#!/usr/bin/env python3
"""Generate the J-space REPRODUCTION dataset + article aggregates on Qwen2.5-0.5B.
Per reasoning probe: base NLL, J-space-ablated NLL, random-subspace-ablated NLL (control), workspace
effective-rank fraction, top-8 energy, per-layer future-influence norms, and logit-lens workspace concepts
by depth. All REAL measurements — feed both the bitacora post and the HF dataset. CPU/float32."""
import json, torch
import jspace_advanced as J

model = J.load()
L = model.config.num_hidden_layers; mid = L // 2
norm, head = model.model.norm, model.lm_head

PROBES = [
 "Q: If a train goes 60 km in 1 hour, how far in 3 hours? 60 times 3 equals",
 "Q: Ana has 5 apples, buys 7 more, gives 3 away. 5 plus 7 is 12, minus 3 is",
 "Q: A spider has 8 legs. How many legs do 3 spiders have? 8 times 3 is",
 "The opposite of hot is cold. The opposite of up is down. The opposite of fast is",
 "Paris is to France as Rome is to",
 "Q: There are 12 eggs in a box. Two boxes and 3 more eggs is 24 plus 3 =",
 "Q: 9 times 6 is",
 "If all cats are animals and Tom is a cat, then Tom is an",
 "Q: Half of 84 is",
 "Water freezes at 0 and boils at 100, so the range is 100 minus 0 =",
 "Q: A car travels 50 km per hour for 4 hours, covering 50 times 4 =",
 "Big is to small as tall is to",
]

def ablate_nll(ids, k=8, layer=mid, random=False):
    G = J.influence_jacobian(model, ids)
    st = J.subspace_stats(G[layer], k)
    Q = st["dirs"]
    if random:
        Q, _ = torch.linalg.qr(torch.randn_like(Q))
    return J.reasoning_nll(model, ids, intervene=(layer, Q)), st

rows = []
agg = {"base": [], "jspace": [], "random": [], "effrank": [], "energy": [], "influence": None}
for p in PROBES:
    ids = J.tok(p, return_tensors="pt").input_ids
    base = J.reasoning_nll(model, ids)
    js, st = ablate_nll(ids, random=False)
    rnd, _ = ablate_nll(ids, random=True)
    # per-layer future-influence norms
    G = J.influence_jacobian(model, ids)
    infl = [round(G[l].norm().item(), 4) for l in range(L)]
    # logit-lens workspace concepts at the final position, by depth
    with torch.no_grad():
        hs = model(ids, output_hidden_states=True).hidden_states
    concepts = {}
    for l in range(0, L + 1, max(1, L // 6)):
        top = head(norm(hs[l][0]))[-1].topk(4).indices.tolist()
        concepts[l] = [J.tok.decode([t]).strip() for t in top]
    row = {"prompt": p, "base_nll": round(base, 4), "jspace_ablated_nll": round(js, 4),
           "random_ablated_nll": round(rnd, 4), "effrank_frac": round(st["eff_rank_frac"], 5),
           "energy_top8": round(st["energy_topk"], 4), "per_layer_influence": infl,
           "workspace_concepts_by_depth": concepts}
    rows.append(row)
    agg["base"].append(base); agg["jspace"].append(js); agg["random"].append(rnd)
    agg["effrank"].append(st["eff_rank_frac"]); agg["energy"].append(st["energy_topk"])
    agg["influence"] = [a + b for a, b in zip(agg["influence"], infl)] if agg["influence"] else list(infl)
    print(f"  {p[:42]:42s} base={base:.2f} Jabl={js:.2f} Rabl={rnd:.2f} effrank={st['eff_rank_frac']:.3f}", flush=True)

n = len(PROBES)
mean = lambda x: sum(x) / len(x)
summary = {
 "model": "Qwen2.5-0.5B-Instruct", "layers": L, "hidden": model.config.hidden_size,
 "n_probes": n, "subspace_k": 8, "ablation_layer": mid,
 "mean_base_nll": round(mean(agg["base"]), 4),
 "mean_jspace_ablated_nll": round(mean(agg["jspace"]), 4),
 "mean_random_ablated_nll": round(mean(agg["random"]), 4),
 "jspace_damage": round(mean(agg["jspace"]) - mean(agg["base"]), 4),
 "random_damage": round(mean(agg["random"]) - mean(agg["base"]), 4),
 "causal_necessity_nats": round((mean(agg["jspace"]) - mean(agg["base"])) - (mean(agg["random"]) - mean(agg["base"])), 4),
 "mean_effrank_frac": round(mean(agg["effrank"]), 5),
 "mean_energy_top8": round(mean(agg["energy"]), 4),
 "mean_per_layer_influence": [round(x / n, 4) for x in agg["influence"]],
}
with open("/Users/kikocisneros/coco_ppl/jspace/jspace_measurements.jsonl", "w") as f:
    for r in rows: f.write(json.dumps(r) + "\n")
json.dump(summary, open("/Users/kikocisneros/coco_ppl/jspace/jspace_summary.json", "w"), indent=2)
print("\n=== SUMMARY ===")
print(json.dumps(summary, indent=2))
print("JSPACE_REPORT_DONE")