length / 2)];
+    const variance = values.reduce((sum, v) => sum + (v - mean) ** 2, 0) / values.length;
+    const std = Math.sqrt(variance);
+    return {
+        mean,
+        median,
+        std,
+        min: sorted[0],
+        max: sorted[sorted.length - 1],
+        count: values.length,
+    };
+}
+/**
+ * Compute simplified silhouette score.
+ */
+function computeSimplifiedSilhouette(embeddings, labels) {
+    const clusteredIndices = labels
+        .map((l, i) => ({ label: l, index: i }))
+        .filter((x) => x.label >= 0);
+    if (clusteredIndices.length < 2)
+        return 0;
+    let totalSilhouette = 0;
+    const sampleSize = Math.min(500, clusteredIndices.length);
+    // Sample points for efficiency
+    const sampled = clusteredIndices
+        .sort(() => Math.random() - 0.5)
+        .slice(0, sampleSize);
+    for (const { label, index } of sampled) {
+        // a(i) = mean distance to same cluster
+        const sameCluster = clusteredIndices.filter((x) => x.label === label && x.index !== index);
+        if (sameCluster.length === 0)
+            continue;
+        const a = sameCluster.reduce((sum, x) => sum + angularDistance(embeddings[index], embeddings[x.index]), 0) / sameCluster.length;
+        // b(i) = min mean distance to other clusters
+        const otherLabels = [...new Set(clusteredIndices.map((x) => x.label))].filter((l) => l !== label);
+        if (otherLabels.length === 0)
+            continue;
+        let b = Infinity;
+        for (const otherLabel of otherLabels) {
+            const otherCluster = clusteredIndices.filter((x) => x.label === otherLabel);
+            const meanDist = otherCluster.reduce((sum, x) => sum + angularDistance(embeddings[index], embeddings[x.index]), 0) / otherCluster.length;
+            b = Math.min(b, meanDist);
+        }
+        const s = (b - a) / Math.max(a, b);
+        totalSilhouette += s;
+    }
+    return totalSilhouette / sampled.length;
+}
+//# sourceMappingURL=run-threshold-sweep.js.map

package/dist/eval/experiments/cluster-threshold/run-threshold-sweep.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"run-threshold-sweep.js","sourceRoot":"","sources":["../../../../src/eval/experiments/cluster-threshold/run-threshold-sweep.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAEH,OAAO,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAC;AACvC,OAAO,EAAE,QAAQ,EAAE,MAAM,6BAA6B,CAAC;AACvD,OAAO,EAAE,QAAQ,EAAE,MAAM,mCAAmC,CAAC;AAC7D,OAAO,EAAE,iBAAiB,EAAE,oBAAoB,EAAE,MAAM,4BAA4B,CAAC;AACrF,OAAO,EAAE,eAAe,EAAE,MAAM,oCAAoC,CAAC;AAyCrE;;GAEG;AACH,MAAM,CAAC,KAAK,UAAU,iBAAiB,CACrC,UAAkB,EAClB,UAKI,EAAE;IAEN,MAAM,EACJ,UAAU,GAAG,CAAC,GAAG,EAAE,IAAI,EAAE,GAAG,EAAE,IAAI,EAAE,GAAG,EAAE,IAAI,EAAE,GAAG,EAAE,IAAI,EAAE,GAAG,CAAC,EAC9D,cAAc,GAAG,CAAC,EAClB,cAAc,GAAG,YAAY,EAC7B,OAAO,GAAG,IAAI,GACf,GAAG,OAAO,CAAC;IAEZ,IAAI,OAAO,EAAE,CAAC;QACZ,OAAO,CAAC,GAAG,CAAC,gDAAgD,CAAC,CAAC;QAC9D,OAAO,CAAC,GAAG,CAAC,uBAAuB,UAAU,KAAK,CAAC,CAAC;IACtD,CAAC;IAED,cAAc;IACd,MAAM,UAAU,GAAG,MAAM,QAAQ,CAAC,UAAU,EAAE,OAAO,CAAC,CAAC;IACvD,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,CAAC;IACtC,wDAAwD;IACxD,MAAM,MAAM,GAAY,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC;IAEvE,IAAI,OAAO,EAAE,CAAC;QACZ,OAAO,CAAC,GAAG,CAAC,UAAU,MAAM,CAAC,MAAM,SAAS,CAAC,CAAC;QAC9C,OAAO,CAAC,GAAG,CAAC,oBAAoB,cAAc,KAAK,CAAC,CAAC;IACvD,CAAC;IAED,mBAAmB;IACnB,MAAM,QAAQ,GAAG,IAAI,QAAQ,EAAE,CAAC;IAChC,MAAM,QAAQ,CAAC,IAAI,CAAC,QAAQ,CAAC,cAAc,CAAC,CAAC,CAAC;IAE9C,MAAM,UAAU,GAAe,EAAE,CAAC;IAClC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QACvC,IAAI,OAAO,IAAI,CAAC,GAAG,EAAE,KAAK,CAAC,EAAE,CAAC;YAC5B,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC,kBAAkB,CAAC,IAAI,MAAM,CAAC,MAAM,EAAE,CAAC,CAAC;QAC/D,CAAC;QACD,MAAM,MAAM,GAAG,MAAM,QAAQ,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,KAAK,CAAC,CAAC;QAC3D,UAAU,CAAC,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,CAAC;IACpC,CAAC;IAED,MAAM,QAAQ,CAAC,OAAO,EAAE,CAAC;IAEzB,IAAI,OAAO,EAAE,CAAC;QACZ,OAAO,CAAC,GAAG,CAAC,kBAAkB,MAAM,CAAC,MAAM,IAAI,MAAM,CAAC,MAAM,WAAW,CAAC,CAAC;QACzE,OAAO,CAAC,GAAG,CAAC,qCAAqC,cAAc,MAAM,CAAC,CAAC;IACzE,CAAC;IAED,cAAc;IACd,MAAM,aAAa,GAAG,iBAAiB,CAAC,UAAU,EAAE,EAAE,cAAc,EAAE,CAAC,CAAC;IACxE,MAAM,UAAU,GAAG,oBAAoB,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC;IAE9D,IAAI,OAAO,EAAE,CAAC;QACZ,OAAO,CAAC,GAAG,CAAC,qBAAqB,aAAa,CAAC,WAAW,EAAE,CAAC,CAAC;QAC9D,OAAO,CAAC,GAAG,CAAC,kBAAkB,CAAC,aAAa,CAAC,UAAU,GAAG,GAAG,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC;IAChF,CAAC;IAED,6BAA6B;IAC7B,IAAI,OAAO,EAAE,CAAC;QACZ,OAAO,CAAC,GAAG,CAAC,mCAAmC,CAAC,CAAC;IACnD,CAAC;IAED,MAAM,kBAAkB,GAAa,EAAE,CAAC;IACxC,MAAM,iBAAiB,GAAa,EAAE,CAAC;IAEvC,8CAA8C;IAC9C,MAAM,kBAAkB,GAAG,IAAI,GAAG,EAAkB,CAAC,CAAC,0BAA0B;IAChF,KAAK,MAAM,CAAC,SAAS,EAAE,OAAO,CAAC,IAAI,UAAU,EAAE,CAAC;QAC9C,IAAI,SAAS,GAAG,CAAC;YAAE,SAAS,CAAC,aAAa;QAC1C,KAAK,MAAM,GAAG,IAAI,OAAO,EAAE,CAAC;YAC1B,kBAAkB,CAAC,GAAG,CAAC,GAAG,EAAE,SAAS,CAAC,CAAC;QACzC,CAAC;IACH,CAAC;IAED,wCAAwC;IACxC,MAAM,SAAS,GAAG,KAAK,CAAC;IACxB,MAAM,eAAe,GAAG,CAAC,GAAG,kBAAkB,CAAC,IAAI,EAAE,CAAC,CAAC;IAEvD,uBAAuB;IACvB,KAAK,MAAM,CAAC,SAAS,EAAE,OAAO,CAAC,IAAI,UAAU,EAAE,CAAC;QAC9C,IAAI,SAAS,GAAG,CAAC;YAAE,SAAS;QAC5B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACxC,KAAK,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;gBAC5C,MAAM,IAAI,GAAG,eAAe,CAAC,UAAU,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,EAAE,UAAU,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;gBAC7E,kBAAkB,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;gBAC9B,IAAI,kBAAkB,CAAC,MAAM,IAAI,SAAS;oBAAE,MAAM;YACpD,CAAC;YACD,IAAI,kBAAkB,CAAC,MAAM,IAAI,SAAS;gBAAE,MAAM;QACpD,CAAC;QACD,IAAI,kBAAkB,CAAC,MAAM,IAAI,SAAS;YAAE,MAAM;IACpD,CAAC;IAED,+BAA+B;IAC/B,MAAM,eAAe,GAAG,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,eAAe,CAAC,MAAM,GAAG,EAAE,CAAC,CAAC;IACzE,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,eAAe,EAAE,CAAC,EAAE,EAAE,CAAC;QACzC,MAAM,CAAC,GAAG,eAAe,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,MAAM,EAAE,GAAG,eAAe,CAAC,MAAM,CAAC,CAAC,CAAC;QAC9E,MAAM,CAAC,GAAG,eAAe,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,MAAM,EAAE,GAAG,eAAe,CAAC,MAAM,CAAC,CAAC,CAAC;QAC9E,IAAI,CAAC,KAAK,CAAC;YAAE,SAAS;QACtB,IAAI,kBAAkB,CAAC,GAAG,CAAC,CAAC,CAAC,KAAK,kBAAkB,CAAC,GAAG,CAAC,CAAC,CAAC;YAAE,SAAS;QAEtE,MAAM,IAAI,GAAG,eAAe,CAAC,UAAU,CAAC,CAAC,CAAC,EAAE,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC;QAC3D,iBAAiB,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IAC/B,CAAC;IAED,IAAI,OAAO,EAAE,CAAC;QACZ,OAAO,CAAC,GAAG,CAAC,2BAA2B,kBAAkB,CAAC,MAAM,EAAE,CAAC,CAAC;QACpE,OAAO,CAAC,GAAG,CAAC,0BAA0B,iBAAiB,CAAC,MAAM,EAAE,CAAC,CAAC;IACpE,CAAC;IAED,8BAA8B;IAC9B,MAAM,WAAW,GAAG,YAAY,CAAC,kBAAkB,CAAC,CAAC;IACrD,MAAM,UAAU,GAAG,YAAY,CAAC,iBAAiB,CAAC,CAAC;IAEnD,IAAI,OAAO,EAAE,CAAC;QACZ,OAAO,CAAC,GAAG,CAAC,qCAAqC,WAAW,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,YAAY,WAAW,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;QACzH,OAAO,CAAC,GAAG,CAAC,kCAAkC,UAAU,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,YAAY,UAAU,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;IACtH,CAAC;IAED,wCAAwC;IACxC,MAAM,UAAU,GAAG,2BAA2B,CAAC,UAAU,EAAE,aAAa,CAAC,MAAM,CAAC,CAAC;IAEjF,IAAI,OAAO,EAAE,CAAC;QACZ,OAAO,CAAC,GAAG,CAAC,uBAAuB,UAAU,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;IAC9D,CAAC;IAED,mBAAmB;IACnB,IAAI,OAAO,EAAE,CAAC;QACZ,OAAO,CAAC,GAAG,CAAC,2BAA2B,CAAC,CAAC;QACzC,OAAO,CAAC,GAAG,CAAC,gEAAgE,CAAC,CAAC;QAC9E,OAAO,CAAC,GAAG,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC,CAAC;IAC9B,CAAC;IAED,MAAM,gBAAgB,GAAsB,EAAE,CAAC;IAE/C,KAAK,MAAM,SAAS,IAAI,UAAU,EAAE,CAAC;QACnC,gEAAgE;QAChE,+CAA+C;QAC/C,qDAAqD;QACrD,qDAAqD;QACrD,yDAAyD;QAEzD,IAAI,EAAE,GAAG,CAAC,EAAE,EAAE,GAAG,CAAC,EAAE,EAAE,GAAG,CAAC,EAAE,EAAE,GAAG,CAAC,CAAC;QAEnC,uCAAuC;QACvC,KAAK,MAAM,IAAI,IAAI,kBAAkB,EAAE,CAAC;YACtC,IAAI,IAAI,GAAG,SAAS,EAAE,CAAC;gBACrB,EAAE,EAAE,CAAC;YACP,CAAC;iBAAM,CAAC;gBACN,EAAE,EAAE,CAAC;YACP,CAAC;QACH,CAAC;QAED,2CAA2C;QAC3C,KAAK,MAAM,IAAI,IAAI,iBAAiB,EAAE,CAAC;YACrC,IAAI,IAAI,GAAG,SAAS,EAAE,CAAC;gBACrB,EAAE,EAAE,CAAC;YACP,CAAC;iBAAM,CAAC;gBACN,EAAE,EAAE,CAAC;YACP,CAAC;QACH,CAAC;QAED,MAAM,SAAS,GAAG,EAAE,GAAG,CAAC,CAAC,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,GAAG,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QAC9C,MAAM,MAAM,GAAG,EAAE,GAAG,CAAC,CAAC,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,GAAG,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QAC3C,MAAM,EAAE,GAAG,SAAS,GAAG,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,SAAS,GAAG,MAAM,GAAG,CAAC,SAAS,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QAEtF,gBAAgB,CAAC,IAAI,CAAC;YACpB,SAAS;YACT,SAAS;YACT,MAAM;YACN,EAAE;YACF,aAAa,EAAE,EAAE;YACjB,cAAc,EAAE,EAAE;YAClB,cAAc,EAAE,EAAE;YAClB,aAAa,EAAE,EAAE;SAClB,CAAC,CAAC;QAEH,IAAI,OAAO,EAAE,CAAC;YACZ,OAAO,CAAC,GAAG,CACT,GAAG,SAAS,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,SAAS,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,EAAE,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,EAAE,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,EAAE,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,EAAE,EAAE,CACxM,CAAC;QACJ,CAAC;IACH,CAAC;IAED,sCAAsC;IACtC,MAAM,IAAI,GAAG,gBAAgB,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAEpE,IAAI,OAAO,EAAE,CAAC;QACZ,OAAO,CAAC,GAAG,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC,CAAC;QAC5B,OAAO,CAAC,GAAG,CAAC,4BAA4B,IAAI,CAAC,SAAS,QAAQ,IAAI,CAAC,EAAE,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC;IACvF,CAAC;IAED,OAAO;QACL,WAAW,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;QACrC,UAAU,EAAE,MAAM,CAAC,MAAM;QACzB,WAAW,EAAE,aAAa,CAAC,WAAW;QACtC,UAAU,EAAE,aAAa,CAAC,UAAU;QACpC,eAAe,EAAE,UAAU;QAC3B,UAAU,EAAE,gBAAgB;QAC5B,oBAAoB,EAAE,IAAI,CAAC,SAAS;QACpC,sBAAsB,EAAE,WAAW;QACnC,qBAAqB,EAAE,UAAU;KAClC,CAAC;AACJ,CAAC;AAED;;GAEG;AACH,SAAS,YAAY,CAAC,MAAgB;IACpC,IAAI,MAAM,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACxB,OAAO,EAAE,IAAI,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,GAAG,EAAE,CAAC,EAAE,GAAG,EAAE,CAAC,EAAE,GAAG,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC;IAClE,CAAC;IAED,MAAM,MAAM,GAAG,CAAC,GAAG,MAAM,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACjD,MAAM,IAAI,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,GAAG,MAAM,CAAC,MAAM,CAAC;IAC/D,MAAM,MAAM,GAAG,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC;IACrD,MAAM,QAAQ,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC,GAAG,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC,GAAG,MAAM,CAAC,MAAM,CAAC;IACrF,MAAM,GAAG,GAAG,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;IAEhC,OAAO;QACL,IAAI;QACJ,MAAM;QACN,GAAG;QACH,GAAG,EAAE,MAAM,CAAC,CAAC,CAAC;QACd,GAAG,EAAE,MAAM,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC;QAC9B,KAAK,EAAE,MAAM,CAAC,MAAM;KACrB,CAAC;AACJ,CAAC;AAED;;GAEG;AACH,SAAS,2BAA2B,CAAC,UAAsB,EAAE,MAAgB;IAC3E,MAAM,gBAAgB,GAAG,MAAM;SAC5B,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC,CAAC;SACvC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,IAAI,CAAC,CAAC,CAAC;IAE/B,IAAI,gBAAgB,CAAC,MAAM,GAAG,CAAC;QAAE,OAAO,CAAC,CAAC;IAE1C,IAAI,eAAe,GAAG,CAAC,CAAC;IACxB,MAAM,UAAU,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,gBAAgB,CAAC,MAAM,CAAC,CAAC;IAE1D,+BAA+B;IAC/B,MAAM,OAAO,GAAG,gBAAgB;SAC7B,IAAI,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,MAAM,EAAE,GAAG,GAAG,CAAC;SAC/B,KAAK,CAAC,CAAC,EAAE,UAAU,CAAC,CAAC;IAExB,KAAK,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,IAAI,OAAO,EAAE,CAAC;QACvC,uCAAuC;QACvC,MAAM,WAAW,GAAG,gBAAgB,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,KAAK,KAAK,IAAI,CAAC,CAAC,KAAK,KAAK,KAAK,CAAC,CAAC;QAC3F,IAAI,WAAW,CAAC,MAAM,KAAK,CAAC;YAAE,SAAS;QAEvC,MAAM,CAAC,GAAG,WAAW,CAAC,MAAM,CAC1B,CAAC,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC,GAAG,GAAG,eAAe,CAAC,UAAU,CAAC,KAAK,CAAC,EAAE,UAAU,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,EACzE,CAAC,CACF,GAAG,WAAW,CAAC,MAAM,CAAC;QAEvB,6CAA6C;QAC7C,MAAM,WAAW,GAAG,CAAC,GAAG,IAAI,GAAG,CAAC,gBAAgB,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,KAAK,KAAK,CAAC,CAAC;QAClG,IAAI,WAAW,CAAC,MAAM,KAAK,CAAC;YAAE,SAAS;QAEvC,IAAI,CAAC,GAAG,QAAQ,CAAC;QACjB,KAAK,MAAM,UAAU,IAAI,WAAW,EAAE,CAAC;YACrC,MAAM,YAAY,GAAG,gBAAgB,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,KAAK,UAAU,CAAC,CAAC;YAC5E,MAAM,QAAQ,GAAG,YAAY,CAAC,MAAM,CAClC,CAAC,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC,GAAG,GAAG,eAAe,CAAC,UAAU,CAAC,KAAK,CAAC,EAAE,UAAU,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,EACzE,CAAC,CACF,GAAG,YAAY,CAAC,MAAM,CAAC;YACxB,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,QAAQ,CAAC,CAAC;QAC5B,CAAC;QAED,MAAM,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;QACnC,eAAe,IAAI,CAAC,CAAC;IACvB,CAAC;IAED,OAAO,eAAe,GAAG,OAAO,CAAC,MAAM,CAAC;AAC1C,CAAC"}

package/dist/eval/experiments/code-focused-mode.d.ts

@@ -0,0 +1,20 @@
+/**
+ * Experiment 5: Code-Focused Render Mode
+ *
+ * Question: Does stripping NL commentary and keeping only code-related content help?
+ *
+ * Re-chunks with renderMode:'code-focused' (existing chunker support —
+ * skips non-code tool results), re-embeds with jina-small,
+ * and compares ROC AUC and silhouette.
+ */
+import type { LabeledPair } from '../annotation-schema.js';
+import type { Corpus } from '../corpus-builder.js';
+import { type SingleModelResult } from './single-model-run.js';
+import { type ExperimentResult } from './types.js';
+/**
+ * Run the code-focused render mode experiment.
+ *
+ * Requires the original corpus config to rebuild with different settings.
+ */
+export declare function runCodeFocusedExperiment(originalCorpus: Corpus, originalPairs: LabeledPair[], baselineResult?: SingleModelResult): Promise<ExperimentResult>;
+//# sourceMappingURL=code-focused-mode.d.ts.map

package/dist/eval/experiments/code-focused-mode.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"code-focused-mode.d.ts","sourceRoot":"","sources":["../../../src/eval/experiments/code-focused-mode.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAGH,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,yBAAyB,CAAC;AAC3D,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,sBAAsB,CAAC;AAGnD,OAAO,EAGL,KAAK,iBAAiB,EACvB,MAAM,uBAAuB,CAAC;AAC/B,OAAO,EAAgB,KAAK,gBAAgB,EAAE,MAAM,YAAY,CAAC;AAIjE;;;;GAIG;AACH,wBAAsB,wBAAwB,CAC5C,cAAc,EAAE,MAAM,EACtB,aAAa,EAAE,WAAW,EAAE,EAC5B,cAAc,CAAC,EAAE,iBAAiB,GACjC,OAAO,CAAC,gBAAgB,CAAC,CAuD3B"}