File size: 8,300 Bytes
f930dac | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 | #!/usr/bin/env python3
"""Experiment 2 (PNS-Bind) evaluation and preregistered adjudication.
This is the study's sealed-confirmation evaluator with the single-use sentinel
removed: the confirmation split was read exactly once during the study, that
use is recorded in `results/sealed_use_record.json`, and the split is public
now, so the refusal machinery would only block reproduction.
Frozen conjunction (preregistration/PREREGISTRATION_E3_BIND.md), required for EVERY seed:
SEM_LATEST >= TXE + 0.05 AND SEM_2HOP >= TXE + 0.05
pooled mean >= TXE + 0.08
delta payload-swap >= 0.10
delta reset64 >= 0.10
Leakage control: payload-zero must collapse accuracy, i.e. the exact address
alone must not carry the answer.
python3 eval/confirmation_eval.py --split e3_conf
"""
import argparse
import json
import subprocess
import sys
from pathlib import Path
import numpy as np
import torch
ROOT = Path(__file__).resolve().parent.parent
sys.path.insert(0, str(ROOT / "src"))
sys.path.insert(0, str(Path(__file__).resolve().parent))
from pns.checkpoint import load_model # noqa: E402
from pns.common import atomic_write_json, eval_root, shards_root # noqa: E402
from pns.model.modules import enum_legal_mask # noqa: E402
from pns.train.loader import iter_eval_batches # noqa: E402
from pns.world.schema import Fam # noqa: E402
from stage_a import to_dev # noqa: E402
SEM = (int(Fam.SEM_LATEST), int(Fam.SEM_2HOP))
INTERVENTIONS = ("none", "payload_swap", "reset64", "payload_zero")
def cache_path(name, split, interv, limit):
suf = "" if limit is None else f"_n{limit}"
return eval_root() / f"E2_{name}_{split}_{interv}{suf}.json"
@torch.no_grad()
def run(name, dev, split, interv="none", limit=None, batch=24, use_cache=True):
"""Evaluate one (checkpoint, intervention) pair. Cached on disk so the
battery is resumable and can be sharded across devices."""
cp = cache_path(name, split, interv, limit)
if use_cache and cp.exists():
return json.loads(cp.read_text())
m, _, _ = load_model(name, dev)
d = m.cfg.d
rows = []
for b in iter_eval_batches(split, shards_root(), batch, limit):
g = to_dev(b, dev)
B, L = g["etype"].shape
st = m.initial_state(B, dev)
if interv == "payload_zero":
st = torch.zeros_like(st)
swap_at = L // 2
with torch.autocast("cuda", dtype=torch.bfloat16):
h = m.recenc.static(g["rec_val_toks"], g["rec_key_toks"], g["rec_store"],
g["rec_kind"], g["rec_key"], g["rec_ent"])
for t in range(L):
if interv == "payload_swap" and t == swap_at:
st = st.roll(1, dims=0)
if interv == "reset64" and t > 0 and t % 64 == 0:
st = m.initial_state(B, dev)
live = g["live"][:, t]
mask = live >= 0
rr = live.clamp(min=0)
bank = torch.gather(h, 1, rr.unsqueeze(-1).expand(-1, -1, d))
bank = m.recenc.finalize(
bank, (t - torch.gather(g["rec_birth"], 1, rr)).clamp(min=0))
st, o = m.step(st, g["tok"][:, t], g["etype"][:, t], g["dt"][:, t],
g["bind_write"][:, t], g["bind_read"][:, t],
g["bind_slot_ent"], g["bind_slot_attr"],
freeze_writes=(interv == "payload_zero"),
rec_bank=bank * mask.unsqueeze(-1), live_mask=mask)
sel = torch.isin(g["family"][:, t], torch.tensor(SEM, device=dev))
if sel.any():
lg = o["enum"][sel].masked_fill(
~enum_legal_mask(g["enum_legal"][sel, t]), -1e9)
ok = (lg.argmax(-1) == g["enum_gold"][sel, t]).cpu().numpy()
for j, i in enumerate(torch.nonzero(sel).flatten().tolist()):
rows.append((int(g["family"][i, t]), float(ok[j]),
int(g["delay"][i, t])))
a = np.array(rows)
out = {"n": len(a), "acc": round(float(a[:, 1].mean()), 4)}
for f, nm in ((int(Fam.SEM_LATEST), "latest"), (int(Fam.SEM_2HOP), "2hop")):
mm = a[:, 0] == f
if mm.sum() > 20:
out[f"acc_{nm}"] = round(float(a[mm, 1].mean()), 4)
out[f"n_{nm}"] = int(mm.sum())
for lo, nm in ((16, "gt16"), (128, "gt128")):
mm = a[:, 2] > lo
if mm.sum() > 20:
out[f"acc_{nm}"] = round(float(a[mm, 1].mean()), 4)
out[f"n_{nm}"] = int(mm.sum())
atomic_write_json(cp, out)
return out
def txe_comparator(split, batch):
"""Memory-free control on the SAME split, via the shared evaluator."""
tag = f"TXE_ON_E3_{split}_none"
rowfile = eval_root() / f"rows_{tag}.npz"
if not rowfile.exists():
subprocess.run([sys.executable, str(Path(__file__).with_name("evaluate.py")),
"--run", "TXE_ON_E3", "--split", split,
"--batch", str(batch)], check=True)
z = np.load(rowfile)
fam, ok = z["fam"], z["ok"]
return {"latest": round(float(ok[fam == int(Fam.SEM_LATEST)].mean()), 4),
"2hop": round(float(ok[fam == int(Fam.SEM_2HOP)].mean()), 4),
"pooled": round(float(ok[np.isin(fam, SEM)].mean()), 4)}
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--split", default="e3_conf",
help="e3_conf (the preregistered confirmation split) or e3_dev")
ap.add_argument("--stage", default="B", choices=["B"],
help="Stage B checkpoints (E3B_*)")
ap.add_argument("--limit", type=int, default=None)
ap.add_argument("--batch", type=int, default=24)
ap.add_argument("--txe-batch", type=int, default=256)
ap.add_argument("--out", default=None)
ap.add_argument("--only", default=None,
help="evaluate just this run (shard the battery), then exit")
ap.add_argument("--interv", default=None,
help="with --only: just this intervention")
args = ap.parse_args()
dev = "cuda"
if args.only:
for iv in ([args.interv] if args.interv else INTERVENTIONS):
r = run(args.only, dev, args.split, iv, args.limit, args.batch)
print(args.only, iv, json.dumps(r), flush=True)
return
rep = {"split": args.split, "TXE": txe_comparator(args.split, args.txe_batch)}
T = rep["TXE"]
for mode in ("bind", "unbound"):
for s in (1, 2, 3):
nm = f"E3B_{mode}_s{s}"
r = {iv: run(nm, dev, args.split, iv, args.limit, args.batch)
for iv in INTERVENTIONS}
r["delta_payload_swap"] = round(r["none"]["acc"] - r["payload_swap"]["acc"], 4)
r["delta_reset64"] = round(r["none"]["acc"] - r["reset64"]["acc"], 4)
r["delta_payload_zero"] = round(r["none"]["acc"] - r["payload_zero"]["acc"], 4)
rep[nm] = r
print(nm, "acc", r["none"]["acc"], "latest", r["none"].get("acc_latest"),
"2hop", r["none"].get("acc_2hop"), "dswap", r["delta_payload_swap"],
"dreset", r["delta_reset64"], flush=True)
v, seeds_ok = {}, []
for s in (1, 2, 3):
r = rep[f"E3B_bind_s{s}"]
c = {"latest": r["none"].get("acc_latest", 0) >= T["latest"] + 0.05,
"2hop": r["none"].get("acc_2hop", 0) >= T["2hop"] + 0.05,
"swap": r["delta_payload_swap"] >= 0.10,
"reset": r["delta_reset64"] >= 0.10}
c["all"] = all(c.values())
seeds_ok.append(c["all"])
v[f"seed{s}"] = c
pooled = float(np.mean([rep[f"E3B_bind_s{s}"]["none"]["acc"] for s in (1, 2, 3)]))
v["pooled_gain"] = round(pooled - T["pooled"], 4)
v["pooled_ok"] = bool(v["pooled_gain"] >= 0.08)
v["CONJUNCTION_MET"] = bool(all(seeds_ok) and v["pooled_ok"])
rep["verdict"] = v
out = Path(args.out) if args.out else eval_root() / f"CONFIRM_{args.split}.json"
atomic_write_json(out, rep)
print("\nCONJUNCTION MET:", v["CONJUNCTION_MET"], "| pooled gain", v["pooled_gain"])
for s in (1, 2, 3):
print(f" seed{s}:", json.dumps(v[f"seed{s}"]))
print("->", out)
if __name__ == "__main__":
main()
|