cross_model.py

#!/usr/bin/env python3
"""
cross_model.py — Cross-Model Drift Analysis
=============================================
Runs AFTER analyze_single.py on 2+ models.
Uses probe bundles + caches to compare drift representations across architectures.

6 Experiments:
  [CM-1] Full-layer CKA matrix (L_A × L_B per pair, not just best layer)
  [CM-2] Drift score correlation (probe A scores vs probe B scores on shared queries)
  [CM-3] Differential facts (queries drifted for A but stable for B)
  [CM-4] Layer correspondence (best layer as % of depth — universal localization?)
  [CM-5] Neuron overlap (same-dim models only: which neuron indices carry drift?)
  [CM-6] Universality score (aggregate metric for paper abstract)

Outputs:
  cross_model_results.json     Complete results
  figures/fig_cm1_cka.png      Layer-wise CKA heatmaps
  figures/fig_cm2_corr.png     Score correlation matrix
  figures/fig_cm3_diff.png     Differential facts scatter
  figures/fig_cm4_layers.png   Layer correspondence bar
  figures/fig_cm5_neurons.png  Neuron overlap (same-dim pairs)
  figures/fig_cm6_summary.png  Universality summary

Usage:
  # Compare two models
  python cross_model.py --models qwen25 llama31

  # All available models
  python cross_model.py --all

  # Quick mode (skip full-layer CKA, just best-layer)
  python cross_model.py --all --quick
"""

import argparse
import json
import logging
import time
import warnings
from pathlib import Path

import numpy as np
import yaml

warnings.filterwarnings("ignore")
logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s [%(levelname)s] %(message)s",
    handlers=[logging.StreamHandler()])
logger = logging.getLogger(__name__)


# ─────────────────────────────────────────────────────────────────────────────
# CONFIG + DATA LOADING
# ─────────────────────────────────────────────────────────────────────────────

def load_config(path="models.yaml"):
    with open(path) as f:
        return yaml.safe_load(f)


def load_cache(model_dir, model_key):
    path = Path(model_dir) / model_key / f"cached_{model_key}.npz"
    if not path.exists():
        logger.error(f"Cache not found: {path}")
        return None
    results = np.load(str(path), allow_pickle=True)["results"].tolist()
    logger.info(f"  [{model_key}] Loaded {len(results)} samples")
    return results


def load_probe_bundle(model_dir, model_key):
    path = Path(model_dir) / model_key / f"probe_bundle_{model_key}.npz"
    if not path.exists():
        logger.warning(f"Probe bundle not found: {path}")
        return None
    d = np.load(str(path), allow_pickle=True)
    bundle = {k: d[k] for k in d.files}
    # Convert scalar items
    for k in ["best_layer", "hidden_dim", "n_samples"]:
        if k in bundle:
            bundle[k] = int(bundle[k])
    for k in ["drift_auroc", "cos_du", "cos_dc"]:
        if k in bundle:
            bundle[k] = float(bundle[k])
    logger.info(f"  [{model_key}] Bundle: layer={bundle.get('best_layer')}, "
                f"dim={bundle.get('hidden_dim')}, "
                f"AUROC={bundle.get('drift_auroc', 0):.4f}")
    return bundle


def load_final_results(model_dir, model_key):
    path = Path(model_dir) / model_key / "final_results.json"
    if not path.exists():
        return None
    with open(path) as f:
        return json.load(f)


# ─────────────────────────────────────────────────────────────────────────────
# PROBE FITTING (lightweight — for scoring shared queries)
# ─────────────────────────────────────────────────────────────────────────────

def soft_threshold(w, lam):
    import torch
    return torch.sign(w) * torch.clamp(torch.abs(w) - lam, min=0.0)


def fit_quick_probe(X_np, y_np, device="cuda:0", lam=1e-3, max_iter=500):
    """Fast probe fit for cross-model scoring."""
    import torch
    X = np.nan_to_num(X_np.astype(np.float32), nan=0., posinf=1e4, neginf=-1e4)
    X = np.clip(X, -1e4, 1e4)
    m = X.mean(0, keepdims=True)
    s = X.std(0, keepdims=True) + 1e-8
    Xt = torch.tensor((X - m) / s, dtype=torch.float32, device=device)
    yt = torch.tensor(y_np.astype(np.float32), device=device)

    w = torch.zeros(Xt.shape[1], device=device)
    b = torch.zeros(1, device=device)
    lr = 1.0
    for _ in range(max_iter):
        z = torch.clamp(Xt @ w + b, -30, 30)
        p = torch.sigmoid(z)
        L = -((yt * torch.log(p + 1e-12)) +
              (1 - yt) * torch.log(1 - p + 1e-12)).mean()
        e = p - yt
        gw = (Xt.T @ e) / len(yt)
        gb = e.mean(keepdim=True)
        wt = soft_threshold(w - lr * gw, lr * lam)
        bt = b - lr * gb
        Lt = -((yt * torch.log(torch.sigmoid(torch.clamp(Xt @ wt + bt, -30, 30)) + 1e-12)) +
               (1 - yt) * torch.log(1 - torch.sigmoid(torch.clamp(Xt @ wt + bt, -30, 30)) + 1e-12)).mean()
        if Lt > L + 1e-4:
            lr *= 0.5
        else:
            lr = min(lr * 1.05, 10.0)
        if (wt - w).abs().max().item() < 1e-6:
            w, b = wt, bt
            break
        w, b = wt, bt

    def score(X_new):
        Xn = np.nan_to_num(X_new.astype(np.float32), nan=0., posinf=1e4, neginf=-1e4)
        Xn = np.clip(Xn, -1e4, 1e4)
        Xn = torch.tensor((Xn - m) / s, dtype=torch.float32, device=device)
        with torch.no_grad():
            return torch.sigmoid(torch.clamp(Xn @ w + b, -30, 30)).cpu().numpy()

    return score, w.cpu().numpy()


# ─────────────────────────────────────────────────────────────────────────────
# [CM-1] CKA ANALYSIS
# ─────────────────────────────────────────────────────────────────────────────

def linear_cka(Xa, Xb):
    """Centered Kernel Alignment between two representation matrices."""
    def _center(K):
        n = K.shape[0]
        H = np.eye(n) - 1.0 / n
        return H @ K @ H
    Ka = _center(Xa @ Xa.T)
    Kb = _center(Xb @ Xb.T)
    num = np.linalg.norm(Ka.T @ Kb, "fro")
    den = np.linalg.norm(Ka, "fro") * np.linalg.norm(Kb, "fro")
    return float(num / (den + 1e-12))


def cka_analysis(res_a, res_b, key_a, key_b, quick=False):
    """
    [CM-1] CKA between two models.
    If quick=False: full L_A × L_B heatmap.
    If quick=True: just best-layer CKA.
    """
    logger.info(f"[CM-1] CKA: {key_a} vs {key_b}")

    # Build shared query lookup
    qa = {r["query"]: r for r in res_a}
    qb = {r["query"]: r for r in res_b}
    shared = sorted(set(qa) & set(qb))
    logger.info(f"  Shared queries: {len(shared)}")

    if len(shared) < 50:
        logger.warning("  Too few shared queries for CKA")
        return None

    # Subsample for speed (CKA is O(n²))
    if len(shared) > 2000:
        np.random.seed(42)
        shared = list(np.random.choice(shared, 2000, replace=False))

    layers_a = sorted(res_a[0]["hidden_states"].keys())
    layers_b = sorted(res_b[0]["hidden_states"].keys())

    if quick:
        # Just best layers
        best_a = layers_a[-5:]  # top 5 layers
        best_b = layers_b[-5:]
    else:
        # Sample layers evenly (max 10 per model for tractability)
        step_a = max(1, len(layers_a) // 10)
        step_b = max(1, len(layers_b) // 10)
        best_a = layers_a[::step_a]
        best_b = layers_b[::step_b]

    cka_mat = np.zeros((len(best_a), len(best_b)))
    for i, la in enumerate(best_a):
        Xa = np.array([qa[q]["hidden_states"][la] for q in shared])
        for j, lb in enumerate(best_b):
            Xb = np.array([qb[q]["hidden_states"][lb] for q in shared])
            cka_mat[i, j] = linear_cka(Xa, Xb)
        if (i + 1) % 3 == 0:
            logger.info(f"  CKA row {i+1}/{len(best_a)}")

    best_cka = float(cka_mat.max())
    logger.info(f"  Best CKA: {best_cka:.4f}")

    return {
        "layers_a": best_a, "layers_b": best_b,
        "cka_matrix": cka_mat.tolist(),
        "best_cka": best_cka,
        "n_shared": len(shared),
    }


# ─────────────────────────────────────────────────────────────────────────────
# [CM-2] SCORE CORRELATION
# ─────────────────────────────────────────────────────────────────────────────

def score_correlation(res_a, res_b, key_a, key_b, bundle_a, bundle_b, device):
    """
    [CM-2] Train probe on each model, score shared queries, correlate.
    """
    from sklearn.metrics import roc_auc_score
    logger.info(f"[CM-2] Score correlation: {key_a} vs {key_b}")

    qa = {r["query"]: r for r in res_a}
    qb = {r["query"]: r for r in res_b}
    shared = sorted(set(qa) & set(qb))
    logger.info(f"  Shared: {len(shared)}")

    if len(shared) < 50:
        return None

    bl_a = int(bundle_a["best_layer"])
    bl_b = int(bundle_b["best_layer"])

    # Train probes on full data
    X_a = np.array([r["hidden_states"][bl_a] for r in res_a])
    y_a = np.array([int(r["is_drifted"]) for r in res_a])
    X_b = np.array([r["hidden_states"][bl_b] for r in res_b])
    y_b = np.array([int(r["is_drifted"]) for r in res_b])

    score_a, _ = fit_quick_probe(X_a, y_a, device)
    score_b, _ = fit_quick_probe(X_b, y_b, device)

    # Score shared queries
    Xa_shared = np.array([qa[q]["hidden_states"][bl_a] for q in shared])
    Xb_shared = np.array([qb[q]["hidden_states"][bl_b] for q in shared])
    sa = score_a(Xa_shared)
    sb = score_b(Xb_shared)

    # Labels for shared
    ya_shared = np.array([int(qa[q]["is_drifted"]) for q in shared])
    yb_shared = np.array([int(qb[q]["is_drifted"]) for q in shared])

    corr = float(np.corrcoef(sa, sb)[0, 1])
    try:
        auroc_a = roc_auc_score(ya_shared, sa)
        auroc_b = roc_auc_score(yb_shared, sb)
    except Exception:
        auroc_a = auroc_b = 0.5

    logger.info(f"  Score corr: {corr:.4f}  "
                f"AUROC_a={auroc_a:.4f}  AUROC_b={auroc_b:.4f}")

    return {
        "correlation": corr,
        "auroc_a_on_shared": auroc_a,
        "auroc_b_on_shared": auroc_b,
        "n_shared": len(shared),
        "scores_a": sa.tolist(),
        "scores_b": sb.tolist(),
    }


# ─────────────────────────────────────────────────────────────────────────────
# [CM-3] DIFFERENTIAL FACTS
# ─────────────────────────────────────────────────────────────────────────────

def differential_facts(res_a, res_b, key_a, key_b, bundle_a, bundle_b, device):
    """
    [CM-3] Queries where is_drifted differs between models.
    Each probe should detect its own model's drift correctly.
    """
    from sklearn.metrics import roc_auc_score
    logger.info(f"[CM-3] Differential facts: {key_a} vs {key_b}")

    qa = {r["query"]: r for r in res_a}
    qb = {r["query"]: r for r in res_b}
    shared = sorted(set(qa) & set(qb))

    # Find differential: drifted for A but not B, or vice versa
    diff_queries = [q for q in shared
                    if qa[q]["is_drifted"] != qb[q]["is_drifted"]]
    logger.info(f"  Shared={len(shared)}, Differential={len(diff_queries)}")

    if len(diff_queries) < 20:
        logger.warning("  Too few differential facts")
        return None

    bl_a = int(bundle_a["best_layer"])
    bl_b = int(bundle_b["best_layer"])

    # Train probes
    X_a = np.array([r["hidden_states"][bl_a] for r in res_a])
    y_a = np.array([int(r["is_drifted"]) for r in res_a])
    X_b = np.array([r["hidden_states"][bl_b] for r in res_b])
    y_b = np.array([int(r["is_drifted"]) for r in res_b])
    score_a, _ = fit_quick_probe(X_a, y_a, device)
    score_b, _ = fit_quick_probe(X_b, y_b, device)

    # Score differential queries
    Xa_d = np.array([qa[q]["hidden_states"][bl_a] for q in diff_queries])
    Xb_d = np.array([qb[q]["hidden_states"][bl_b] for q in diff_queries])
    sa = score_a(Xa_d)
    sb = score_b(Xb_d)
    la = np.array([int(qa[q]["is_drifted"]) for q in diff_queries])
    lb = np.array([int(qb[q]["is_drifted"]) for q in diff_queries])

    try:
        auroc_a = roc_auc_score(la, sa)
    except Exception:
        auroc_a = 0.5
    try:
        auroc_b = roc_auc_score(lb, sb)
    except Exception:
        auroc_b = 0.5

    # Anti-correlation: when A says drifted and B says stable,
    # score_a should be high and score_b should be low
    score_corr = float(np.corrcoef(sa, sb)[0, 1])

    # Count categories
    a_only = sum(1 for q in diff_queries
                 if qa[q]["is_drifted"] and not qb[q]["is_drifted"])
    b_only = sum(1 for q in diff_queries
                 if not qa[q]["is_drifted"] and qb[q]["is_drifted"])

    logger.info(f"  AUROC_a={auroc_a:.4f}  AUROC_b={auroc_b:.4f}  "
                f"score_corr={score_corr:.4f}")
    logger.info(f"  A-only drifted: {a_only}  B-only drifted: {b_only}")

    return {
        "n_differential": len(diff_queries),
        "n_shared": len(shared),
        "a_only_drifted": a_only,
        "b_only_drifted": b_only,
        "auroc_a": auroc_a,
        "auroc_b": auroc_b,
        "score_correlation": score_corr,
        "scores_a": sa.tolist(),
        "scores_b": sb.tolist(),
        "labels_a": la.tolist(),
        "labels_b": lb.tolist(),
    }


# ─────────────────────────────────────────────────────────────────────────────
# [CM-4] LAYER CORRESPONDENCE
# ─────────────────────────────────────────────────────────────────────────────

def layer_correspondence(all_bundles, all_final):
    """
    [CM-4] Best drift layer as fraction of total depth.
    If all models peak at ~80%, drift localization is universal.
    """
    logger.info("[CM-4] Layer correspondence")
    data = {}
    for key in all_bundles:
        bl = int(all_bundles[key]["best_layer"])
        total = int(all_bundles[key].get("hidden_dim", 0))
        # Get total layers from final results
        fr = all_final.get(key, {})
        n_layers = fr.get("best_layer_results", {}).get("layer", bl) + 1
        # Better: look at probe stability layers
        stab = fr.get("probe_stability", {})
        if "layers" in stab and len(stab["layers"]) > 0:
            n_layers = max(stab["layers"]) + 1

        frac = bl / max(n_layers, 1)
        auroc = float(all_bundles[key].get("drift_auroc", 0))
        data[key] = {
            "best_layer": bl,
            "n_layers": n_layers,
            "fraction": frac,
            "auroc": auroc,
        }
        logger.info(f"  {key}: L{bl}/{n_layers} = {frac:.1%}  "
                    f"AUROC={auroc:.4f}")

    fracs = [v["fraction"] for v in data.values()]
    mean_frac = float(np.mean(fracs))
    std_frac = float(np.std(fracs))
    logger.info(f"  Mean fraction: {mean_frac:.1%} +/- {std_frac:.1%}")

    return {
        "per_model": data,
        "mean_fraction": mean_frac,
        "std_fraction": std_frac,
    }


# ─────────────────────────────────────────────────────────────────────────────
# [CM-5] NEURON OVERLAP (same-dim models only)
# ─────────────────────────────────────────────────────────────────────────────

def neuron_overlap(bundle_a, bundle_b, key_a, key_b):
    """
    [CM-5] For same-dimension models: do the same neuron indices carry drift?
    """
    dim_a = int(bundle_a["hidden_dim"])
    dim_b = int(bundle_b["hidden_dim"])

    if dim_a != dim_b:
        logger.info(f"[CM-5] {key_a}({dim_a}) vs {key_b}({dim_b}): "
                    f"dim mismatch, skipping")
        return None

    logger.info(f"[CM-5] Neuron overlap: {key_a} vs {key_b} (dim={dim_a})")

    w_a = bundle_a["w_drift"]
    w_b = bundle_b["w_drift"]

    active_a = set(np.where(w_a != 0)[0])
    active_b = set(np.where(w_b != 0)[0])

    inter = len(active_a & active_b)
    union = len(active_a | active_b)
    jacc = inter / union if union > 0 else 0.0

    # Cosine of weight vectors (even though from different models)
    cos = float(np.dot(w_a, w_b) / (np.linalg.norm(w_a) * np.linalg.norm(w_b) + 1e-12))

    # Top-k overlap
    top100_a = set(np.argsort(np.abs(w_a))[-100:])
    top100_b = set(np.argsort(np.abs(w_b))[-100:])
    top100_overlap = len(top100_a & top100_b) / 100.0

    logger.info(f"  Active: A={len(active_a)}, B={len(active_b)}")
    logger.info(f"  Jaccard: {jacc:.4f}  Cosine: {cos:.4f}  "
                f"Top-100 overlap: {top100_overlap:.2%}")

    return {
        "dim": dim_a,
        "n_active_a": len(active_a),
        "n_active_b": len(active_b),
        "intersection": inter,
        "union": union,
        "jaccard": jacc,
        "cosine": cos,
        "top100_overlap": top100_overlap,
    }


# ─────────────────────────────────────────────────────────────────────────────
# [CM-6] UNIVERSALITY SCORE
# ─────────────────────────────────────────────────────────────────────────────

def universality_score(all_cka, all_corr, all_diff, all_layer_corr,
                       n_bootstrap=1000):
    """
    [CM-6] Aggregate metric: geometric mean of CKA, score correlation,
    differential AUROC, and layer consistency.
    """
    logger.info("[CM-6] Universality score")

    components = {}

    # Mean best CKA across pairs
    cka_vals = [v["best_cka"] for v in all_cka.values() if v]
    if cka_vals:
        components["mean_cka"] = float(np.mean(cka_vals))

    # Mean score correlation
    corr_vals = [v["correlation"] for v in all_corr.values() if v]
    if corr_vals:
        components["mean_score_corr"] = float(np.mean(corr_vals))

    # Mean differential AUROC
    diff_aurocs = []
    for v in all_diff.values():
        if v:
            diff_aurocs.extend([v["auroc_a"], v["auroc_b"]])
    if diff_aurocs:
        components["mean_diff_auroc"] = float(np.mean(diff_aurocs))

    # Layer consistency (1 - std of fractions)
    if all_layer_corr:
        components["layer_consistency"] = float(
            1.0 - all_layer_corr.get("std_fraction", 0.5))

    if not components:
        return None

    vals = list(components.values())
    # Geometric mean
    geo_mean = float(np.exp(np.mean(np.log(np.clip(vals, 1e-6, None)))))

    # Bootstrap CI
    boot = []
    for _ in range(n_bootstrap):
        idx = np.random.choice(len(vals), len(vals), replace=True)
        boot.append(np.exp(np.mean(np.log(np.clip(np.array(vals)[idx], 1e-6, None)))))
    ci_lo = float(np.percentile(boot, 2.5))
    ci_hi = float(np.percentile(boot, 97.5))

    logger.info(f"  Components: {components}")
    logger.info(f"  Universality: {geo_mean:.4f} [{ci_lo:.4f}, {ci_hi:.4f}]")

    return {
        "components": components,
        "universality_score": geo_mean,
        "ci_95": [ci_lo, ci_hi],
    }


# ─────────────────────────────────────────────────────────────────────────────
# FIGURES
# ─────────────────────────────────────────────────────────────────────────────

def save_cross_figures(out_dir, keys, all_cka, all_corr, all_diff,
                       layer_data, neuron_data, univ_data):
    import matplotlib
    matplotlib.use("Agg")
    import matplotlib.pyplot as plt

    fig_dir = Path(out_dir) / "figures"
    fig_dir.mkdir(parents=True, exist_ok=True)

    P = {"drift": "#e74c3c", "unc": "#3498db", "corr": "#2ecc71",
         "null": "#9b59b6", "neu": "#e67e22"}

    # ── CM-1: CKA heatmaps ───────────────────────────────────────────────
    cka_pairs = [(k, v) for k, v in all_cka.items() if v]
    if cka_pairs:
        n_pairs = len(cka_pairs)
        fig, axes = plt.subplots(1, n_pairs, figsize=(8 * n_pairs, 7))
        if n_pairs == 1:
            axes = [axes]
        fig.suptitle("[CM-1] Cross-Model CKA", fontsize=16, fontweight="bold")
        for ax, (pair_key, data) in zip(axes, cka_pairs):
            mat = np.array(data["cka_matrix"])
            im = ax.imshow(mat, cmap="viridis", vmin=0, vmax=1, aspect="auto")
            la = data["layers_a"]
            lb = data["layers_b"]
            step_a = max(1, len(la) // 6)
            step_b = max(1, len(lb) // 6)
            ax.set_xticks(range(0, len(lb), step_b))
            ax.set_yticks(range(0, len(la), step_a))
            ax.set_xticklabels([lb[i] for i in range(0, len(lb), step_b)])
            ax.set_yticklabels([la[i] for i in range(0, len(la), step_a)])
            parts = pair_key.split("_vs_")
            ax.set(xlabel=f"{parts[1]} layer", ylabel=f"{parts[0]} layer",
                   title=f"{pair_key}\nbest={data['best_cka']:.3f}")
            plt.colorbar(im, ax=ax, shrink=0.8)
        plt.tight_layout()
        plt.savefig(fig_dir / "fig_cm1_cka.png", dpi=300, bbox_inches="tight")
        plt.close()
        logger.info("  fig_cm1 saved")

    # ── CM-2: Score correlation matrix ────────────────────────────────────
    if len(keys) >= 2 and all_corr:
        n = len(keys)
        mat = np.eye(n)
        for pair_key, data in all_corr.items():
            if data is None:
                continue
            parts = pair_key.split("_vs_")
            if len(parts) == 2:
                i = keys.index(parts[0]) if parts[0] in keys else -1
                j = keys.index(parts[1]) if parts[1] in keys else -1
                if i >= 0 and j >= 0:
                    mat[i, j] = mat[j, i] = data["correlation"]

        fig, ax = plt.subplots(figsize=(8, 7))
        im = ax.imshow(mat, cmap="RdBu_r", vmin=-1, vmax=1)
        ax.set_xticks(range(n))
        ax.set_yticks(range(n))
        ax.set_xticklabels(keys, fontsize=12, rotation=20)
        ax.set_yticklabels(keys, fontsize=12)
        for i in range(n):
            for j in range(n):
                c = "white" if abs(mat[i, j]) > 0.5 else "black"
                ax.text(j, i, f"{mat[i,j]:.3f}", ha="center", va="center",
                        fontsize=13, fontweight="bold", color=c)
        ax.set_title("[CM-2] Drift Score Correlation Matrix", fontsize=14)
        plt.colorbar(im, ax=ax, shrink=0.8)
        plt.tight_layout()
        plt.savefig(fig_dir / "fig_cm2_corr.png", dpi=300, bbox_inches="tight")
        plt.close()
        logger.info("  fig_cm2 saved")

    # ── CM-3: Differential facts ──────────────────────────────────────────
    diff_pairs = [(k, v) for k, v in all_diff.items() if v]
    if diff_pairs:
        n_pairs = min(len(diff_pairs), 4)
        fig, axes = plt.subplots(1, n_pairs, figsize=(7 * n_pairs, 6))
        if n_pairs == 1:
            axes = [axes]
        fig.suptitle("[CM-3] Differential Facts", fontsize=16, fontweight="bold")
        for ax, (pair_key, data) in zip(axes, diff_pairs[:n_pairs]):
            sa = np.array(data["scores_a"])
            sb = np.array(data["scores_b"])
            la = np.array(data["labels_a"])
            lb = np.array(data["labels_b"])
            # Color by which model says drifted
            a_drifted = la.astype(bool) & ~lb.astype(bool)
            b_drifted = ~la.astype(bool) & lb.astype(bool)
            ax.scatter(sa[a_drifted], sb[a_drifted], c=P["drift"], alpha=0.5,
                       s=30, label="A=drifted, B=stable")
            ax.scatter(sa[b_drifted], sb[b_drifted], c=P["unc"], alpha=0.5,
                       s=30, label="A=stable, B=drifted")
            ax.plot([0, 1], [0, 1], "k--", alpha=0.3)
            ax.axhline(0.5, color="gray", ls=":", alpha=0.3)
            ax.axvline(0.5, color="gray", ls=":", alpha=0.3)
            parts = pair_key.split("_vs_")
            ax.set(xlabel=f"{parts[0]} score", ylabel=f"{parts[1]} score",
                   title=f"{pair_key}\nr={data['score_correlation']:.3f}")
            ax.legend(fontsize=8)
            ax.grid(alpha=0.2)
        plt.tight_layout()
        plt.savefig(fig_dir / "fig_cm3_diff.png", dpi=300, bbox_inches="tight")
        plt.close()
        logger.info("  fig_cm3 saved")

    # ── CM-4: Layer correspondence ────────────────────────────────────────
    if layer_data and "per_model" in layer_data:
        pm = layer_data["per_model"]
        models = sorted(pm.keys())
        fig, axes = plt.subplots(1, 2, figsize=(14, 6))
        fig.suptitle("[CM-4] Layer Correspondence", fontsize=14,
                     fontweight="bold")

        # Absolute layers
        x = np.arange(len(models))
        bls = [pm[m]["best_layer"] for m in models]
        nls = [pm[m]["n_layers"] for m in models]
        ax = axes[0]
        ax.bar(x, bls, color=P["drift"], edgecolor="black", lw=0.5,
               label="Best layer")
        ax.bar(x, [n - b for b, n in zip(bls, nls)], bottom=bls,
               color="#ecf0f1", edgecolor="black", lw=0.5, label="Remaining")
        ax.set_xticks(x)
        ax.set_xticklabels(models, fontsize=11)
        ax.set(ylabel="Layer", title="Best Drift Layer (absolute)")
        ax.legend()
        ax.grid(alpha=0.3, axis="y")

        # Fraction
        ax = axes[1]
        fracs = [pm[m]["fraction"] for m in models]
        bars = ax.bar(x, fracs, color=P["neu"], edgecolor="black", lw=0.5)
        ax.axhline(layer_data["mean_fraction"], color="red", ls="--", lw=2,
                    label=f"Mean: {layer_data['mean_fraction']:.1%}")
        ax.fill_between(
            [-0.5, len(models) - 0.5],
            layer_data["mean_fraction"] - layer_data["std_fraction"],
            layer_data["mean_fraction"] + layer_data["std_fraction"],
            alpha=0.2, color="red")
        ax.set_xticks(x)
        ax.set_xticklabels(models, fontsize=11)
        ax.set(ylabel="Fraction of depth", title="Best Layer as % of Depth",
               ylim=(0, 1))
        ax.legend()
        ax.grid(alpha=0.3, axis="y")
        plt.tight_layout()
        plt.savefig(fig_dir / "fig_cm4_layers.png",
                    dpi=300, bbox_inches="tight")
        plt.close()
        logger.info("  fig_cm4 saved")

    # ── CM-6: Summary ─────────────────────────────────────────────────────
    if univ_data:
        fig, ax = plt.subplots(figsize=(10, 6))
        comp = univ_data["components"]
        names = list(comp.keys())
        vals = list(comp.values())
        x = np.arange(len(names))
        colors = [P["drift"], P["unc"], P["corr"], P["neu"]][:len(names)]
        ax.bar(x, vals, color=colors, edgecolor="black", lw=0.5, alpha=0.8)
        ax.axhline(univ_data["universality_score"], color="red", ls="--",
                    lw=2.5,
                    label=f"Geo mean: {univ_data['universality_score']:.3f} "
                          f"[{univ_data['ci_95'][0]:.3f}, "
                          f"{univ_data['ci_95'][1]:.3f}]")
        ax.set_xticks(x)
        ax.set_xticklabels([n.replace("_", "\n") for n in names], fontsize=10)
        ax.set(ylabel="Score", title="[CM-6] Universality Score Components",
               ylim=(0, 1.1))
        ax.legend(fontsize=11)
        ax.grid(alpha=0.3, axis="y")
        plt.tight_layout()
        plt.savefig(fig_dir / "fig_cm6_summary.png",
                    dpi=300, bbox_inches="tight")
        plt.close()
        logger.info("  fig_cm6 saved")

    logger.info(f"All cross-model figures -> {fig_dir}")


# ─────────────────────────────────────────────────────────────────────────────
# MAIN
# ─────────────────────────────────────────────────────────────────────────────

def main():
    p = argparse.ArgumentParser(
        description="Cross-model drift analysis",
        formatter_class=argparse.ArgumentDefaultsHelpFormatter)
    p.add_argument("--models", nargs="+", default=None,
                   help="Model keys to compare")
    p.add_argument("--all", action="store_true",
                   help="Use all models with available caches")
    p.add_argument("--config", default="models.yaml")
    p.add_argument("--output_dir", default=None)
    p.add_argument("--device", default="cuda:0")
    p.add_argument("--quick", action="store_true",
                   help="Skip full-layer CKA, just best-layer")
    args = p.parse_args()

    cfg = load_config(args.config)
    defaults = cfg.get("defaults", {})
    output_dir = args.output_dir or defaults.get("output_dir",
                    "data/experiments/v4")

    # Determine which models to use
    if args.all:
        model_keys = list(cfg["models"].keys())
    elif args.models:
        model_keys = args.models
    else:
        logger.error("Specify --models or --all")
        return

    # Load caches and bundles
    all_results = {}
    all_bundles = {}
    all_final = {}
    for key in model_keys:
        res = load_cache(output_dir, key)
        bundle = load_probe_bundle(output_dir, key)
        final = load_final_results(output_dir, key)
        if res and bundle:
            all_results[key] = res
            all_bundles[key] = bundle
            if final:
                all_final[key] = final

    keys = sorted(all_results.keys())
    logger.info(f"\nModels available: {keys}")

    if len(keys) < 2:
        logger.error("Need at least 2 models with caches + bundles")
        return

    cross_dir = Path(output_dir) / "cross_model"
    cross_dir.mkdir(parents=True, exist_ok=True)

    # Run all 6 experiments
    all_cka = {}
    all_corr = {}
    all_diff = {}
    all_neuron = {}

    for i, ka in enumerate(keys):
        for j, kb in enumerate(keys):
            if i >= j:
                continue
            pair = f"{ka}_vs_{kb}"
            logger.info(f"\n{'─'*50}")
            logger.info(f"  {pair}")
            logger.info(f"{'─'*50}")

            # [CM-1] CKA
            all_cka[pair] = cka_analysis(
                all_results[ka], all_results[kb], ka, kb, quick=args.quick)

            # [CM-2] Score correlation
            all_corr[pair] = score_correlation(
                all_results[ka], all_results[kb], ka, kb,
                all_bundles[ka], all_bundles[kb], args.device)

            # [CM-3] Differential facts
            all_diff[pair] = differential_facts(
                all_results[ka], all_results[kb], ka, kb,
                all_bundles[ka], all_bundles[kb], args.device)

            # [CM-5] Neuron overlap
            all_neuron[pair] = neuron_overlap(
                all_bundles[ka], all_bundles[kb], ka, kb)

    # [CM-4] Layer correspondence
    layer_data = layer_correspondence(all_bundles, all_final)

    # [CM-6] Universality score
    univ_data = universality_score(all_cka, all_corr, all_diff, layer_data)

    # Save results
    results = {
        "models": keys,
        "cka": {k: v for k, v in all_cka.items()},
        "score_correlation": {k: v for k, v in all_corr.items()},
        "differential_facts": {k: v for k, v in all_diff.items()},
        "neuron_overlap": {k: v for k, v in all_neuron.items() if v},
        "layer_correspondence": layer_data,
        "universality": univ_data,
        "timestamp": datetime.now().isoformat(),
    }

    from datetime import datetime
    out_path = cross_dir / "cross_model_results.json"
    with open(out_path, "w") as f:
        json.dump(results, f, indent=2, default=str)
    logger.info(f"\nResults saved: {out_path}")

    # Figures
    save_cross_figures(str(cross_dir), keys, all_cka, all_corr, all_diff,
                       layer_data, all_neuron, univ_data)

    # Print summary
    print(f"\n{'='*70}")
    print(f"  CROSS-MODEL SUMMARY")
    print(f"{'='*70}")
    for pair, data in all_corr.items():
        if data:
            print(f"  {pair}: score_corr={data['correlation']:.4f}")
    for pair, data in all_diff.items():
        if data:
            print(f"  {pair}: diff_AUROC_a={data['auroc_a']:.4f}  "
                  f"diff_AUROC_b={data['auroc_b']:.4f}  "
                  f"n_diff={data['n_differential']}")
    if layer_data:
        print(f"\n  Layer correspondence: "
              f"{layer_data['mean_fraction']:.1%} +/- "
              f"{layer_data['std_fraction']:.1%}")
    if univ_data:
        print(f"\n  UNIVERSALITY SCORE: "
              f"{univ_data['universality_score']:.4f} "
              f"[{univ_data['ci_95'][0]:.4f}, {univ_data['ci_95'][1]:.4f}]")
    print(f"{'='*70}")


if __name__ == "__main__":
    main()