@claude-flow/cli 3.10.4 → 3.10.6

package/dist/src/benchmarks/gaia-hardness/features.js

@@ -0,0 +1,170 @@
+/**
+ * GAIA Hardness Predictor — Feature Extraction (ADR-136 Track Q)
+ *
+ * Extracts a 17-dimensional feature vector from a GaiaQuestion for use
+ * in the linear hardness classifier.  Deliberately avoids external
+ * dependencies (no spacy, no heavy NLP) — all extraction is regex-based
+ * with O(1) per question.
+ *
+ * Feature vector layout (17 dims):
+ *   [0]  question length in characters (normalised / 500)
+ *   [1]  question length in words      (normalised / 100)
+ *   [2]  sentence count                (normalised / 5)
+ *   [3]  question word: "what"         (0/1)
+ *   [4]  question word: "how"          (0/1)
+ *   [5]  question word: "who/when/where" (0/1)
+ *   [6]  question word: "calculate/compute/how many/what percentage" (0/1)
+ *   [7]  has numeric token             (0/1)
+ *   [8]  has year token (4-digit)      (0/1)
+ *   [9]  has comparison keyword        (0/1)
+ *   [10] estimated named-entity count  (normalised / 5)
+ *   [11] digit-token count             (normalised / 5)
+ *   [12] multi-hop signal              (0/1)  — bridge/relative clause pattern
+ *   [13] requires math                 (0/1)  — "how many/much/percentage/calculate/compute"
+ *   [14] temporal chain                (0/1)  — "before/after X happened / since / until"
+ *   [15] tool implication count        (normalised / 4)  — PDF/image/video/URL markers
+ *   [16] file attachment present       (0/1)
+ *
+ * All continuous values are min-max normalised to [0, 1] using fixed
+ * divisors chosen so typical GAIA questions fall in [0.1, 0.9].
+ *
+ * Refs: ADR-136, #2156
+ */
+// ---------------------------------------------------------------------------
+// Feature labels (in order, matches values array)
+// ---------------------------------------------------------------------------
+export const FEATURE_LABELS = [
+    'len_chars_norm',
+    'len_words_norm',
+    'sentence_count_norm',
+    'qword_what',
+    'qword_how',
+    'qword_who_when_where',
+    'qword_calc_compute',
+    'has_numeric',
+    'has_year',
+    'has_comparison',
+    'entity_count_norm',
+    'digit_token_norm',
+    'multi_hop_signal',
+    'requires_math',
+    'temporal_chain',
+    'tool_implication_norm',
+    'has_file_attachment',
+];
+// ---------------------------------------------------------------------------
+// Regex catalogue (compiled once at module load)
+// ---------------------------------------------------------------------------
+// Named entity proxy: capitalised word sequences (not at sentence start)
+// e.g. "Barack Obama", "New York City", "United States of America"
+const RE_ENTITY = /(?<!^|\.\s{1,3})(?:[A-Z][a-z]+(?:\s+[A-Z][a-z]+){0,4})/g;
+const RE_YEAR = /\b(1[0-9]{3}|20[0-9]{2})\b/g;
+const RE_DIGIT_TOKEN = /\b\d[\d,._]*/g;
+const RE_NUMERIC = /\b\d/;
+const RE_COMPARISON = /\b(more|fewer|greater|less|compare|versus|vs\.?|differ|increase|decrease|ratio|proportion)\b/i;
+const RE_MULTI_HOP = /\b(the\s+\w+\s+that\s+|which\s+was\s+|who\s+was\s+|whose\s+|of\s+the\s+\w+\s+that\s+)/i;
+// Requires math: stems without trailing \b to handle "calculate", "multiply", "division", etc.
+const RE_MATH = /\b(how\s+many|how\s+much|what\s+percentage|what\s+fraction|what\s+proportion|calculat|comput|multipli|divis|subtract|sum\s+of|total\s+of|average\s+of|product\s+of|express\s+as\s+a\s+decimal)/i;
+const RE_TEMPORAL = /\b(before|after|since|until|during|between\s+\d|in\s+the\s+year|by\s+the\s+time|at\s+the\s+time)\b/i;
+// Tool implication markers
+const RE_TOOL_FILE = /\b(pdf|\.docx|\.xlsx|\.csv|file|attachment|document|spreadsheet)\b/i;
+const RE_TOOL_IMAGE = /\b(image|photo|picture|screenshot|figure|diagram|chart)\b/i;
+const RE_TOOL_VIDEO = /\b(video|youtube|clip|footage)\b/i;
+const RE_TOOL_URL = /https?:\/\//i;
+// Question-word detection (first word of question, or full-text for calc/compute)
+const RE_QWORD_WHAT = /^(what|which)\b/i;
+const RE_QWORD_HOW = /^how\b/i;
+const RE_QWORD_WHO_WHEN_WHERE = /^(who|when|where|why|whom)\b/i;
+// calc/compute: fires if question contains calculate/compute anywhere (stems, no trailing \b)
+const RE_QWORD_CALC = /\b(calculat|comput|how\s+many|what\s+percentage|how\s+much)/i;
+// Sentence boundary (rough — period/exclamation/question-mark followed by space+capital)
+const RE_SENTENCE = /[.!?]\s+[A-Z]/g;
+// ---------------------------------------------------------------------------
+// Helper: clamp value to [0, 1]
+// ---------------------------------------------------------------------------
+function clamp(v, max) {
+    return Math.min(v / max, 1.0);
+}
+// ---------------------------------------------------------------------------
+// Public API
+// ---------------------------------------------------------------------------
+/**
+ * Extract the 17-dimensional feature vector from a GaiaQuestion.
+ * All features are in [0, 1].  Never throws.
+ */
+export function extractFeatures(q) {
+    const text = q.question ?? '';
+    // ── Syntactic scalars ──────────────────────────────────────────────────
+    const lenChars = text.length;
+    const words = text.trim().split(/\s+/).filter(Boolean);
+    const lenWords = words.length;
+    const sentenceMatches = text.match(RE_SENTENCE);
+    const sentenceCount = 1 + (sentenceMatches ? sentenceMatches.length : 0);
+    // ── Question-word one-hot ─────────────────────────────────────────────
+    const firstWord = text.trim();
+    const qwWhat = RE_QWORD_WHAT.test(firstWord) ? 1 : 0;
+    const qwHow = RE_QWORD_HOW.test(firstWord) ? 1 : 0;
+    const qwWhoWhenWhere = RE_QWORD_WHO_WHEN_WHERE.test(firstWord) ? 1 : 0;
+    const qwCalcCompute = RE_QWORD_CALC.test(text) ? 1 : 0;
+    // ── Lexical booleans ──────────────────────────────────────────────────
+    const hasNumeric = RE_NUMERIC.test(text) ? 1 : 0;
+    // Count year tokens
+    const yearMatches = text.match(RE_YEAR);
+    const hasYear = yearMatches && yearMatches.length > 0 ? 1 : 0;
+    const hasComparison = RE_COMPARISON.test(text) ? 1 : 0;
+    // ── Named entity proxy (capitalised word sequences) ───────────────────
+    // Reset lastIndex for global regex
+    RE_ENTITY.lastIndex = 0;
+    let entityCount = 0;
+    // eslint-disable-next-line no-constant-condition
+    while (true) {
+        const m = RE_ENTITY.exec(text);
+        if (!m)
+            break;
+        entityCount++;
+    }
+    // ── Digit token count ─────────────────────────────────────────────────
+    RE_DIGIT_TOKEN.lastIndex = 0;
+    const digitTokenMatches = text.match(RE_DIGIT_TOKEN);
+    const digitTokenCount = digitTokenMatches ? digitTokenMatches.length : 0;
+    // ── Multi-hop signal ──────────────────────────────────────────────────
+    const multiHopSignal = RE_MULTI_HOP.test(text) ? 1 : 0;
+    // ── Math requirement ──────────────────────────────────────────────────
+    const requiresMath = RE_MATH.test(text) ? 1 : 0;
+    // ── Temporal chain ────────────────────────────────────────────────────
+    const temporalChain = RE_TEMPORAL.test(text) ? 1 : 0;
+    // ── Tool implication (PDF/image/video/URL → file_read/web_browse/image_describe) ──
+    let toolImplication = 0;
+    if (RE_TOOL_FILE.test(text) || q.file_name)
+        toolImplication++;
+    if (RE_TOOL_IMAGE.test(text))
+        toolImplication++;
+    if (RE_TOOL_VIDEO.test(text))
+        toolImplication++;
+    if (RE_TOOL_URL.test(text))
+        toolImplication++;
+    // ── File attachment ───────────────────────────────────────────────────
+    const hasFileAttachment = q.file_name && q.file_name.trim() ? 1 : 0;
+    // ── Assemble normalised vector ────────────────────────────────────────
+    const values = [
+        clamp(lenChars, 500), // [0]
+        clamp(lenWords, 100), // [1]
+        clamp(sentenceCount, 5), // [2]
+        qwWhat, // [3]
+        qwHow, // [4]
+        qwWhoWhenWhere, // [5]
+        qwCalcCompute, // [6]
+        hasNumeric, // [7]
+        hasYear, // [8]
+        hasComparison, // [9]
+        clamp(entityCount, 5), // [10]
+        clamp(digitTokenCount, 5), // [11]
+        multiHopSignal, // [12]
+        requiresMath, // [13]
+        temporalChain, // [14]
+        clamp(toolImplication, 4), // [15]
+        hasFileAttachment ? 1 : 0, // [16]
+    ];
+    return { values, labels: FEATURE_LABELS };
+}
+//# sourceMappingURL=features.js.map

package/dist/src/benchmarks/gaia-hardness/features.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"features.js","sourceRoot":"","sources":["../../../../src/benchmarks/gaia-hardness/features.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+BG;AAeH,8EAA8E;AAC9E,kDAAkD;AAClD,8EAA8E;AAE9E,MAAM,CAAC,MAAM,cAAc,GAAsB;IAC/C,gBAAgB;IAChB,gBAAgB;IAChB,qBAAqB;IACrB,YAAY;IACZ,WAAW;IACX,sBAAsB;IACtB,oBAAoB;IACpB,aAAa;IACb,UAAU;IACV,gBAAgB;IAChB,mBAAmB;IACnB,kBAAkB;IAClB,kBAAkB;IAClB,eAAe;IACf,gBAAgB;IAChB,uBAAuB;IACvB,qBAAqB;CACb,CAAC;AAEX,8EAA8E;AAC9E,iDAAiD;AACjD,8EAA8E;AAE9E,yEAAyE;AACzE,mEAAmE;AACnE,MAAM,SAAS,GAAG,yDAAyD,CAAC;AAE5E,MAAM,OAAO,GAAG,6BAA6B,CAAC;AAC9C,MAAM,cAAc,GAAG,eAAe,CAAC;AACvC,MAAM,UAAU,GAAG,MAAM,CAAC;AAC1B,MAAM,aAAa,GAAG,+FAA+F,CAAC;AACtH,MAAM,YAAY,GAAG,wFAAwF,CAAC;AAC9G,+FAA+F;AAC/F,MAAM,OAAO,GAAG,iMAAiM,CAAC;AAClN,MAAM,WAAW,GAAG,qGAAqG,CAAC;AAE1H,2BAA2B;AAC3B,MAAM,YAAY,GAAG,qEAAqE,CAAC;AAC3F,MAAM,aAAa,GAAG,4DAA4D,CAAC;AACnF,MAAM,aAAa,GAAG,mCAAmC,CAAC;AAC1D,MAAM,WAAW,GAAG,cAAc,CAAC;AAEnC,kFAAkF;AAClF,MAAM,aAAa,GAAG,kBAAkB,CAAC;AACzC,MAAM,YAAY,GAAG,SAAS,CAAC;AAC/B,MAAM,uBAAuB,GAAG,+BAA+B,CAAC;AAChE,8FAA8F;AAC9F,MAAM,aAAa,GAAG,8DAA8D,CAAC;AAErF,yFAAyF;AACzF,MAAM,WAAW,GAAG,gBAAgB,CAAC;AAErC,8EAA8E;AAC9E,gCAAgC;AAChC,8EAA8E;AAE9E,SAAS,KAAK,CAAC,CAAS,EAAE,GAAW;IACnC,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,GAAG,EAAE,GAAG,CAAC,CAAC;AAChC,CAAC;AAED,8EAA8E;AAC9E,aAAa;AACb,8EAA8E;AAE9E;;;GAGG;AACH,MAAM,UAAU,eAAe,CAAC,CAAe;IAC7C,MAAM,IAAI,GAAG,CAAC,CAAC,QAAQ,IAAI,EAAE,CAAC;IAE9B,0EAA0E;IAC1E,MAAM,QAAQ,GAAG,IAAI,CAAC,MAAM,CAAC;IAC7B,MAAM,KAAK,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC;IACvD,MAAM,QAAQ,GAAG,KAAK,CAAC,MAAM,CAAC;IAC9B,MAAM,eAAe,GAAG,IAAI,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC;IAChD,MAAM,aAAa,GAAG,CAAC,GAAG,CAAC,eAAe,CAAC,CAAC,CAAC,eAAe,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAEzE,yEAAyE;IACzE,MAAM,SAAS,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;IAC9B,MAAM,MAAM,GAAG,aAAa,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IACrD,MAAM,KAAK,GAAG,YAAY,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IACnD,MAAM,cAAc,GAAG,uBAAuB,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IACvE,MAAM,aAAa,GAAG,aAAa,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAEvD,yEAAyE;IACzE,MAAM,UAAU,GAAG,UAAU,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAEjD,oBAAoB;IACpB,MAAM,WAAW,GAAG,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;IACxC,MAAM,OAAO,GAAG,WAAW,IAAI,WAAW,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAE9D,MAAM,aAAa,GAAG,aAAa,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAEvD,yEAAyE;IACzE,mCAAmC;IACnC,SAAS,CAAC,SAAS,GAAG,CAAC,CAAC;IACxB,IAAI,WAAW,GAAG,CAAC,CAAC;IACpB,iDAAiD;IACjD,OAAO,IAAI,EAAE,CAAC;QACZ,MAAM,CAAC,GAAG,SAAS,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QAC/B,IAAI,CAAC,CAAC;YAAE,MAAM;QACd,WAAW,EAAE,CAAC;IAChB,CAAC;IAED,yEAAyE;IACzE,cAAc,CAAC,SAAS,GAAG,CAAC,CAAC;IAC7B,MAAM,iBAAiB,GAAG,IAAI,CAAC,KAAK,CAAC,cAAc,CAAC,CAAC;IACrD,MAAM,eAAe,GAAG,iBAAiB,CAAC,CAAC,CAAC,iBAAiB,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC;IAEzE,yEAAyE;IACzE,MAAM,cAAc,GAAG,YAAY,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAEvD,yEAAyE;IACzE,MAAM,YAAY,GAAG,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAEhD,yEAAyE;IACzE,MAAM,aAAa,GAAG,WAAW,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAErD,qFAAqF;IACrF,IAAI,eAAe,GAAG,CAAC,CAAC;IACxB,IAAI,YAAY,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,SAAS;QAAE,eAAe,EAAE,CAAC;IAC9D,IAAI,aAAa,CAAC,IAAI,CAAC,IAAI,CAAC;QAAE,eAAe,EAAE,CAAC;IAChD,IAAI,aAAa,CAAC,IAAI,CAAC,IAAI,CAAC;QAAE,eAAe,EAAE,CAAC;IAChD,IAAI,WAAW,CAAC,IAAI,CAAC,IAAI,CAAC;QAAE,eAAe,EAAE,CAAC;IAE9C,yEAAyE;IACzE,MAAM,iBAAiB,GAAG,CAAC,CAAC,SAAS,IAAI,CAAC,CAAC,SAAS,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAEpE,yEAAyE;IACzE,MAAM,MAAM,GAAa;QACvB,KAAK,CAAC,QAAQ,EAAE,GAAG,CAAC,EAAS,MAAM;QACnC,KAAK,CAAC,QAAQ,EAAE,GAAG,CAAC,EAAS,MAAM;QACnC,KAAK,CAAC,aAAa,EAAE,CAAC,CAAC,EAAM,MAAM;QACnC,MAAM,EAAuB,MAAM;QACnC,KAAK,EAAwB,MAAM;QACnC,cAAc,EAAgB,MAAM;QACpC,aAAa,EAAgB,MAAM;QACnC,UAAU,EAAmB,MAAM;QACnC,OAAO,EAAsB,MAAM;QACnC,aAAa,EAAgB,MAAM;QACnC,KAAK,CAAC,WAAW,EAAE,CAAC,CAAC,EAAQ,OAAO;QACpC,KAAK,CAAC,eAAe,EAAE,CAAC,CAAC,EAAI,OAAO;QACpC,cAAc,EAAe,OAAO;QACpC,YAAY,EAAiB,OAAO;QACpC,aAAa,EAAgB,OAAO;QACpC,KAAK,CAAC,eAAe,EAAE,CAAC,CAAC,EAAI,OAAO;QACpC,iBAAiB,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,EAAI,OAAO;KACrC,CAAC;IAEF,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,cAA0B,EAAE,CAAC;AACxD,CAAC"}

package/dist/src/benchmarks/gaia-hardness/predictor.d.ts

@@ -0,0 +1,105 @@
+/**
+ * GAIA Hardness Predictor — Linear Classifier (ADR-136 Track Q)
+ *
+ * Classifies GAIA questions into easy / medium / hard using a
+ * hand-rolled logistic regression (no external ML dependencies).
+ *
+ * Training:
+ *   `predictor.train(labeledData)` — fits weights via gradient descent
+ *   on cross-entropy loss using the 17-dim feature vectors.
+ *
+ * Inference:
+ *   `predictor.predict(question)` — returns difficulty class + confidence
+ *   + a ComputeBudget that drives model/turns/voting choices in gaia-bench.
+ *
+ * Cold-start:
+ *   When untrained (weights = null), classifies everything as "medium".
+ *   This is the correct safe default: no wasted Haiku-on-hard, no missed
+ *   Sonnet-on-easy.
+ *
+ * Compute budget policy (from ADR-136 Track Q research):
+ *   easy   → Haiku,  max 4 turns,  1 attempt
+ *   medium → Sonnet, max 8 turns,  1 attempt
+ *   hard   → Sonnet, max 12 turns, 3-vote (Track A)
+ *
+ * Conservative threshold:
+ *   If in doubt, classify UP (medium→hard preferred over medium→easy).
+ *   `conservativeMode: true` (default) shifts the easy/medium boundary
+ *   so fewer questions fall into "easy".
+ *
+ * Refs: ADR-136, ADR-135, #2156
+ */
+import type { GaiaQuestion } from '../gaia-loader.js';
+import { type FeatureVector } from './features.js';
+export type DifficultyClass = 'easy' | 'medium' | 'hard';
+export interface ComputeBudget {
+    model: 'haiku' | 'sonnet';
+    maxTurns: number;
+    votingAttempts: number;
+}
+export interface PredictionResult {
+    difficulty: DifficultyClass;
+    confidence: number;
+    budget: ComputeBudget;
+    features: FeatureVector;
+}
+export interface LabeledExample {
+    question: GaiaQuestion;
+    wasCorrect: boolean;
+    /** Number of turns the agent used (optional; used to refine label). */
+    turns?: number;
+}
+export declare const COMPUTE_BUDGETS: Record<DifficultyClass, ComputeBudget>;
+export declare class HardnessPredictor {
+    /**
+     * Weight matrix: weights[classIdx][featureIdx].
+     * null = untrained (cold-start: return medium for everything).
+     */
+    private weights;
+    /** Bias terms per class. */
+    private biases;
+    /** Whether conservative mode is active (default: true). */
+    private readonly conservativeMode;
+    constructor(options?: {
+        conservativeMode?: boolean;
+    });
+    /**
+     * Returns true when the predictor has been trained and is ready
+     * to make non-trivial predictions.
+     */
+    get isTrained(): boolean;
+    /**
+     * Train the linear classifier using labelled examples from prior runs.
+     *
+     * Labelling strategy (weak supervision):
+     *   - All correct + turns ≤ median turns  → easy
+     *   - All correct + turns > median turns  → medium
+     *   - Incorrect                           → hard
+     *
+     * With < 10 examples, refuses to train (cold-start is safer).
+     * With 10-50 examples, trains but sets `conservativeMode`-threshold high.
+     */
+    train(labeledData: LabeledExample[]): void;
+    /**
+     * Predict the hardness class of a single GAIA question.
+     *
+     * Cold-start (untrained): returns medium with confidence=0.5.
+     */
+    predict(question: GaiaQuestion): PredictionResult;
+    /**
+     * Export weights as a plain JSON-serialisable object.
+     * Returns null if untrained.
+     */
+    export(): {
+        weights: number[][];
+        biases: number[];
+    } | null;
+    /**
+     * Import previously exported weights.
+     */
+    import(state: {
+        weights: number[][];
+        biases: number[];
+    }): void;
+}
+//# sourceMappingURL=predictor.d.ts.map

package/dist/src/benchmarks/gaia-hardness/predictor.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"predictor.d.ts","sourceRoot":"","sources":["../../../../src/benchmarks/gaia-hardness/predictor.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAC;AACtD,OAAO,EAAmB,KAAK,aAAa,EAAE,MAAM,eAAe,CAAC;AAMpE,MAAM,MAAM,eAAe,GAAG,MAAM,GAAG,QAAQ,GAAG,MAAM,CAAC;AAEzD,MAAM,WAAW,aAAa;IAC5B,KAAK,EAAE,OAAO,GAAG,QAAQ,CAAC;IAC1B,QAAQ,EAAE,MAAM,CAAC;IACjB,cAAc,EAAE,MAAM,CAAC;CACxB;AAED,MAAM,WAAW,gBAAgB;IAC/B,UAAU,EAAE,eAAe,CAAC;IAC5B,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,EAAE,aAAa,CAAC;IACtB,QAAQ,EAAE,aAAa,CAAC;CACzB;AAED,MAAM,WAAW,cAAc;IAC7B,QAAQ,EAAE,YAAY,CAAC;IACvB,UAAU,EAAE,OAAO,CAAC;IACpB,uEAAuE;IACvE,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB;AAMD,eAAO,MAAM,eAAe,EAAE,MAAM,CAAC,eAAe,EAAE,aAAa,CAgBlE,CAAC;AAyCF,qBAAa,iBAAiB;IAC5B;;;OAGG;IACH,OAAO,CAAC,OAAO,CAA2B;IAE1C,4BAA4B;IAC5B,OAAO,CAAC,MAAM,CAAyB;IAEvC,2DAA2D;IAC3D,OAAO,CAAC,QAAQ,CAAC,gBAAgB,CAAU;gBAE/B,OAAO,GAAE;QAAE,gBAAgB,CAAC,EAAE,OAAO,CAAA;KAAO;IAIxD;;;OAGG;IACH,IAAI,SAAS,IAAI,OAAO,CAEvB;IAID;;;;;;;;;;OAUG;IACH,KAAK,CAAC,WAAW,EAAE,cAAc,EAAE,GAAG,IAAI;IA+E1C;;;;OAIG;IACH,OAAO,CAAC,QAAQ,EAAE,YAAY,GAAG,gBAAgB;IAgDjD;;;OAGG;IACH,MAAM,IAAI;QAAE,OAAO,EAAE,MAAM,EAAE,EAAE,CAAC;QAAC,MAAM,EAAE,MAAM,EAAE,CAAA;KAAE,GAAG,IAAI;IAK1D;;OAEG;IACH,MAAM,CAAC,KAAK,EAAE;QAAE,OAAO,EAAE,MAAM,EAAE,EAAE,CAAC;QAAC,MAAM,EAAE,MAAM,EAAE,CAAA;KAAE,GAAG,IAAI;CAc/D"}

package/dist/src/benchmarks/gaia-hardness/predictor.js

@@ -0,0 +1,260 @@
+/**
+ * GAIA Hardness Predictor — Linear Classifier (ADR-136 Track Q)
+ *
+ * Classifies GAIA questions into easy / medium / hard using a
+ * hand-rolled logistic regression (no external ML dependencies).
+ *
+ * Training:
+ *   `predictor.train(labeledData)` — fits weights via gradient descent
+ *   on cross-entropy loss using the 17-dim feature vectors.
+ *
+ * Inference:
+ *   `predictor.predict(question)` — returns difficulty class + confidence
+ *   + a ComputeBudget that drives model/turns/voting choices in gaia-bench.
+ *
+ * Cold-start:
+ *   When untrained (weights = null), classifies everything as "medium".
+ *   This is the correct safe default: no wasted Haiku-on-hard, no missed
+ *   Sonnet-on-easy.
+ *
+ * Compute budget policy (from ADR-136 Track Q research):
+ *   easy   → Haiku,  max 4 turns,  1 attempt
+ *   medium → Sonnet, max 8 turns,  1 attempt
+ *   hard   → Sonnet, max 12 turns, 3-vote (Track A)
+ *
+ * Conservative threshold:
+ *   If in doubt, classify UP (medium→hard preferred over medium→easy).
+ *   `conservativeMode: true` (default) shifts the easy/medium boundary
+ *   so fewer questions fall into "easy".
+ *
+ * Refs: ADR-136, ADR-135, #2156
+ */
+import { extractFeatures } from './features.js';
+// ---------------------------------------------------------------------------
+// Compute budget policy
+// ---------------------------------------------------------------------------
+export const COMPUTE_BUDGETS = {
+    easy: {
+        model: 'haiku',
+        maxTurns: 4,
+        votingAttempts: 1,
+    },
+    medium: {
+        model: 'sonnet',
+        maxTurns: 8,
+        votingAttempts: 1,
+    },
+    hard: {
+        model: 'sonnet',
+        maxTurns: 12,
+        votingAttempts: 3,
+    },
+};
+// ---------------------------------------------------------------------------
+// Internal constants
+// ---------------------------------------------------------------------------
+const FEATURE_DIM = 17;
+const NUM_CLASSES = 3; // easy=0, medium=1, hard=2
+const CLASS_NAMES = ['easy', 'medium', 'hard'];
+// Learning-rate and regularisation tuned for ~300-example datasets
+const LEARNING_RATE = 0.05;
+const REGULARISATION_LAMBDA = 0.01;
+const TRAINING_EPOCHS = 200;
+// Conservative mode: shift the probability threshold so fewer questions
+// are classified as "easy" (reduces risk of underpowering hard questions).
+// With conservativeMode: the easy threshold is 0.6 (not 0.5).
+const EASY_THRESHOLD_CONSERVATIVE = 0.55;
+// ---------------------------------------------------------------------------
+// Math helpers (pure functions, no deps)
+// ---------------------------------------------------------------------------
+function softmax(logits) {
+    const maxLogit = Math.max(...logits);
+    const exps = logits.map((l) => Math.exp(l - maxLogit));
+    const sumExps = exps.reduce((a, b) => a + b, 0);
+    return exps.map((e) => e / sumExps);
+}
+function dot(a, b) {
+    let sum = 0;
+    for (let i = 0; i < a.length; i++)
+        sum += a[i] * b[i];
+    return sum;
+}
+// ---------------------------------------------------------------------------
+// HardnessPredictor
+// ---------------------------------------------------------------------------
+export class HardnessPredictor {
+    /**
+     * Weight matrix: weights[classIdx][featureIdx].
+     * null = untrained (cold-start: return medium for everything).
+     */
+    weights = null;
+    /** Bias terms per class. */
+    biases = null;
+    /** Whether conservative mode is active (default: true). */
+    conservativeMode;
+    constructor(options = {}) {
+        this.conservativeMode = options.conservativeMode ?? true;
+    }
+    /**
+     * Returns true when the predictor has been trained and is ready
+     * to make non-trivial predictions.
+     */
+    get isTrained() {
+        return this.weights !== null;
+    }
+    // ── Training ─────────────────────────────────────────────────────────────
+    /**
+     * Train the linear classifier using labelled examples from prior runs.
+     *
+     * Labelling strategy (weak supervision):
+     *   - All correct + turns ≤ median turns  → easy
+     *   - All correct + turns > median turns  → medium
+     *   - Incorrect                           → hard
+     *
+     * With < 10 examples, refuses to train (cold-start is safer).
+     * With 10-50 examples, trains but sets `conservativeMode`-threshold high.
+     */
+    train(labeledData) {
+        if (labeledData.length < 10) {
+            // Too few examples for meaningful generalisation.
+            this.weights = null;
+            this.biases = null;
+            return;
+        }
+        // Derive labels using weak-supervision strategy.
+        const allTurns = labeledData
+            .filter((d) => d.turns !== undefined)
+            .map((d) => d.turns);
+        const medianTurns = allTurns.length > 0 ? median(allTurns) : 6;
+        const X = [];
+        const y = [];
+        for (const example of labeledData) {
+            const fv = extractFeatures(example.question);
+            X.push(fv.values);
+            let classIdx;
+            if (example.wasCorrect) {
+                const t = example.turns ?? medianTurns;
+                classIdx = t <= medianTurns ? 0 : 1; // easy=0, medium=1
+            }
+            else {
+                classIdx = 2; // hard
+            }
+            y.push(classIdx);
+        }
+        // Initialise weights to 0.
+        const W = Array.from({ length: NUM_CLASSES }, () => new Array(FEATURE_DIM).fill(0));
+        const b = new Array(NUM_CLASSES).fill(0);
+        // Mini-batch gradient descent (batch = full dataset for small sizes).
+        for (let epoch = 0; epoch < TRAINING_EPOCHS; epoch++) {
+            // Accumulate gradients.
+            const dW = Array.from({ length: NUM_CLASSES }, () => new Array(FEATURE_DIM).fill(0));
+            const db = new Array(NUM_CLASSES).fill(0);
+            for (let n = 0; n < X.length; n++) {
+                const x = X[n];
+                const trueClass = y[n];
+                // Compute logits and softmax probabilities.
+                const logits = W.map((w, k) => dot(w, x) + b[k]);
+                const probs = softmax(logits);
+                // Cross-entropy gradient for each class.
+                for (let k = 0; k < NUM_CLASSES; k++) {
+                    const grad = probs[k] - (k === trueClass ? 1 : 0);
+                    for (let f = 0; f < FEATURE_DIM; f++) {
+                        dW[k][f] += grad * x[f];
+                    }
+                    db[k] += grad;
+                }
+            }
+            // Update weights with L2 regularisation.
+            const N = X.length;
+            for (let k = 0; k < NUM_CLASSES; k++) {
+                for (let f = 0; f < FEATURE_DIM; f++) {
+                    W[k][f] -= LEARNING_RATE * (dW[k][f] / N + REGULARISATION_LAMBDA * W[k][f]);
+                }
+                b[k] -= LEARNING_RATE * (db[k] / N);
+            }
+        }
+        this.weights = W;
+        this.biases = b;
+    }
+    // ── Inference ─────────────────────────────────────────────────────────────
+    /**
+     * Predict the hardness class of a single GAIA question.
+     *
+     * Cold-start (untrained): returns medium with confidence=0.5.
+     */
+    predict(question) {
+        const features = extractFeatures(question);
+        if (!this.weights || !this.biases) {
+            // Cold-start fallback: medium for everything.
+            return {
+                difficulty: 'medium',
+                confidence: 0.5,
+                budget: COMPUTE_BUDGETS.medium,
+                features,
+            };
+        }
+        const logits = this.weights.map((w, k) => dot(w, features.values) + this.biases[k]);
+        const probs = softmax(logits);
+        // Conservative mode: down-weight easy probability.
+        let adjustedProbs = [...probs];
+        if (this.conservativeMode) {
+            // Transfer a fraction of easy probability to medium.
+            const easyExcess = Math.max(0, probs[0] - EASY_THRESHOLD_CONSERVATIVE);
+            adjustedProbs[0] = probs[0] - easyExcess;
+            adjustedProbs[1] = probs[1] + easyExcess;
+        }
+        // Re-normalise after adjustment.
+        const sumAdj = adjustedProbs.reduce((a, b) => a + b, 0);
+        adjustedProbs = adjustedProbs.map((p) => p / sumAdj);
+        // Pick argmax.
+        let bestClass = 0;
+        for (let k = 1; k < NUM_CLASSES; k++) {
+            if (adjustedProbs[k] > adjustedProbs[bestClass])
+                bestClass = k;
+        }
+        const difficulty = CLASS_NAMES[bestClass];
+        const confidence = adjustedProbs[bestClass];
+        return {
+            difficulty,
+            confidence,
+            budget: COMPUTE_BUDGETS[difficulty],
+            features,
+        };
+    }
+    // ── Serialisation ─────────────────────────────────────────────────────────
+    /**
+     * Export weights as a plain JSON-serialisable object.
+     * Returns null if untrained.
+     */
+    export() {
+        if (!this.weights || !this.biases)
+            return null;
+        return { weights: this.weights, biases: this.biases };
+    }
+    /**
+     * Import previously exported weights.
+     */
+    import(state) {
+        if (!Array.isArray(state.weights) ||
+            state.weights.length !== NUM_CLASSES ||
+            !Array.isArray(state.biases) ||
+            state.biases.length !== NUM_CLASSES) {
+            throw new Error(`Invalid weight state: expected ${NUM_CLASSES}×${FEATURE_DIM} matrix + ${NUM_CLASSES} biases`);
+        }
+        this.weights = state.weights.map((row) => [...row]);
+        this.biases = [...state.biases];
+    }
+}
+// ---------------------------------------------------------------------------
+// Utility
+// ---------------------------------------------------------------------------
+function median(values) {
+    if (values.length === 0)
+        return 0;
+    const sorted = [...values].sort((a, b) => a - b);
+    const mid = Math.floor(sorted.length / 2);
+    return sorted.length % 2 === 0
+        ? (sorted[mid - 1] + sorted[mid]) / 2
+        : sorted[mid];
+}
+//# sourceMappingURL=predictor.js.map

package/dist/src/benchmarks/gaia-hardness/predictor.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"predictor.js","sourceRoot":"","sources":["../../../../src/benchmarks/gaia-hardness/predictor.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AAGH,OAAO,EAAE,eAAe,EAAsB,MAAM,eAAe,CAAC;AA4BpE,8EAA8E;AAC9E,wBAAwB;AACxB,8EAA8E;AAE9E,MAAM,CAAC,MAAM,eAAe,GAA2C;IACrE,IAAI,EAAE;QACJ,KAAK,EAAE,OAAO;QACd,QAAQ,EAAE,CAAC;QACX,cAAc,EAAE,CAAC;KAClB;IACD,MAAM,EAAE;QACN,KAAK,EAAE,QAAQ;QACf,QAAQ,EAAE,CAAC;QACX,cAAc,EAAE,CAAC;KAClB;IACD,IAAI,EAAE;QACJ,KAAK,EAAE,QAAQ;QACf,QAAQ,EAAE,EAAE;QACZ,cAAc,EAAE,CAAC;KAClB;CACF,CAAC;AAEF,8EAA8E;AAC9E,qBAAqB;AACrB,8EAA8E;AAE9E,MAAM,WAAW,GAAG,EAAE,CAAC;AACvB,MAAM,WAAW,GAAG,CAAC,CAAC,CAAC,2BAA2B;AAClD,MAAM,WAAW,GAAsB,CAAC,MAAM,EAAE,QAAQ,EAAE,MAAM,CAAC,CAAC;AAElE,mEAAmE;AACnE,MAAM,aAAa,GAAG,IAAI,CAAC;AAC3B,MAAM,qBAAqB,GAAG,IAAI,CAAC;AACnC,MAAM,eAAe,GAAG,GAAG,CAAC;AAE5B,wEAAwE;AACxE,2EAA2E;AAC3E,8DAA8D;AAC9D,MAAM,2BAA2B,GAAG,IAAI,CAAC;AAEzC,8EAA8E;AAC9E,yCAAyC;AACzC,8EAA8E;AAE9E,SAAS,OAAO,CAAC,MAAgB;IAC/B,MAAM,QAAQ,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,MAAM,CAAC,CAAC;IACrC,MAAM,IAAI,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,QAAQ,CAAC,CAAC,CAAC;IACvD,MAAM,OAAO,GAAG,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC;IAChD,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,OAAO,CAAC,CAAC;AACtC,CAAC;AAED,SAAS,GAAG,CAAC,CAAW,EAAE,CAAW;IACnC,IAAI,GAAG,GAAG,CAAC,CAAC;IACZ,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE;QAAE,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;IACtD,OAAO,GAAG,CAAC;AACb,CAAC;AAED,8EAA8E;AAC9E,oBAAoB;AACpB,8EAA8E;AAE9E,MAAM,OAAO,iBAAiB;IAC5B;;;OAGG;IACK,OAAO,GAAsB,IAAI,CAAC;IAE1C,4BAA4B;IACpB,MAAM,GAAoB,IAAI,CAAC;IAEvC,2DAA2D;IAC1C,gBAAgB,CAAU;IAE3C,YAAY,UAA0C,EAAE;QACtD,IAAI,CAAC,gBAAgB,GAAG,OAAO,CAAC,gBAAgB,IAAI,IAAI,CAAC;IAC3D,CAAC;IAED;;;OAGG;IACH,IAAI,SAAS;QACX,OAAO,IAAI,CAAC,OAAO,KAAK,IAAI,CAAC;IAC/B,CAAC;IAED,4EAA4E;IAE5E;;;;;;;;;;OAUG;IACH,KAAK,CAAC,WAA6B;QACjC,IAAI,WAAW,CAAC,MAAM,GAAG,EAAE,EAAE,CAAC;YAC5B,kDAAkD;YAClD,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC;YACpB,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC;YACnB,OAAO;QACT,CAAC;QAED,iDAAiD;QACjD,MAAM,QAAQ,GAAG,WAAW;aACzB,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,KAAK,SAAS,CAAC;aACpC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAe,CAAC,CAAC;QACjC,MAAM,WAAW,GAAG,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QAE/D,MAAM,CAAC,GAAe,EAAE,CAAC;QACzB,MAAM,CAAC,GAAa,EAAE,CAAC;QAEvB,KAAK,MAAM,OAAO,IAAI,WAAW,EAAE,CAAC;YAClC,MAAM,EAAE,GAAG,eAAe,CAAC,OAAO,CAAC,QAAQ,CAAC,CAAC;YAC7C,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,MAAM,CAAC,CAAC;YAElB,IAAI,QAAgB,CAAC;YACrB,IAAI,OAAO,CAAC,UAAU,EAAE,CAAC;gBACvB,MAAM,CAAC,GAAG,OAAO,CAAC,KAAK,IAAI,WAAW,CAAC;gBACvC,QAAQ,GAAG,CAAC,IAAI,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,mBAAmB;YAC1D,CAAC;iBAAM,CAAC;gBACN,QAAQ,GAAG,CAAC,CAAC,CAAC,OAAO;YACvB,CAAC;YACD,CAAC,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;QACnB,CAAC;QAED,2BAA2B;QAC3B,MAAM,CAAC,GAAe,KAAK,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,WAAW,EAAE,EAAE,GAAG,EAAE,CAC7D,IAAI,KAAK,CAAC,WAAW,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAC/B,CAAC;QACF,MAAM,CAAC,GAAa,IAAI,KAAK,CAAC,WAAW,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QAEnD,sEAAsE;QACtE,KAAK,IAAI,KAAK,GAAG,CAAC,EAAE,KAAK,GAAG,eAAe,EAAE,KAAK,EAAE,EAAE,CAAC;YACrD,wBAAwB;YACxB,MAAM,EAAE,GAAe,KAAK,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,WAAW,EAAE,EAAE,GAAG,EAAE,CAC9D,IAAI,KAAK,CAAC,WAAW,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAC/B,CAAC;YACF,MAAM,EAAE,GAAa,IAAI,KAAK,CAAC,WAAW,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;YAEpD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;gBAClC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;gBACf,MAAM,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;gBAEvB,4CAA4C;gBAC5C,MAAM,MAAM,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;gBACjD,MAAM,KAAK,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC;gBAE9B,yCAAyC;gBACzC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,WAAW,EAAE,CAAC,EAAE,EAAE,CAAC;oBACrC,MAAM,IAAI,GAAG,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,KAAK,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;oBAClD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,WAAW,EAAE,CAAC,EAAE,EAAE,CAAC;wBACrC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,IAAI,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;oBAC1B,CAAC;oBACD,EAAE,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC;gBAChB,CAAC;YACH,CAAC;YAED,yCAAyC;YACzC,MAAM,CAAC,GAAG,CAAC,CAAC,MAAM,CAAC;YACnB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,WAAW,EAAE,CAAC,EAAE,EAAE,CAAC;gBACrC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,WAAW,EAAE,CAAC,EAAE,EAAE,CAAC;oBACrC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,aAAa,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,GAAG,qBAAqB,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;gBAC9E,CAAC;gBACD,CAAC,CAAC,CAAC,CAAC,IAAI,aAAa,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;YACtC,CAAC;QACH,CAAC;QAED,IAAI,CAAC,OAAO,GAAG,CAAC,CAAC;QACjB,IAAI,CAAC,MAAM,GAAG,CAAC,CAAC;IAClB,CAAC;IAED,6EAA6E;IAE7E;;;;OAIG;IACH,OAAO,CAAC,QAAsB;QAC5B,MAAM,QAAQ,GAAG,eAAe,CAAC,QAAQ,CAAC,CAAC;QAE3C,IAAI,CAAC,IAAI,CAAC,OAAO,IAAI,CAAC,IAAI,CAAC,MAAM,EAAE,CAAC;YAClC,8CAA8C;YAC9C,OAAO;gBACL,UAAU,EAAE,QAAQ;gBACpB,UAAU,EAAE,GAAG;gBACf,MAAM,EAAE,eAAe,CAAC,MAAM;gBAC9B,QAAQ;aACT,CAAC;QACJ,CAAC;QAED,MAAM,MAAM,GAAG,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,CAAC,EAAE,QAAQ,CAAC,MAAM,CAAC,GAAG,IAAI,CAAC,MAAO,CAAC,CAAC,CAAC,CAAC,CAAC;QACrF,MAAM,KAAK,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC;QAE9B,mDAAmD;QACnD,IAAI,aAAa,GAAG,CAAC,GAAG,KAAK,CAAC,CAAC;QAC/B,IAAI,IAAI,CAAC,gBAAgB,EAAE,CAAC;YAC1B,qDAAqD;YACrD,MAAM,UAAU,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,CAAC,CAAC,GAAG,2BAA2B,CAAC,CAAC;YACvE,aAAa,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,CAAC,GAAG,UAAU,CAAC;YACzC,aAAa,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,CAAC,GAAG,UAAU,CAAC;QAC3C,CAAC;QAED,iCAAiC;QACjC,MAAM,MAAM,GAAG,aAAa,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC;QACxD,aAAa,GAAG,aAAa,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,MAAM,CAAC,CAAC;QAErD,eAAe;QACf,IAAI,SAAS,GAAG,CAAC,CAAC;QAClB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,WAAW,EAAE,CAAC,EAAE,EAAE,CAAC;YACrC,IAAI,aAAa,CAAC,CAAC,CAAC,GAAG,aAAa,CAAC,SAAS,CAAC;gBAAE,SAAS,GAAG,CAAC,CAAC;QACjE,CAAC;QAED,MAAM,UAAU,GAAG,WAAW,CAAC,SAAS,CAAC,CAAC;QAC1C,MAAM,UAAU,GAAG,aAAa,CAAC,SAAS,CAAC,CAAC;QAE5C,OAAO;YACL,UAAU;YACV,UAAU;YACV,MAAM,EAAE,eAAe,CAAC,UAAU,CAAC;YACnC,QAAQ;SACT,CAAC;IACJ,CAAC;IAED,6EAA6E;IAE7E;;;OAGG;IACH,MAAM;QACJ,IAAI,CAAC,IAAI,CAAC,OAAO,IAAI,CAAC,IAAI,CAAC,MAAM;YAAE,OAAO,IAAI,CAAC;QAC/C,OAAO,EAAE,OAAO,EAAE,IAAI,CAAC,OAAO,EAAE,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,CAAC;IACxD,CAAC;IAED;;OAEG;IACH,MAAM,CAAC,KAAgD;QACrD,IACE,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,OAAO,CAAC;YAC7B,KAAK,CAAC,OAAO,CAAC,MAAM,KAAK,WAAW;YACpC,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,MAAM,CAAC;YAC5B,KAAK,CAAC,MAAM,CAAC,MAAM,KAAK,WAAW,EACnC,CAAC;YACD,MAAM,IAAI,KAAK,CACb,kCAAkC,WAAW,IAAI,WAAW,aAAa,WAAW,SAAS,CAC9F,CAAC;QACJ,CAAC;QACD,IAAI,CAAC,OAAO,GAAG,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,CAAC;QACpD,IAAI,CAAC,MAAM,GAAG,CAAC,GAAG,KAAK,CAAC,MAAM,CAAC,CAAC;IAClC,CAAC;CACF;AAED,8EAA8E;AAC9E,UAAU;AACV,8EAA8E;AAE9E,SAAS,MAAM,CAAC,MAAgB;IAC9B,IAAI,MAAM,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,CAAC,CAAC;IAClC,MAAM,MAAM,GAAG,CAAC,GAAG,MAAM,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACjD,MAAM,GAAG,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;IAC1C,OAAO,MAAM,CAAC,MAAM,GAAG,CAAC,KAAK,CAAC;QAC5B,CAAC,CAAC,CAAC,MAAM,CAAC,GAAG,GAAG,CAAC,CAAC,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC;QACrC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;AAClB,CAAC"}

package/dist/src/benchmarks/gaia-hardness/predictor.smoke.d.ts

@@ -0,0 +1,20 @@
+/**
+ * Smoke tests for gaia-hardness predictor — ADR-136 Track Q
+ *
+ * All tests are mock-based (no live API calls, no HF token, $0 cost).
+ *
+ * Test matrix:
+ *   1. Feature extraction: math question fires requiresMath + qword_calc
+ *   2. Feature extraction: simple factual question has low complexity signal
+ *   3. Feature extraction: question with file attachment fires has_file_attachment
+ *   4. Feature extraction: multi-hop question fires multi_hop_signal
+ *   5. Cold-start fallback: untrained predictor returns medium with confidence=0.5
+ *   6. Train on mock data: predictor becomes trained after ≥ 10 examples
+ *   7. Trained predictor: math-heavy question predicts harder than trivial question
+ *   8. Export/import: weights round-trip through JSON without loss
+ *
+ * Refs: ADR-136, #2156
+ */
+declare function runSmoke(): Promise<void>;
+export { runSmoke };
+//# sourceMappingURL=predictor.smoke.d.ts.map

package/dist/src/benchmarks/gaia-hardness/predictor.smoke.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"predictor.smoke.d.ts","sourceRoot":"","sources":["../../../../src/benchmarks/gaia-hardness/predictor.smoke.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AA2PH,iBAAe,QAAQ,IAAI,OAAO,CAAC,IAAI,CAAC,CA+BvC;AAaD,OAAO,EAAE,QAAQ,EAAE,CAAC"}