year_controlled_analysis.py

"""
Year-Controlled Analysis — Qwen2.5 (v2 — all issues fixed)
=============================================================
Uses EXISTING cached hidden states. No model loading needed.

Fixes:
  - Year type normalization (int vs str)
  - Source/format confound tracking
  - Zip verification (query match between cache and dataset)
  - Proper category handling

Usage:
    cd ~/svd_kg/knowledge_drift
    python year_controlled_analysis.py
"""

import json, os, sys, time
import numpy as np
import torch
import torch.nn as nn
from collections import Counter, defaultdict

print("=" * 70)
print("  YEAR-CONTROLLED ANALYSIS v2 — Qwen2.5")
print("  Using cached hidden states (no GPU extraction needed)")
print("=" * 70)

# ============================================================
# LOAD CACHED STATES + DATASET
# ============================================================

CACHE_PATH = "data/experiments/tier1_qwen25_v2/cached_states.npz"
DATASET_PATH = "data/tier1_qwen25.json"
OUTPUT_DIR = "data/experiments/tier1_qwen25_v2/year_controlled"
os.makedirs(OUTPUT_DIR, exist_ok=True)

print("\nLoading cached hidden states...")
t0 = time.time()
results = np.load(CACHE_PATH, allow_pickle=True)["results"].tolist()
print(f"  Loaded {len(results)} cached results in {time.time()-t0:.1f}s")

print("Loading dataset for metadata...")
with open(DATASET_PATH) as f:
    ds = json.load(f)
samples_meta = ds.get("samples", ds)
print(f"  Loaded {len(samples_meta)} dataset samples")

# ============================================================
# VERIFY ZIP ALIGNMENT (critical safety check)
# ============================================================

print("\nVerifying cache-dataset alignment...")
mismatches = 0
for i in range(min(100, len(results))):
    cached_q = results[i].get("query", "")[:50]
    meta_q = samples_meta[i].get("query", "")[:50]
    if cached_q != meta_q:
        mismatches += 1
        if mismatches <= 3:
            print(f"  MISMATCH at {i}: cache='{cached_q}' vs meta='{meta_q}'")

if mismatches > 0:
    print(f"  ⚠️ {mismatches}/100 mismatches! Zip alignment is BROKEN.")
    print("  Cannot proceed — cached states don't match dataset order.")
    sys.exit(1)
else:
    print(f"  ✅ First 100 queries match. Alignment verified.")

# ============================================================
# ATTACH METADATA FROM DATASET TO CACHED RESULTS
# ============================================================

print("\nAttaching metadata...")
for r, s in zip(results, samples_meta):
    # Normalize year to int
    raw_year = s.get("year", 0)
    try:
        r["year"] = int(raw_year) if raw_year else 0
    except (ValueError, TypeError):
        r["year"] = 0
    
    r["dataset_source"] = s.get("dataset_source", "unknown")
    r["drift_date"] = s.get("drift_date", "")
    r["sample_id"] = s.get("sample_id", "")
    
    # Extract year mentioned in query text
    query_year = None
    q = r.get("query", "")
    for y in range(2010, 2027):
        if str(y) in q:
            query_year = y
            break
    r["query_year"] = query_year
    r["has_year_in_query"] = query_year is not None

# Print summary of attached metadata
year_counts = Counter(r["year"] for r in results)
source_counts = Counter(r["dataset_source"] for r in results)
qy_counts = Counter(r["has_year_in_query"] for r in results)

print(f"  Years: {dict(sorted(year_counts.items()))}")
print(f"  Sources: {dict(source_counts)}")
print(f"  Has year in query: {dict(qy_counts)}")
print(f"  Drifted: {sum(1 for r in results if r['is_drifted'])}")
print(f"  Stable: {sum(1 for r in results if not r['is_drifted'])}")

# ============================================================
# PROBING FUNCTIONS (minimal, from disentanglement_v2.py)
# ============================================================

def numpy_auroc(y_true, y_score):
    y_true = np.asarray(y_true, dtype=np.float64)
    y_score = np.asarray(y_score, dtype=np.float64)
    if len(np.unique(y_true)) < 2:
        return 0.5
    n_pos = y_true.sum()
    n_neg = len(y_true) - n_pos
    if n_pos == 0 or n_neg == 0:
        return 0.5
    asc = np.argsort(y_score, kind='stable')
    y_sorted = y_true[asc]
    s_sorted = y_score[asc]
    _, inverse, counts = np.unique(s_sorted, return_inverse=True, return_counts=True)
    cumcounts = np.cumsum(counts)
    avg_ranks = np.empty(len(y_true), dtype=np.float64)
    for i, c in enumerate(counts):
        start = cumcounts[i] - c
        end = cumcounts[i]
        avg_ranks[inverse == i] = (start + end + 1) / 2.0
    rank_sum = avg_ranks[y_sorted == 1].sum()
    auroc = (rank_sum - n_pos * (n_pos + 1) / 2) / (n_pos * n_neg)
    return float(np.clip(auroc, 0.0, 1.0))

def stratified_kfold(y, n_splits, seed=42):
    rng = np.random.RandomState(seed)
    y = np.asarray(y)
    folds = [[] for _ in range(n_splits)]
    for cls in np.unique(y):
        idx = np.where(y == cls)[0].copy()
        rng.shuffle(idx)
        for i, v in enumerate(idx):
            folds[i % n_splits].append(v)
    splits = []
    for i in range(n_splits):
        val_idx = np.array(folds[i], dtype=np.int64)
        train_idx = np.concatenate([np.array(folds[j], dtype=np.int64) for j in range(n_splits) if j != i])
        splits.append((train_idx, val_idx))
    return splits

def prepare_tensors(X_np, y_np, device):
    X = np.nan_to_num(np.clip(X_np.astype(np.float32), -1e4, 1e4))
    mean, std = X.mean(0, keepdims=True), X.std(0, keepdims=True) + 1e-8
    return (torch.tensor((X - mean) / std, dtype=torch.float32, device=device),
            torch.tensor(y_np.astype(np.float32), device=device), mean, std)

class LinearProbeGPU:
    def __init__(self, input_dim, C=1.0, lr=0.05, max_iter=500, device="cuda"):
        self.C, self.lr, self.max_iter, self.device = C, lr, max_iter, device
        self.model = nn.Linear(input_dim, 1, bias=True).to(device)
        self.coef_ = None
    def fit(self, X_t, y_t):
        X_t, y_t = X_t.contiguous(), y_t.contiguous()
        nn.init.zeros_(self.model.weight); nn.init.zeros_(self.model.bias)
        wd = 1.0 / (self.C * len(y_t) + 1e-8)
        opt = torch.optim.LBFGS(self.model.parameters(), lr=self.lr,
                                 max_iter=self.max_iter, tolerance_grad=1e-5, tolerance_change=1e-7)
        crit = nn.BCEWithLogitsLoss()
        def closure():
            opt.zero_grad()
            loss = crit(self.model(X_t).squeeze(1), y_t) + wd * self.model.weight.pow(2).sum()
            loss.backward(); return loss
        opt.step(closure)
        self.coef_ = [self.model.weight.detach().cpu().numpy().flatten()]
        return self
    def predict_proba(self, X_t):
        with torch.no_grad():
            p = torch.sigmoid(self.model(X_t.contiguous()).squeeze(1)).cpu().numpy()
        return np.column_stack([1 - p, p])

def best_auroc(X_np, y_np, device="cuda", n_splits=3):
    """Try multiple C values, return best AUROC."""
    best = 0.0
    for C in [0.01, 0.1, 1.0]:
        X_t, y_t, mean, std = prepare_tensors(X_np, y_np, device)
        dim = X_t.shape[1]
        mc = min(int((y_np == 0).sum()), int((y_np == 1).sum()))
        ns = min(n_splits, mc)
        aurocs = []
        if ns >= 2:
            for tr, va in stratified_kfold(y_np, ns):
                tr_t = torch.from_numpy(tr).long().to(device)
                va_t = torch.from_numpy(va).long().to(device)
                p = LinearProbeGPU(dim, C=C, device=device)
                p.fit(X_t[tr_t].clone().contiguous(), y_t[tr_t].clone().contiguous())
                probs = p.predict_proba(X_t[va_t].clone().contiguous())[:, 1]
                if len(np.unique(y_np[va])) > 1:
                    aurocs.append(numpy_auroc(y_np[va], probs))
        a = float(np.mean(aurocs)) if aurocs else 0.5
        if a > best:
            best = a
    return best

# ============================================================
# DEVICE
# ============================================================

device = "cuda:0" if torch.cuda.is_available() else "cpu"
print(f"\nProbe device: {device}")

# ============================================================
# TEST LAYERS
# ============================================================

test_layers = [0, 3, 7, 13, 20, 27]

def run_analysis(name, subset, test_layers, device, baseline=None):
    """Run probe on a subset and print results."""
    n_d = sum(1 for r in subset if r["is_drifted"])
    n_s = len(subset) - n_d
    print(f"  Samples: {len(subset)}, Drifted: {n_d}, Stable: {n_s}")
    
    if n_d < 10 or n_s < 10:
        print(f"  ⚠️ Too few samples to probe (need >=10 per class)")
        return {}
    
    results_dict = {}
    if baseline:
        print(f"\n  {'Layer':>5s}  {'Baseline':>8s}  {name:>8s}  {'Delta':>7s}")
    else:
        print(f"\n  {'Layer':>5s}  {'AUROC':>7s}")
    print("  " + "-" * 40)
    
    for layer in test_layers:
        X = np.array([r["hidden_states"][layer] for r in subset])
        y = np.array([1 if r["is_drifted"] else 0 for r in subset])
        auroc = best_auroc(X, y, device=device)
        results_dict[layer] = auroc
        
        if baseline:
            delta = auroc - baseline.get(layer, 0)
            print(f"  {layer:5d}  {baseline.get(layer, 0):8.4f}  {auroc:8.4f}  {delta:+7.4f}")
        else:
            print(f"  {layer:5d}  {auroc:7.4f}")
    
    return results_dict


# ============================================================
# ANALYSIS 0: SOURCE/FORMAT CONFOUND CHECK
# ============================================================

print("\n" + "=" * 70)
print("  ANALYSIS 0: SOURCE & FORMAT CONFOUND CHECK")
print("=" * 70)

print("\n  Drift rate by data source:")
for src in sorted(source_counts.keys()):
    src_results = [r for r in results if r["dataset_source"] == src]
    n_d = sum(1 for r in src_results if r["is_drifted"])
    n_s = len(src_results) - n_d
    pct = n_d / len(src_results) * 100 if src_results else 0
    has_yr = sum(1 for r in src_results if r["has_year_in_query"])
    print(f"    {src:25s}: {n_d:5d}/{len(src_results):5d} drifted ({pct:5.1f}%), {has_yr} with year in query")

print("\n  Drift rate by query format:")
for has_yr in [True, False]:
    fmt_results = [r for r in results if r["has_year_in_query"] == has_yr]
    n_d = sum(1 for r in fmt_results if r["is_drifted"])
    label = "with year" if has_yr else "no year"
    pct = n_d / len(fmt_results) * 100 if fmt_results else 0
    print(f"    {label:25s}: {n_d:5d}/{len(fmt_results):5d} drifted ({pct:5.1f}%)")

# ============================================================
# ANALYSIS 1: FULL DATASET (baseline)
# ============================================================

print("\n" + "=" * 70)
print("  ANALYSIS 1: FULL DATASET (baseline)")
print("=" * 70)

baseline = run_analysis("Full", results, test_layers, device)

# ============================================================
# ANALYSIS 2: YEAR-CONTROLLED (2024+ only)
# ============================================================

print("\n" + "=" * 70)
print("  ANALYSIS 2: YEAR-CONTROLLED (2024+ only)")
print("  Both drifted and stable share same year range")
print("=" * 70)

subset_2024 = [r for r in results if r["year"] >= 2024]

print("\n  Year breakdown in 2024+ subset:")
for yr in sorted(set(r["year"] for r in subset_2024)):
    d = sum(1 for r in subset_2024 if r["year"] == yr and r["is_drifted"])
    s = sum(1 for r in subset_2024 if r["year"] == yr and not r["is_drifted"])
    print(f"    {yr}: {d} drifted, {s} stable")

print("\n  Sources in 2024+ subset:")
for src in sorted(set(r["dataset_source"] for r in subset_2024)):
    n = sum(1 for r in subset_2024 if r["dataset_source"] == src)
    nd = sum(1 for r in subset_2024 if r["dataset_source"] == src and r["is_drifted"])
    print(f"    {src}: {n} total, {nd} drifted")

yc_results = run_analysis("2024+", subset_2024, test_layers, device, baseline)

# ============================================================
# ANALYSIS 3: SINGLE-YEAR (2025 only)
# ============================================================

print("\n" + "=" * 70)
print("  ANALYSIS 3: SINGLE-YEAR (2025 only)")
print("  Zero year variation — strictest year control")
print("=" * 70)

subset_2025 = [r for r in results if r["year"] == 2025]
sy_results = run_analysis("2025", subset_2025, test_layers, device, baseline)

# ============================================================
# ANALYSIS 4: FORMAT-CONTROLLED (only queries WITH year prefix)
# ============================================================

print("\n" + "=" * 70)
print("  ANALYSIS 4: FORMAT-CONTROLLED")
print("  Only queries with year in text (removes TempLAMA cloze format)")
print("=" * 70)

subset_withyear = [r for r in results if r["has_year_in_query"]]
print(f"\n  Sources in this subset:")
for src in sorted(set(r["dataset_source"] for r in subset_withyear)):
    n = sum(1 for r in subset_withyear if r["dataset_source"] == src)
    nd = sum(1 for r in subset_withyear if r["dataset_source"] == src and r["is_drifted"])
    print(f"    {src}: {n} total, {nd} drifted")

fc_results = run_analysis("WithYear", subset_withyear, test_layers, device, baseline)

# ============================================================
# ANALYSIS 5: COMBINED STRICTEST (2024+ AND format-matched)
# ============================================================

print("\n" + "=" * 70)
print("  ANALYSIS 5: STRICTEST CONTROL")
print("  2024+ only, queries with year, excludes known_drift edge cases")
print("=" * 70)

subset_strict = [r for r in results 
                 if r["year"] >= 2024 
                 and r["has_year_in_query"]
                 and r["category"] in ("unknown_drift", "no_drift", "stable")]
strict_results = run_analysis("Strict", subset_strict, test_layers, device, baseline)

# ============================================================
# ANALYSIS 6: PER-YEAR AUROC (Layer 7)
# ============================================================

print("\n" + "=" * 70)
print("  ANALYSIS 6: PER-YEAR DRIFT AUROC (Layer 7)")
print("=" * 70)

best_layer = 7
all_years = sorted(set(r["year"] for r in results if r["year"] > 0))
print(f"\n  {'Year':>6s}  {'N':>6s}  {'Drifted':>8s}  {'Stable':>8s}  {'AUROC':>7s}")
print("  " + "-" * 45)
for yr in all_years:
    yr_results = [r for r in results if r["year"] == yr]
    n_d = sum(1 for r in yr_results if r["is_drifted"])
    n_s = len(yr_results) - n_d
    if n_d >= 5 and n_s >= 5:
        X = np.array([r["hidden_states"][best_layer] for r in yr_results])
        y = np.array([1 if r["is_drifted"] else 0 for r in yr_results])
        auroc = best_auroc(X, y, device=device)
        print(f"  {yr:6d}  {len(yr_results):6d}  {n_d:8d}  {n_s:8d}  {auroc:7.4f}")
    elif n_d == 0:
        print(f"  {yr:6d}  {len(yr_results):6d}  {n_d:8d}  {n_s:8d}  {'--':>7s} (all stable)")
    else:
        print(f"  {yr:6d}  {len(yr_results):6d}  {n_d:8d}  {n_s:8d}  {'--':>7s}")

# ============================================================
# ANALYSIS 7: PER-RELATION on year-controlled subset (Layer 7)
# ============================================================

print("\n" + "=" * 70)
print("  ANALYSIS 7: PER-RELATION AUROC (2024+ only, Layer 7)")
print("=" * 70)

rels = Counter(r["relation"] for r in subset_2024)
print(f"\n  {'Relation':30s}  {'N':>5s}  {'Drifted':>7s}  {'Stable':>7s}  {'AUROC':>7s}")
print("  " + "-" * 65)
for rel in sorted(rels.keys()):
    rel_results = [r for r in subset_2024 if r["relation"] == rel]
    n_d = sum(1 for r in rel_results if r["is_drifted"])
    n_s = len(rel_results) - n_d
    if n_d >= 5 and n_s >= 5:
        X = np.array([r["hidden_states"][best_layer] for r in rel_results])
        y = np.array([1 if r["is_drifted"] else 0 for r in rel_results])
        auroc = best_auroc(X, y, device=device)
        print(f"  {rel:30s}  {len(rel_results):5d}  {n_d:7d}  {n_s:7d}  {auroc:7.4f}")
    else:
        print(f"  {rel:30s}  {len(rel_results):5d}  {n_d:7d}  {n_s:7d}  {'N/A':>7s}")

# ============================================================
# ANALYSIS 8: NAIVE YEAR BASELINE
# ============================================================

print("\n" + "=" * 70)
print("  ANALYSIS 8: NAIVE YEAR BASELINE")
print("  How well does 'year >= 2024' predict drift? (no probe needed)")
print("=" * 70)

y_true = np.array([1 if r["is_drifted"] else 0 for r in results])
y_year = np.array([1 if r["year"] >= 2024 else 0 for r in results])

tp = int(((y_true == 1) & (y_year == 1)).sum())
fp = int(((y_true == 0) & (y_year == 1)).sum())
tn = int(((y_true == 0) & (y_year == 0)).sum())
fn = int(((y_true == 1) & (y_year == 0)).sum())
acc = (tp + tn) / len(y_true)
prec = tp / (tp + fp) if (tp + fp) > 0 else 0
rec = tp / (tp + fn) if (tp + fn) > 0 else 0
year_auroc = numpy_auroc(y_true, y_year.astype(np.float64))

print(f"  Year >= 2024 as drift predictor:")
print(f"    AUROC:     {year_auroc:.4f}")
print(f"    Accuracy:  {acc:.4f}")
print(f"    Precision: {prec:.4f}")
print(f"    Recall:    {rec:.4f}")
print(f"    TP={tp}, FP={fp}, TN={tn}, FN={fn}")
print(f"\n  Compare: Probe AUROC at layer 7 = {baseline.get(7, 0):.4f}")
print(f"  Gap: {baseline.get(7, 0) - year_auroc:.4f} (probe advantage over year baseline)")

# ============================================================
# SUMMARY
# ============================================================

print("\n" + "=" * 70)
print("  SUMMARY TABLE")
print("=" * 70)

print(f"\n  {'Condition':30s}  " + "  ".join(f"L{l:2d}" for l in test_layers))
print("  " + "-" * (30 + 8 * len(test_layers)))

for name, res in [("Full dataset", baseline), 
                   ("2024+ only", yc_results),
                   ("2025 only", sy_results),
                   ("Format-controlled", fc_results),
                   ("Strictest control", strict_results)]:
    if res:
        vals = "  ".join(f"{res.get(l, 0):.3f}" for l in test_layers)
        print(f"  {name:30s}  {vals}")

print(f"\n  {'Year baseline AUROC':30s}  {year_auroc:.3f}")

print(f"""
  INTERPRETATION:
  - If 2024+ AUROC ~ Full AUROC      → year is NOT the main signal
  - If 2024+ AUROC << Full AUROC     → year WAS helping, check if still >0.7
  - If 2025-only AUROC still high    → probe works with zero year variation  
  - If year baseline ~ probe AUROC   → probe might just read years
  - If probe AUROC >> year baseline  → probe captures MORE than year info
  - If per-relation AUROCs are high  → signal is real within content domains
""")

# Save
summary = {
    "baseline": baseline,
    "year_controlled_2024": yc_results,
    "single_year_2025": sy_results,
    "format_controlled": fc_results,
    "strictest": strict_results,
    "year_baseline_auroc": float(year_auroc),
}
with open(os.path.join(OUTPUT_DIR, "year_controlled_results.json"), "w") as f:
    json.dump(summary, f, indent=2, default=str)

print(f"  Results saved to {OUTPUT_DIR}/year_controlled_results.json")
print("=" * 70)