#!/usr/bin/env python3
# =============================================================================
#  Project   : modelmap
#  File      : experiments/micro/expC_causal_verification/implementation/benchmark_v3.py
#  Purpose   : Run #3 — replication of the causal direction-erasure profile
#  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 #3 (hypothesis registered before this run).

Five direction sources (set A, set B, 3 bootstrap resamples of A); shared
random-direction controls; pairwise Spearman replication of the per-layer
specific-damage profile + the 2–15 vs 20–27 band claim per source.
"""

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 = 779
BOOTS = 3

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")]
NEW_NEAR = ["across the yard", "above the workbench", "below the landing", "outside the depot",
            "around the corner", "beyond the fence", "beneath the awning", "atop the cabinet"]


def spearman(a, b):
    ra = np.argsort(np.argsort(a)).astype(float)
    rb = np.argsort(np.argsort(b)).astype(float)
    ra -= ra.mean(); rb -= rb.mean()
    return float((ra * rb).sum() / np.sqrt((ra**2).sum() * (rb**2).sum()))


def directions_from(reps, labels, idx=None):
    """Per-layer unit diff-of-means directions (+grand means)."""
    if idx is not None:
        reps, labels = reps[idx], labels[idx]
    dirs, mus = [], []
    for layer in range(reps.shape[1]):
        x = reps[:, layer, :]
        mu = x.mean(0)
        u = x[labels == "correct"].mean(0) - x[labels == "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 NEW_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"baseline margin {base_m:+.4f} | pairs {len(pairs)}")

    # ---- direction sources
    sources: dict[str, tuple] = {}
    reps_cache = {}
    for setname in ("A", "B"):
        items = [json.loads(l) for l in
                 (ROOT / "benchmarks" / "promptsets" / f"agreement_{setname}.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_{setname}…", flush=True)
        reps = capture_pooled(model, taps, toks)["mean"]
        reps_cache[setname] = (reps, labels)
        sources[f"set{setname}"] = directions_from(reps, labels)
    repsA, labelsA = reps_cache["A"]
    for b in range(BOOTS):
        idx = np.random.default_rng(100 + b).integers(0, len(labelsA), len(labelsA))
        sources[f"bootA{b}"] = directions_from(repsA, labelsA, idx)

    # ---- shared random-direction control per layer
    rand_damage = []
    for layer in range(n_layers):
        ms = []
        for s in range(N_RANDOM_DIRS):
            ru = np.random.default_rng(1000 * layer + s).standard_normal(repsA.shape[-1])
            ru /= np.linalg.norm(ru)
            taps[layer].edit = erase_fn(ru.astype(np.float32), sources["setA"][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)

    # ---- per-source agreement-direction scans
    profiles = {}
    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
        band_early = float(np.mean(prof[2:16]))
        band_late = float(np.mean(prof[20:28]))
        print(f"{name:8s} early(2-15)={band_early:+.3f} late(20-27)={band_late:+.3f} "
              f"ratio={band_early / max(band_late, 1e-6):.1f}", flush=True)

    names = list(profiles)
    rhos = [spearman(np.array(profiles[a]), np.array(profiles[b]))
            for i, a in enumerate(names) for b in names[i + 1:]]
    mean_rho = float(np.mean(rhos))
    band_ok = all(np.mean(profiles[n][2:16]) >= 3 * max(np.mean(profiles[n][20:28]), 1e-6)
                  for n in names)
    survives = bool(mean_rho >= 0.7 and band_ok)
    print(f"replication: mean pairwise rho={mean_rho:.3f} (min {min(rhos):.3f}) "
          f"band-claim-all-sources={band_ok} -> publishable={survives}")

    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": 3,
        "scope": "replication of the direction-erasure causal profile (5 sources)",
        "commit": commit,
        "config": {"model": MODEL, "sources": names, "n_random_dirs": N_RANDOM_DIRS,
                   "n_pairs": len(pairs), "seed": SEED, "boots": BOOTS},
        "manifest": manifest(),
        "baseline_margin": base_m,
        "random_direction_damage": rand_damage,
        "profiles": profiles,
        "replication": {"pairwise_rhos": rhos, "mean_rho": mean_rho,
                        "min_rho": float(min(rhos)), "band_claim_all": band_ok,
                        "publishable": survives},
        "wall_seconds": round(time.time() - t0, 1),
    }, indent=2) + "\n")
    print(f"results -> {outdir / 'results.json'}")
    return 0


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