claude-turing 3.2.0 → 3.3.0

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "turing",
-  "version": "3.2.0",
-  "description": "Autonomous ML research harness — the autoresearch loop as a formal protocol. 47 commands, 2 specialized agents, model debugging (xray + sensitivity + calibration), pre-training intelligence (sanity checks + baseline generation + leakage detection), meta-intelligence (cross-project knowledge transfer + methodology audit), scaling & efficiency (scaling laws + compute budget + model distillation), model composition (ensemble + pipeline stitch + warm-start), deep analysis (experiment diff + live training monitor + regression gate), experiment orchestration (batch queue + smart retry + branching), literature integration + paper drafting, production model export, performance profiling, smart checkpoints, experiment intelligence, statistical rigor, tree-search hypothesis exploration, cost-performance frontier, model cards, model registry, hypothesis database with novelty guard, anti-cheating guardrails, and the taste-leverage loop. Inspired by Karpathy's autoresearch and the scientific method itself.",
+  "version": "3.3.0",
+  "description": "Autonomous ML research harness — the autoresearch loop as a formal protocol. 49 commands, 2 specialized agents, feature & training intelligence (feature selection + curriculum optimization), model debugging (xray + sensitivity + calibration), pre-training intelligence (sanity checks + baseline generation + leakage detection), meta-intelligence (cross-project knowledge transfer + methodology audit), scaling & efficiency (scaling laws + compute budget + model distillation), model composition (ensemble + pipeline stitch + warm-start), deep analysis (experiment diff + live training monitor + regression gate), experiment orchestration (batch queue + smart retry + branching), literature integration + paper drafting, production model export, performance profiling, smart checkpoints, experiment intelligence, statistical rigor, tree-search hypothesis exploration, cost-performance frontier, model cards, model registry, hypothesis database with novelty guard, anti-cheating guardrails, and the taste-leverage loop. Inspired by Karpathy's autoresearch and the scientific method itself.",
   "author": {
     "name": "pragnition"
   },

package/README.md

@@ -358,6 +358,8 @@ The index (`hypotheses.yaml`) is the lightweight queue. The detail files (`hypot
 | `/turing:xray [exp-id]` | Internal model diagnostics — gradient flow, dead neurons, weight distributions, tree analysis |
 | `/turing:sensitivity [exp-id]` | Hyperparameter sensitivity — rank parameters by impact, detect non-monotonic responses |
 | `/turing:calibrate [exp-id]` | Probability calibration — ECE/MCE, reliability diagrams, Platt/isotonic/temperature scaling |
+| `/turing:feature [--method]` | Automated feature selection — multi-method consensus ranking, redundancy, interactions |
+| `/turing:curriculum [exp-id]` | Training curriculum optimization — difficulty scoring, strategy comparison, mislabeled sample detection |
 
 And for fully hands-off operation:
 
@@ -542,11 +544,11 @@ Each project gets independent config, data, experiments, models, and agent memor
 
 ## Architecture of Turing Itself
 
-47 commands, 2 agents, 10 config files, 66 template scripts, model registry, artifact contract, cost-performance frontier, model cards, tree-search exploration, statistical rigor, experiment intelligence, performance profiling, smart checkpoints, production model export, literature integration, paper section drafting, experiment orchestration (queue + retry + fork), deep analysis (diff + watch + regress), model composition (ensemble + stitch + warm), scaling & efficiency (scale + budget + distill), meta-intelligence (transfer + audit), pre-training intelligence (sanity + baseline + leak), model debugging (xray + sensitivity + calibrate), 16 ADRs. See [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) for the full codemap.
+49 commands, 2 agents, 10 config files, 68 template scripts, model registry, artifact contract, cost-performance frontier, model cards, tree-search exploration, statistical rigor, experiment intelligence, performance profiling, smart checkpoints, production model export, literature integration, paper section drafting, experiment orchestration (queue + retry + fork), deep analysis (diff + watch + regress), model composition (ensemble + stitch + warm), scaling & efficiency (scale + budget + distill), meta-intelligence (transfer + audit), pre-training intelligence (sanity + baseline + leak), model debugging (xray + sensitivity + calibrate), feature & training intelligence (feature + curriculum), 16 ADRs. See [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) for the full codemap.
 
 ```
 turing/
-├── commands/              46 skill files (core + taste-leverage + reporting + exploration + statistical rigor + experiment intelligence + performance + deployment + research workflow + orchestration + deep analysis + model composition + scaling & efficiency + meta-intelligence + pre-training intelligence + model debugging)
+├── commands/              48 skill files (core + taste-leverage + reporting + exploration + statistical rigor + experiment intelligence + performance + deployment + research workflow + orchestration + deep analysis + model composition + scaling & efficiency + meta-intelligence + pre-training intelligence + model debugging + feature & training intelligence)
 ├── agents/                2 agents (researcher: read/write, evaluator: read-only)
 ├── config/                8 files (lifecycle, taxonomy, archetypes, novelty aliases)
 ├── templates/             Scaffolded into user projects by /turing:init

package/commands/curriculum.md

@@ -0,0 +1,43 @@
+---
+name: curriculum
+description: Training curriculum optimization — order data by difficulty, compare easy-to-hard vs hard-to-easy vs self-paced strategies.
+disable-model-invocation: true
+argument-hint: "[exp-id] [--strategies easy-to-hard,random]"
+allowed-tools: Read, Bash(*), Grep, Glob
+---
+
+Does the order your model sees data matter? Find out systematically.
+
+## Steps
+
+1. **Activate environment:**
+   ```bash
+   source .venv/bin/activate
+   ```
+
+2. **Parse arguments from `$ARGUMENTS`:**
+   - Optional experiment ID
+   - `--strategies "easy_to_hard,hard_to_easy,self_paced,random"` — strategies to test
+   - `--json` — raw JSON output
+
+3. **Run curriculum analysis:**
+   ```bash
+   python scripts/curriculum_optimizer.py $ARGUMENTS
+   ```
+
+4. **Strategies tested:**
+   - **Random:** standard shuffling (control)
+   - **Easy-to-hard:** classic curriculum learning
+   - **Hard-to-easy:** anti-curriculum
+   - **Self-paced:** start easy, gradually include harder samples
+
+5. **Report includes:** strategy comparison table with metric, convergence epoch, and speedup vs random; impossible sample detection (likely mislabeled)
+
+6. **Saved output:** report in `experiments/curriculum/<exp-id>-curriculum.yaml`
+
+## Examples
+
+```
+/turing:curriculum exp-042                      # All strategies
+/turing:curriculum --strategies easy_to_hard,random  # Specific strategies
+```

package/commands/feature.md

@@ -0,0 +1,42 @@
+---
+name: feature
+description: Automated feature selection — multi-method importance consensus, redundancy detection, and interaction feature generation.
+disable-model-invocation: true
+argument-hint: "[--method all|importance] [--top-k 20]"
+allowed-tools: Read, Bash(*), Grep, Glob
+---
+
+Systematically evaluate which features matter and which are noise.
+
+## Steps
+
+1. **Activate environment:**
+   ```bash
+   source .venv/bin/activate
+   ```
+
+2. **Parse arguments from `$ARGUMENTS`:**
+   - `--method all|importance|selection|generation` — analysis type (default: all)
+   - `--top-k 20` — number of top features to consider
+   - `--json` — raw JSON output
+
+3. **Run feature analysis:**
+   ```bash
+   python scripts/feature_intelligence.py $ARGUMENTS
+   ```
+
+4. **Report includes:**
+   - Consensus ranking: features ranked by number of methods placing them in top-K
+   - Per-method ranks: mutual information, L1, tree-based
+   - Redundant pairs: features with |r| > 0.95
+   - Candidate interaction features from top consensus set
+   - Drop recommendation for zero-consensus features
+
+5. **Saved output:** report in `experiments/features/features-*.yaml`
+
+## Examples
+
+```
+/turing:feature                      # Full analysis
+/turing:feature --top-k 10           # Top-10 consensus
+```

package/commands/turing.md

@@ -56,6 +56,8 @@ You are the Turing ML research router. Detect the user's intent and route to the
 | "xray", "model internals", "dead neurons", "gradient flow", "weight distribution", "inside the model" | `/turing:xray` | Analyze |
 | "sensitivity", "which params matter", "hyperparameter importance", "parameter ranking" | `/turing:sensitivity` | Analyze |
 | "calibrate", "calibration", "ECE", "reliability diagram", "overconfident", "probability calibration" | `/turing:calibrate` | Analyze |
+| "feature", "features", "feature selection", "feature importance", "which features matter", "redundant features" | `/turing:feature` | Analyze |
+| "curriculum", "training order", "easy to hard", "data ordering", "curriculum learning" | `/turing:curriculum` | Optimize |
 
 ## Sub-commands
 
@@ -108,6 +110,8 @@ You are the Turing ML research router. Detect the user's intent and route to the
 | `/turing:xray [exp-id]` | Internal model diagnostics: gradient flow, dead neurons, weight distributions, tree analysis | (inline) |
 | `/turing:sensitivity [exp-id]` | Hyperparameter sensitivity analysis: rank parameters by impact, detect non-monotonic responses | (inline) |
 | `/turing:calibrate [exp-id]` | Probability calibration: ECE/MCE, reliability diagrams, Platt/isotonic/temperature scaling | (inline) |
+| `/turing:feature [--method]` | Automated feature selection: multi-method consensus ranking, redundancy, interaction generation | (inline) |
+| `/turing:curriculum [exp-id]` | Training curriculum optimization: difficulty scoring, strategy comparison, impossible sample detection | (inline) |
 
 ## Proactive Detection
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-turing",
-  "version": "3.2.0",
+  "version": "3.3.0",
   "type": "module",
   "description": "Autonomous ML research harness for Claude Code. The autoresearch loop as a formal protocol — iteratively trains, evaluates, and improves ML models with structured experiment tracking, convergence detection, immutable evaluation infrastructure, and safety guardrails.",
   "bin": {

package/src/install.js

@@ -31,6 +31,7 @@ const SUB_COMMANDS = [
   "transfer", "audit",
   "sanity", "baseline", "leak",
   "xray", "sensitivity", "calibrate",
+  "feature", "curriculum",
 ];
 
 export async function install(opts = {}) {

package/src/verify.js

@@ -61,6 +61,8 @@ const EXPECTED_COMMANDS = [
   "xray/SKILL.md",
   "sensitivity/SKILL.md",
   "calibrate/SKILL.md",
+  "feature/SKILL.md",
+  "curriculum/SKILL.md",
 ];
 
 const EXPECTED_AGENTS = ["ml-researcher.md", "ml-evaluator.md"];

package/templates/scripts/curriculum_optimizer.py

@@ -0,0 +1,337 @@
+#!/usr/bin/env python3
+"""Training curriculum optimization for the autoresearch pipeline.
+
+Orders training data by difficulty and measures whether curriculum
+learning improves convergence speed or final performance. Tests
+easy-to-hard, hard-to-easy, self-paced, and random strategies.
+
+Usage:
+    python scripts/curriculum_optimizer.py exp-042
+    python scripts/curriculum_optimizer.py --strategies easy-to-hard,random
+    python scripts/curriculum_optimizer.py --json
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from datetime import datetime, timezone
+from pathlib import Path
+
+import numpy as np
+import yaml
+
+from scripts.turing_io import load_config
+
+DEFAULT_STRATEGIES = ["random", "easy_to_hard", "hard_to_easy", "self_paced"]
+IMPOSSIBLE_THRESHOLD = 0.9  # Samples with difficulty > this across all strategies
+
+
+# --- Difficulty Scoring ---
+
+
+def score_difficulty_by_loss(
+    losses: np.ndarray,
+) -> np.ndarray:
+    """Score sample difficulty by loss value (higher loss = harder).
+
+    Normalizes to [0, 1].
+    """
+    if len(losses) == 0:
+        return np.array([])
+
+    min_loss = np.min(losses)
+    max_loss = np.max(losses)
+    if max_loss == min_loss:
+        return np.zeros(len(losses))
+
+    return (losses - min_loss) / (max_loss - min_loss)
+
+
+def score_difficulty_by_margin(
+    margins: np.ndarray,
+) -> np.ndarray:
+    """Score sample difficulty by margin (smaller margin = harder).
+
+    Margins = distance from decision boundary. Normalizes to [0, 1].
+    """
+    if len(margins) == 0:
+        return np.array([])
+
+    min_m = np.min(margins)
+    max_m = np.max(margins)
+    if max_m == min_m:
+        return np.full(len(margins), 0.5)
+
+    # Invert: small margin = high difficulty
+    return 1.0 - (margins - min_m) / (max_m - min_m)
+
+
+def score_difficulty_by_disagreement(
+    multi_seed_predictions: list[np.ndarray],
+    labels: np.ndarray,
+) -> np.ndarray:
+    """Score difficulty by prediction disagreement across seeds.
+
+    Samples where different seeds disagree are "hard" (and possibly mislabeled).
+    """
+    if not multi_seed_predictions or len(labels) == 0:
+        return np.array([])
+
+    n_samples = len(labels)
+    n_seeds = len(multi_seed_predictions)
+
+    agreement = np.zeros(n_samples)
+    for preds in multi_seed_predictions:
+        if len(preds) == n_samples:
+            agreement += (preds == labels).astype(float)
+
+    agreement /= n_seeds  # Fraction of seeds that got it right
+
+    # Disagreement = difficulty
+    return 1.0 - agreement
+
+
+# --- Curriculum Strategies ---
+
+
+def apply_curriculum(
+    indices: np.ndarray,
+    difficulties: np.ndarray,
+    strategy: str,
+) -> np.ndarray:
+    """Reorder sample indices according to curriculum strategy.
+
+    Args:
+        indices: Original sample indices.
+        difficulties: Difficulty scores [0, 1] per sample.
+        strategy: Curriculum strategy name.
+
+    Returns:
+        Reordered indices.
+    """
+    if len(indices) == 0:
+        return indices
+
+    if strategy == "random":
+        np.random.shuffle(indices)
+        return indices
+
+    elif strategy == "easy_to_hard":
+        order = np.argsort(difficulties)
+        return indices[order]
+
+    elif strategy == "hard_to_easy":
+        order = np.argsort(difficulties)[::-1]
+        return indices[order]
+
+    elif strategy == "self_paced":
+        # Start with easiest 20%, then gradually include harder
+        order = np.argsort(difficulties)
+        n = len(order)
+        # Shuffle within difficulty bands
+        bands = [order[:n // 5], order[n // 5:2 * n // 5],
+                 order[2 * n // 5:3 * n // 5], order[3 * n // 5:4 * n // 5],
+                 order[4 * n // 5:]]
+        result = []
+        for band in bands:
+            np.random.shuffle(band)
+            result.extend(band)
+        return np.array(result)
+
+    return indices
+
+
+def detect_impossible_samples(
+    difficulties: np.ndarray,
+    threshold: float = IMPOSSIBLE_THRESHOLD,
+) -> list[int]:
+    """Find samples that are consistently difficult (likely mislabeled).
+
+    Returns list of sample indices.
+    """
+    return [int(i) for i in range(len(difficulties)) if difficulties[i] > threshold]
+
+
+# --- Strategy Comparison ---
+
+
+def compare_strategies(
+    strategy_results: dict[str, dict],
+    primary_metric: str = "accuracy",
+) -> dict:
+    """Compare curriculum strategy results.
+
+    Args:
+        strategy_results: {strategy_name: {metric_value, convergence_epoch, ...}}
+
+    Returns:
+        Comparison report with best strategy and verdict.
+    """
+    if not strategy_results:
+        return {"best_strategy": None, "verdict": "no_data"}
+
+    # Find baseline (random)
+    baseline = strategy_results.get("random", {})
+    baseline_metric = baseline.get("metric_value", 0)
+    baseline_epochs = baseline.get("convergence_epoch")
+
+    results = []
+    for name, data in strategy_results.items():
+        metric = data.get("metric_value", 0)
+        epochs = data.get("convergence_epoch")
+        speedup = None
+        if epochs and baseline_epochs and baseline_epochs > 0:
+            speedup = round(1 - epochs / baseline_epochs, 4)
+
+        results.append({
+            "strategy": name,
+            "metric_value": round(metric, 6) if metric else None,
+            "convergence_epoch": epochs,
+            "delta_vs_random": round(metric - baseline_metric, 6) if metric and baseline_metric else None,
+            "speedup": speedup,
+        })
+
+    # Find best by metric
+    with_metric = [r for r in results if r["metric_value"] is not None]
+    best = max(with_metric, key=lambda r: r["metric_value"]) if with_metric else None
+
+    verdict = "no_improvement"
+    if best and best.get("delta_vs_random") and best["delta_vs_random"] > 0.005:
+        verdict = "curriculum_helps"
+    elif best and best.get("speedup") and best["speedup"] > 0.1:
+        verdict = "faster_convergence"
+
+    return {
+        "results": results,
+        "best_strategy": best.get("strategy") if best else None,
+        "verdict": verdict,
+    }
+
+
+# --- Full Pipeline ---
+
+
+def curriculum_analysis(
+    difficulties: np.ndarray | None = None,
+    strategy_results: dict[str, dict] | None = None,
+    exp_id: str | None = None,
+    config_path: str = "config.yaml",
+) -> dict:
+    """Run curriculum analysis."""
+    config = load_config(config_path)
+    primary_metric = config.get("evaluation", {}).get("primary_metric", "accuracy")
+
+    report = {
+        "generated_at": datetime.now(timezone.utc).isoformat(),
+        "experiment_id": exp_id,
+        "primary_metric": primary_metric,
+    }
+
+    if difficulties is not None:
+        impossible = detect_impossible_samples(difficulties)
+        report["difficulty_stats"] = {
+            "n_samples": len(difficulties),
+            "mean_difficulty": round(float(np.mean(difficulties)), 4),
+            "n_impossible": len(impossible),
+            "impossible_indices": impossible[:20],
+        }
+
+    if strategy_results:
+        comparison = compare_strategies(strategy_results, primary_metric)
+        report["comparison"] = comparison
+    else:
+        report["note"] = "Provide strategy_results for comparison. Use /turing:curriculum to run strategies."
+        report["available_strategies"] = DEFAULT_STRATEGIES
+
+    return report
+
+
+# --- Report Formatting ---
+
+
+def save_curriculum_report(report: dict, output_dir: str = "experiments/curriculum") -> Path:
+    out_path = Path(output_dir)
+    out_path.mkdir(parents=True, exist_ok=True)
+    exp_id = report.get("experiment_id", "unknown")
+    filepath = out_path / f"{exp_id}-curriculum.yaml"
+    clean = json.loads(json.dumps(report, default=str))
+    with open(filepath, "w") as f:
+        yaml.dump(clean, f, default_flow_style=False, sort_keys=False)
+    return filepath
+
+
+def format_curriculum_report(report: dict) -> str:
+    if "error" in report:
+        return f"ERROR: {report['error']}"
+
+    exp_id = report.get("experiment_id", "?")
+    metric = report.get("primary_metric", "metric")
+
+    lines = [f"# Curriculum Analysis: {exp_id}", "",
+             f"*Generated {report.get('generated_at', 'N/A')[:19]}*", ""]
+
+    # Difficulty stats
+    diff_stats = report.get("difficulty_stats")
+    if diff_stats:
+        lines.extend([
+            "## Difficulty Distribution",
+            f"- **Samples:** {diff_stats['n_samples']}",
+            f"- **Mean difficulty:** {diff_stats['mean_difficulty']:.4f}",
+            f"- **Impossible samples:** {diff_stats['n_impossible']} (likely mislabeled)",
+            "",
+        ])
+
+    # Strategy comparison
+    comparison = report.get("comparison")
+    if comparison:
+        results = comparison.get("results", [])
+        if results:
+            lines.extend(["## Strategy Comparison", "",
+                         f"| Strategy | {metric} | Δ vs Random | Speedup |",
+                         "|----------|--------|-------------|---------|"])
+            best_name = comparison.get("best_strategy")
+            for r in results:
+                val = f"{r['metric_value']:.4f}" if r.get("metric_value") is not None else "N/A"
+                delta = f"{r['delta_vs_random']:+.4f}" if r.get("delta_vs_random") is not None else "—"
+                speedup = f"{r['speedup']:+.0%}" if r.get("speedup") is not None else "—"
+                marker = " ← BEST" if r["strategy"] == best_name else ""
+                lines.append(f"| {r['strategy']} | {val} | {delta} | {speedup} |{marker}")
+            lines.append("")
+
+        verdict_labels = {
+            "curriculum_helps": "Curriculum learning improves final performance",
+            "faster_convergence": "Curriculum learning converges faster (similar final performance)",
+            "no_improvement": "No significant improvement from curriculum ordering",
+        }
+        verdict = comparison.get("verdict", "?")
+        lines.extend(["## Verdict", "", f"**{verdict_labels.get(verdict, verdict.upper())}**"])
+    elif report.get("note"):
+        lines.append(f"*{report['note']}*")
+
+    return "\n".join(lines)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Training curriculum optimization")
+    parser.add_argument("exp_id", nargs="?", help="Experiment ID")
+    parser.add_argument("--strategies", help="Comma-separated strategies")
+    parser.add_argument("--config", default="config.yaml")
+    parser.add_argument("--json", action="store_true")
+    args = parser.parse_args()
+
+    report = curriculum_analysis(exp_id=args.exp_id, config_path=args.config)
+
+    if "error" not in report:
+        filepath = save_curriculum_report(report)
+        print(f"Saved to {filepath}", file=sys.stderr)
+
+    if args.json:
+        print(json.dumps(report, indent=2, default=str))
+    else:
+        print(format_curriculum_report(report))
+
+
+if __name__ == "__main__":
+    main()

package/templates/scripts/feature_intelligence.py

@@ -0,0 +1,369 @@
+#!/usr/bin/env python3
+"""Automated feature selection and generation for the autoresearch pipeline.
+
+Runs multiple feature importance methods (mutual information, permutation,
+L1, tree-based), computes consensus ranking, detects redundancy, and
+generates candidate interaction features.
+
+Usage:
+    python scripts/feature_intelligence.py
+    python scripts/feature_intelligence.py --method all
+    python scripts/feature_intelligence.py --method importance --top-k 15
+    python scripts/feature_intelligence.py --json
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from datetime import datetime, timezone
+from pathlib import Path
+
+import numpy as np
+import yaml
+
+from scripts.turing_io import load_config
+
+DEFAULT_TOP_K = 20
+REDUNDANCY_THRESHOLD = 0.95
+
+
+# --- Importance Methods ---
+
+
+def mutual_information_ranking(
+    X: np.ndarray,
+    y: np.ndarray,
+    feature_names: list[str] | None = None,
+) -> list[dict]:
+    """Rank features by mutual information with the target."""
+    from sklearn.feature_selection import mutual_info_classif, mutual_info_regression
+
+    if feature_names is None:
+        feature_names = [f"feature_{i}" for i in range(X.shape[1])]
+
+    # Detect task type from target
+    unique_vals = len(np.unique(y))
+    if unique_vals <= 20:  # Classification heuristic
+        scores = mutual_info_classif(X, y, random_state=42)
+    else:
+        scores = mutual_info_regression(X, y, random_state=42)
+
+    ranked = sorted(
+        [{"feature": feature_names[i], "score": round(float(scores[i]), 6), "rank": 0}
+         for i in range(len(scores))],
+        key=lambda x: x["score"], reverse=True,
+    )
+    for i, r in enumerate(ranked):
+        r["rank"] = i + 1
+
+    return ranked
+
+
+def l1_ranking(
+    X: np.ndarray,
+    y: np.ndarray,
+    feature_names: list[str] | None = None,
+) -> list[dict]:
+    """Rank features by L1 regularization coefficient magnitude."""
+    from sklearn.linear_model import Lasso, LogisticRegression
+    from sklearn.preprocessing import StandardScaler
+
+    if feature_names is None:
+        feature_names = [f"feature_{i}" for i in range(X.shape[1])]
+
+    scaler = StandardScaler()
+    X_scaled = scaler.fit_transform(X)
+
+    unique_vals = len(np.unique(y))
+    if unique_vals <= 20:
+        model = LogisticRegression(penalty="l1", solver="liblinear", C=1.0, max_iter=1000)
+        model.fit(X_scaled, y)
+        coefs = np.abs(model.coef_).mean(axis=0) if model.coef_.ndim > 1 else np.abs(model.coef_.ravel())
+    else:
+        model = Lasso(alpha=0.01, max_iter=1000)
+        model.fit(X_scaled, y)
+        coefs = np.abs(model.coef_)
+
+    ranked = sorted(
+        [{"feature": feature_names[i], "score": round(float(coefs[i]), 6), "rank": 0}
+         for i in range(len(coefs))],
+        key=lambda x: x["score"], reverse=True,
+    )
+    for i, r in enumerate(ranked):
+        r["rank"] = i + 1
+
+    return ranked
+
+
+def tree_importance_ranking(
+    X: np.ndarray,
+    y: np.ndarray,
+    feature_names: list[str] | None = None,
+) -> list[dict]:
+    """Rank features by tree-based importance."""
+    from sklearn.ensemble import RandomForestClassifier, RandomForestRegressor
+
+    if feature_names is None:
+        feature_names = [f"feature_{i}" for i in range(X.shape[1])]
+
+    unique_vals = len(np.unique(y))
+    if unique_vals <= 20:
+        model = RandomForestClassifier(n_estimators=100, random_state=42, n_jobs=-1)
+    else:
+        model = RandomForestRegressor(n_estimators=100, random_state=42, n_jobs=-1)
+
+    model.fit(X, y)
+    importances = model.feature_importances_
+
+    ranked = sorted(
+        [{"feature": feature_names[i], "score": round(float(importances[i]), 6), "rank": 0}
+         for i in range(len(importances))],
+        key=lambda x: x["score"], reverse=True,
+    )
+    for i, r in enumerate(ranked):
+        r["rank"] = i + 1
+
+    return ranked
+
+
+# --- Consensus ---
+
+
+def compute_consensus(
+    rankings: dict[str, list[dict]],
+    top_k: int = DEFAULT_TOP_K,
+) -> list[dict]:
+    """Compute consensus ranking across multiple methods.
+
+    A feature's consensus score = number of methods that place it in top-K.
+    """
+    n_methods = len(rankings)
+    feature_scores = {}
+
+    for method_name, ranking in rankings.items():
+        top_features = {r["feature"] for r in ranking[:top_k]}
+        for feat in top_features:
+            if feat not in feature_scores:
+                feature_scores[feat] = {"feature": feat, "methods": {}, "consensus": 0}
+            feature_scores[feat]["methods"][method_name] = next(
+                (r["rank"] for r in ranking if r["feature"] == feat), None
+            )
+            feature_scores[feat]["consensus"] += 1
+
+    # Add features not in any top-K
+    all_features = set()
+    for ranking in rankings.values():
+        for r in ranking:
+            all_features.add(r["feature"])
+
+    for feat in all_features:
+        if feat not in feature_scores:
+            feature_scores[feat] = {
+                "feature": feat,
+                "methods": {m: next((r["rank"] for r in rk if r["feature"] == feat), None) for m, rk in rankings.items()},
+                "consensus": 0,
+            }
+
+    result = sorted(feature_scores.values(), key=lambda x: (-x["consensus"], x["feature"]))
+
+    for r in result:
+        r["consensus_str"] = f"{r['consensus']}/{n_methods}"
+        if r["consensus"] == n_methods:
+            r["consensus_str"] += " ★"
+        elif r["consensus"] == 0:
+            r["consensus_str"] += " — DROP"
+
+    return result
+
+
+# --- Redundancy Detection ---
+
+
+def detect_redundancy(
+    X: np.ndarray,
+    feature_names: list[str] | None = None,
+    threshold: float = REDUNDANCY_THRESHOLD,
+) -> list[dict]:
+    """Detect highly correlated feature pairs."""
+    if feature_names is None:
+        feature_names = [f"feature_{i}" for i in range(X.shape[1])]
+
+    n = X.shape[1]
+    if n < 2:
+        return []
+
+    corr = np.corrcoef(X.T)
+    redundant = []
+
+    for i in range(n):
+        for j in range(i + 1, n):
+            c = abs(corr[i, j])
+            if not np.isnan(c) and c > threshold:
+                redundant.append({
+                    "feature_a": feature_names[i],
+                    "feature_b": feature_names[j],
+                    "correlation": round(float(c), 4),
+                })
+
+    return sorted(redundant, key=lambda x: -x["correlation"])
+
+
+# --- Feature Generation ---
+
+
+def generate_interaction_features(
+    top_features: list[str],
+    max_interactions: int = 10,
+) -> list[dict]:
+    """Generate candidate interaction features from top consensus features."""
+    candidates = []
+
+    for i, fa in enumerate(top_features[:5]):
+        for fb in top_features[i + 1:6]:
+            if len(candidates) >= max_interactions:
+                break
+            candidates.append({"name": f"{fa}_x_{fb}", "type": "product", "features": [fa, fb]})
+            candidates.append({"name": f"{fa}_div_{fb}", "type": "ratio", "features": [fa, fb]})
+
+    return candidates[:max_interactions]
+
+
+# --- Full Pipeline ---
+
+
+def feature_analysis(
+    X: np.ndarray | None = None,
+    y: np.ndarray | None = None,
+    feature_names: list[str] | None = None,
+    method: str = "all",
+    top_k: int = DEFAULT_TOP_K,
+    config_path: str = "config.yaml",
+) -> dict:
+    """Run feature intelligence analysis."""
+    config = load_config(config_path)
+
+    if X is None or y is None:
+        return {"error": "Provide X and y arrays for feature analysis",
+                "note": "Run with --data train.npz to analyze features"}
+
+    if feature_names is None:
+        feature_names = [f"feature_{i}" for i in range(X.shape[1])]
+
+    rankings = {}
+
+    if method in ("all", "importance"):
+        rankings["mutual_info"] = mutual_information_ranking(X, y, feature_names)
+        rankings["l1"] = l1_ranking(X, y, feature_names)
+        rankings["tree"] = tree_importance_ranking(X, y, feature_names)
+
+    if not rankings:
+        return {"error": f"Unknown method: {method}"}
+
+    consensus = compute_consensus(rankings, top_k)
+    redundant = detect_redundancy(X, feature_names)
+
+    top_consensus_features = [c["feature"] for c in consensus if c["consensus"] > 0][:top_k]
+    interactions = generate_interaction_features(top_consensus_features)
+
+    drop_candidates = [c for c in consensus if c["consensus"] == 0]
+
+    report = {
+        "generated_at": datetime.now(timezone.utc).isoformat(),
+        "n_features": X.shape[1],
+        "top_k": top_k,
+        "rankings": {k: v[:top_k] for k, v in rankings.items()},
+        "consensus": consensus[:top_k * 2],
+        "drop_candidates": drop_candidates,
+        "n_drop": len(drop_candidates),
+        "redundant_pairs": redundant,
+        "interaction_candidates": interactions,
+        "recommendation": f"Drop {len(drop_candidates)} features with 0/{len(rankings)} consensus ({len(drop_candidates)/X.shape[1]*100:.0f}% of features)" if drop_candidates else "All features contribute to at least one method",
+    }
+
+    return report
+
+
+# --- Report Formatting ---
+
+
+def save_feature_report(report: dict, output_dir: str = "experiments/features") -> Path:
+    out_path = Path(output_dir)
+    out_path.mkdir(parents=True, exist_ok=True)
+    date = datetime.now(timezone.utc).strftime("%Y-%m-%d")
+    filepath = out_path / f"features-{date}.yaml"
+    clean = json.loads(json.dumps(report, default=str))
+    with open(filepath, "w") as f:
+        yaml.dump(clean, f, default_flow_style=False, sort_keys=False)
+    return filepath
+
+
+def format_feature_report(report: dict) -> str:
+    if "error" in report:
+        return f"ERROR: {report['error']}\n{report.get('note', '')}"
+
+    lines = ["# Feature Intelligence", "",
+             f"*Generated {report.get('generated_at', 'N/A')[:19]}*",
+             f"**{report.get('n_features', 0)} features analyzed, top-{report.get('top_k', 20)}**", ""]
+
+    # Consensus table
+    consensus = report.get("consensus", [])
+    if consensus:
+        methods = set()
+        for c in consensus:
+            methods.update(c.get("methods", {}).keys())
+        method_names = sorted(methods)
+
+        header = "| Feature |" + "|".join(f" {m} Rank " for m in method_names) + "| Consensus |"
+        sep = "|---------|" + "|".join("-------" for _ in method_names) + "|-----------|"
+        lines.extend(["## Consensus Ranking", "", header, sep])
+        for c in consensus[:15]:
+            ranks = "|".join(f" {c['methods'].get(m, '—')} " for m in method_names)
+            lines.append(f"| {c['feature']} |{ranks}| {c['consensus_str']} |")
+        lines.append("")
+
+    # Redundancy
+    redundant = report.get("redundant_pairs", [])
+    if redundant:
+        lines.extend(["## Redundant Pairs", ""])
+        for r in redundant[:5]:
+            lines.append(f"- **{r['feature_a']}** ↔ **{r['feature_b']}**: r={r['correlation']}")
+        lines.append("")
+
+    # Interactions
+    interactions = report.get("interaction_candidates", [])
+    if interactions:
+        lines.extend(["## Candidate Interactions", ""])
+        for i in interactions[:5]:
+            lines.append(f"- `{i['name']}` ({i['type']}: {' × '.join(i['features'])})")
+        lines.append("")
+
+    # Recommendation
+    lines.extend(["## Recommendation", "", report.get("recommendation", "")])
+
+    return "\n".join(lines)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Automated feature selection")
+    parser.add_argument("--method", choices=["all", "importance", "selection", "generation"], default="all")
+    parser.add_argument("--top-k", type=int, default=DEFAULT_TOP_K)
+    parser.add_argument("--config", default="config.yaml")
+    parser.add_argument("--json", action="store_true")
+    args = parser.parse_args()
+
+    report = feature_analysis(method=args.method, top_k=args.top_k, config_path=args.config)
+
+    if "error" not in report:
+        filepath = save_feature_report(report)
+        print(f"Saved to {filepath}", file=sys.stderr)
+
+    if args.json:
+        print(json.dumps(report, indent=2, default=str))
+    else:
+        print(format_feature_report(report))
+
+
+if __name__ == "__main__":
+    main()

package/templates/scripts/scaffold.py

@@ -124,6 +124,8 @@ TEMPLATE_DIRS = {
         "model_xray.py",
         "sensitivity_analysis.py",
         "calibration.py",
+        "feature_intelligence.py",
+        "curriculum_optimizer.py",
     ],
     "tests": ["__init__.py", "conftest.py"],
 }
@@ -160,6 +162,8 @@ DIRECTORIES_TO_CREATE = [
     "experiments/xrays",
     "experiments/sensitivity",
     "experiments/calibration",
+    "experiments/features",
+    "experiments/curriculum",
     "experiments/logs",
     "models/best",
     "models/archive",