#!/usr/bin/env python3
# =============================================================================
#  Project   : modelmap
#  File      : experiments/micro/expC_causal_verification/implementation/benchmark_v5.py
#  Purpose   : Run #5 — Level-3 path: dose-response steering of the direction
#  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 #5 (hypothesis registered before this run).

Activation-addition steering of the diff-of-means agreement direction at
three early-band layers, doses alpha in {-2,-1,0,+1,+2} x sigma_l; strict
monotonicity + halving criterion + random-direction specificity control.
"""

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"
TEST_LAYERS = (4, 8, 12)
DOSES = (-2.0, -1.0, 0.0, 1.0, 2.0)
N_RANDOM_DIRS = 3
SEED = 781

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

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

    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))

    # direction + per-layer projection std from agreement_A (the v1 object)
    items = [json.loads(l) for l in
             (ROOT / "benchmarks" / "promptsets" / "agreement_A.jsonl").read_text().splitlines()]
    toks = [tokenizer.encode(it["text"]) for it in items]
    labels = np.array([it["label"] for it in items])
    print("capture agreement_A (direction + sigma)…", flush=True)
    reps = capture_pooled(model, taps, toks)["mean"]
    dirs, sigmas = {}, {}
    for layer in TEST_LAYERS:
        x = reps[:, layer, :]
        u = x[labels == "correct"].mean(0) - x[labels == "violated"].mean(0)
        u = u / (np.linalg.norm(u) + 1e-8)
        dirs[layer] = u
        sigmas[layer] = float((x @ u).std())

    def add_fn(u_np, delta):
        d = mx.array((u_np * delta).astype(np.float32))
        def fn(out):
            return (out.astype(mx.float32) + d).astype(out.dtype)
        return fn

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

    results = {"per_layer": {}, "baseline_margin": base_m}
    mono_ok, halve_ok, spec_ok = True, True, True
    for layer in TEST_LAYERS:
        dose_margins = {}
        for a in DOSES:
            if a == 0.0:
                dose_margins[a] = base_m
                continue
            taps[layer].edit = add_fn(dirs[layer], a * sigmas[layer])
            dose_margins[a] = margin()
            taps[layer].edit = None
        seq = [dose_margins[a] for a in DOSES]
        mono = all(seq[i] < seq[i + 1] for i in range(len(seq) - 1))
        halve = dose_margins[-2.0] <= 0.5 * base_m
        rand_changes = []
        for s in range(N_RANDOM_DIRS):
            ru = np.random.default_rng(3000 * layer + s).standard_normal(reps.shape[-1])
            ru /= np.linalg.norm(ru)
            for a in (-2.0, 2.0):
                taps[layer].edit = add_fn(ru.astype(np.float32), a * sigmas[layer])
                rand_changes.append(abs(margin() - base_m))
                taps[layer].edit = None
        spec = float(np.mean(rand_changes)) < 0.25 * base_m
        mono_ok &= mono; halve_ok &= halve; spec_ok &= spec
        results["per_layer"][layer] = {
            "sigma": sigmas[layer],
            "dose_margins": {str(a): dose_margins[a] for a in DOSES},
            "monotone": mono, "halved_at_minus2": halve,
            "random_dir_mean_abs_change": float(np.mean(rand_changes)),
            "specific": spec,
        }
        print(f"L{layer:02d} doses " +
              " ".join(f"{a:+.0f}σ:{dose_margins[a]:+.2f}" for a in DOSES) +
              f" | mono={mono} halve={halve} randΔ={np.mean(rand_changes):.2f} spec={spec}",
              flush=True)

    passes = bool(mono_ok and halve_ok and spec_ok)
    print(f"LEVEL-3 CRITERION: monotone={mono_ok} halved={halve_ok} specific={spec_ok} "
          f"-> passes={passes}")

    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": 5,
        "scope": "dose-response steering (Level-3 path) at early-band layers",
        "commit": commit,
        "config": {"model": MODEL, "test_layers": list(TEST_LAYERS), "doses": list(DOSES),
                   "n_random_dirs": N_RANDOM_DIRS, "n_pairs": len(pairs), "seed": SEED},
        "manifest": manifest(),
        "results": results,
        "criterion": {"monotone_all": mono_ok, "halved_all": halve_ok,
                      "specific_all": spec_ok, "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())