#!/usr/bin/env python3
# =============================================================================
#  Project   : modelmap
#  File      : experiments/micro/expC_causal_verification/implementation/benchmark_v4.py
#  Purpose   : Run #4 — the narrowed early-band claim, everything fresh
#  Author    : Simon-Pierre Boucher
#  Contact   : contact@spboucher.ai
#  Website   : https://modelmap.io
#  Created   : 2026-08-12
#  Modified  : 2026-08-12
#  Platform  : macOS / Apple Silicon (arm64) — MLX / Metal
#  License   : All rights reserved (research code)
# =============================================================================
"""expC run #4 (hypothesis registered before this run).

Band claim: early-band (layers 2-15) mean specific damage >= 2.5 in EVERY
of six fresh direction sources (disjoint halves of A and B + two fresh
bootstraps), on a fresh behavioral bank. Late band reported, no claim.
"""

from __future__ import annotations

import json
import random
import subprocess
import sys
import time
from pathlib import Path

import numpy as np

ROOT = Path(__file__).resolve().parents[4]
sys.path.insert(0, str(ROOT / "src"))
sys.path.insert(0, str(ROOT / "benchmarks"))
from hardware_manifest import manifest

from modelmap.capture.mlx_capture import capture_pooled, install_taps

MODEL = "mlx-community/Qwen3-0.6B-4bit"
N_RANDOM_DIRS = 3
SEED = 780
EARLY = slice(2, 16)
LATE = slice(20, 28)
BAR = 2.5

NOUN_PAIRS = [("key", "keys"), ("crate", "crates"), ("report", "reports"), ("valve", "valves"),
              ("ticket", "tickets"), ("ladder", "ladders"), ("sample", "samples"), ("cable", "cables"),
              ("permit", "permits"), ("beacon", "beacons"), ("filter", "filters"), ("stamp", "stamps")]
FRESH_NEAR = ["past the boiler room", "near the loading dock", "behind the ticket booth",
              "under the mezzanine", "beside the flagpole", "opposite the greenhouse",
              "inside the stairwell", "along the towpath"]


def directions_from(reps, labels, idx):
    dirs, mus = [], []
    r, l = reps[idx], labels[idx]
    for layer in range(reps.shape[1]):
        x = r[:, layer, :]
        mu = x.mean(0)
        u = x[l == "correct"].mean(0) - x[l == "violated"].mean(0)
        u = u / (np.linalg.norm(u) + 1e-8)
        dirs.append(u); mus.append(mu)
    return dirs, mus


def main() -> int:
    import mlx.core as mx
    from mlx_lm import load

    t0 = time.time()
    model, tokenizer = load(MODEL)
    taps = install_taps(model)
    n_layers = len(taps)

    rng = random.Random(SEED)
    combos = [(n, loc) for n in NOUN_PAIRS for loc in FRESH_NEAR]
    rng.shuffle(combos)
    pairs = []
    for (sg, pl), loc in combos:
        pairs.append({"prefix": f"The {sg} {loc}", "singular": True})
        pairs.append({"prefix": f"The {pl} {loc}", "singular": False})
    prefix_ids = [tokenizer.encode(p["prefix"]) for p in pairs]
    id_is, id_are = tokenizer.encode(" is")[0], tokenizer.encode(" are")[0]

    def margin() -> float:
        out = []
        for ids, p in zip(prefix_ids, pairs):
            logits = model(mx.array([ids]))[0, -1, :]
            mx.eval(logits)
            m = float(logits[id_is] - logits[id_are])
            out.append(m if p["singular"] else -m)
        return float(np.mean(out))

    def erase_fn(u_np, mu_np):
        u = mx.array(u_np.astype(np.float32))
        mu = mx.array(mu_np.astype(np.float32))
        def fn(out):
            h = out.astype(mx.float32)
            coef = ((h - mu) * u).sum(axis=-1, keepdims=True)
            return (h - coef * u).astype(out.dtype)
        return fn

    base_m = margin()
    print(f"fresh bank: {len(pairs)} pairs | baseline margin {base_m:+.4f}", flush=True)

    # fresh direction sources: disjoint halves + fresh bootstraps
    reps_cache = {}
    for s in ("A", "B"):
        items = [json.loads(l) for l in
                 (ROOT / "benchmarks" / "promptsets" / f"agreement_{s}.jsonl").read_text().splitlines()]
        toks = [tokenizer.encode(it["text"]) for it in items]
        labels = np.array([it["label"] for it in items])
        print(f"capture agreement_{s}…", flush=True)
        reps_cache[s] = (capture_pooled(model, taps, toks)["mean"], labels)

    sources = {}
    for s in ("A", "B"):
        reps, labels = reps_cache[s]
        n = len(labels)
        perm = np.random.default_rng(SEED + ord(s)).permutation(n)
        sources[f"{s}half1"] = directions_from(reps, labels, perm[: n // 2])
        sources[f"{s}half2"] = directions_from(reps, labels, perm[n // 2:])
        boot = np.random.default_rng(200 + ord(s)).integers(0, n, n)
        sources[f"boot{s}"] = directions_from(reps, labels, boot)
    del sources["bootB"]  # keep 6 by hypothesis? A/B halves (4) + bootA + bootB = 6 — keep both
    sources["bootB"] = directions_from(reps_cache["B"][0], reps_cache["B"][1],
                                       np.random.default_rng(202).integers(0, len(reps_cache["B"][1]),
                                                                           len(reps_cache["B"][1])))

    rand_damage = []
    d_model = reps_cache["A"][0].shape[-1]
    for layer in range(n_layers):
        ms = []
        for s in range(N_RANDOM_DIRS):
            ru = np.random.default_rng(2000 * layer + s).standard_normal(d_model)
            ru /= np.linalg.norm(ru)
            taps[layer].edit = erase_fn(ru.astype(np.float32), sources["Ahalf1"][1][layer])
            ms.append(margin())
            taps[layer].edit = None
        rand_damage.append(base_m - float(np.mean(ms)))
    print("random-direction controls done", flush=True)

    profiles, bands = {}, {}
    for name, (dirs, mus) in sources.items():
        prof = []
        for layer in range(n_layers):
            taps[layer].edit = erase_fn(dirs[layer], mus[layer])
            m = margin()
            taps[layer].edit = None
            prof.append((base_m - m) - rand_damage[layer])
        profiles[name] = prof
        bands[name] = {"early_mean": float(np.mean(prof[EARLY])),
                       "late_mean": float(np.mean(prof[LATE]))}
        print(f"{name:8s} early={bands[name]['early_mean']:+.3f} late={bands[name]['late_mean']:+.3f}",
              flush=True)

    early_means = [bands[n]["early_mean"] for n in sources]
    passes = bool(all(e >= BAR for e in early_means))
    print(f"BAND CLAIM (early mean >= {BAR} in every source): "
          f"min={min(early_means):+.3f} -> passes={passes}")

    arr = np.array(list(profiles.values()))
    commit = subprocess.run(["git", "rev-parse", "HEAD"], cwd=ROOT,
                            capture_output=True, text=True, check=False).stdout.strip()
    ts = time.strftime("%Y%m%dT%H%M%SZ", time.gmtime())
    outdir = ROOT / "results" / "expC_causal_verification" / ts
    outdir.mkdir(parents=True)
    (outdir / "results.json").write_text(json.dumps({
        "experiment": "expC_causal_verification", "run": 4,
        "scope": "narrowed early-band claim, fresh sources + fresh behavioral bank",
        "commit": commit,
        "config": {"model": MODEL, "sources": list(sources), "bar": BAR,
                   "early_layers": [2, 15], "late_layers": [20, 27],
                   "n_random_dirs": N_RANDOM_DIRS, "n_pairs": len(pairs), "seed": SEED},
        "manifest": manifest(),
        "baseline_margin": base_m,
        "random_direction_damage": rand_damage,
        "profiles": profiles,
        "bands": bands,
        "profile_mean": arr.mean(axis=0).tolist(),
        "profile_min": arr.min(axis=0).tolist(),
        "band_claim": {"bar": BAR, "early_means": early_means,
                       "min_early_mean": float(min(early_means)), "passes": passes},
        "wall_seconds": round(time.time() - t0, 1),
    }, indent=2) + "\n")
    print(f"results -> {outdir / 'results.json'}")
    return 0


if __name__ == "__main__":
    sys.exit(main())