File size: 7,811 Bytes
47709ed | 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 | #!/usr/bin/env python3
"""Paired, lock-aware statistical analysis for the QH-024 replacement study."""
from __future__ import annotations
import hashlib
import json
import math
from datetime import datetime, timezone
from pathlib import Path
import numpy as np
ROOT = Path(__file__).resolve().parents[1]
ARTIFACTS = ROOT / "artifacts"
RUN_DIR = ARTIFACTS / "wikitext_qh024"
LOCK_PATH = ARTIFACTS / "wikitext-qh024-distillation-lock.json"
FILES = {
"base": RUN_DIR / "wikitext-base-s20260838-n512-e32.json",
"svd512": RUN_DIR / "wikitext-svd512-s20260838-n512-e32.json",
"qh024": RUN_DIR / "wikitext-qh024-s20260838-n512-e32.json",
"cc024": RUN_DIR / "wikitext-cc024-s20260838-n512-e32.json",
}
OUTPUT = RUN_DIR / "wikitext-qh024-paired-analysis.json"
def sha256(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as handle:
for chunk in iter(lambda: handle.read(8 << 20), b""):
digest.update(chunk)
return digest.hexdigest()
def load_rows(path: Path, key: str = "evaluation_active") -> tuple[dict, dict[int, float]]:
payload = json.loads(path.read_text(encoding="utf-8"))
rows = {
int(row["block_index"]): float(row["mean_token_nll"])
for row in payload[key]["rows"]
}
return payload, rows
def main() -> None:
lock = json.loads(LOCK_PATH.read_text(encoding="utf-8"))
samples = int(lock["evaluation_protocol"]["paired_bootstrap_samples"])
margin = float(lock["evaluation_protocol"]["compression_noninferiority_margin_nll"])
seed = int(lock["seed"])
payloads: dict[str, dict] = {}
rows: dict[str, dict[int, float]] = {}
for name, path in FILES.items():
payloads[name], rows[name] = load_rows(path)
expected = [int(value) for value in lock["eval_indices"]]
for name, values in rows.items():
if list(values) != expected:
raise AssertionError(f"{name} evaluation order does not match the lock")
if payloads[name]["dataset_lock_sha256"] != sha256(LOCK_PATH):
raise AssertionError(f"{name} lock hash mismatch")
if payloads[name]["dataset"]["test_rows_inspected_or_used_for_fitness"] != 0:
raise AssertionError(f"{name} reports test leakage")
qh_zero_payload, qh_zero = load_rows(FILES["qh024"], "evaluation_zero_residual")
cc_zero_payload, cc_zero = load_rows(FILES["cc024"], "evaluation_zero_residual")
del qh_zero_payload, cc_zero_payload
if qh_zero != rows["svd512"] or cc_zero != rows["svd512"]:
raise AssertionError("active-checkpoint zero branch is not identical to SVD-512")
arrays = {
name: np.asarray([values[index] for index in expected], dtype=np.float64)
for name, values in rows.items()
}
arrays["qh024_zero"] = np.asarray([qh_zero[index] for index in expected], dtype=np.float64)
rng = np.random.default_rng(seed)
bootstrap_indices = rng.integers(0, len(expected), size=(samples, len(expected)))
sign_rng = np.random.default_rng(seed + 1)
signs = sign_rng.choice(np.asarray([-1.0, 1.0]), size=(samples, len(expected)))
comparisons = {
"svd512_minus_base": ("svd512", "base"),
"qh024_minus_base": ("qh024", "base"),
"cc024_minus_base": ("cc024", "base"),
"qh024_minus_svd512": ("qh024", "svd512"),
"cc024_minus_svd512": ("cc024", "svd512"),
"qh024_minus_cc024": ("qh024", "cc024"),
"qh024_active_minus_own_zero": ("qh024", "qh024_zero"),
}
results: dict[str, dict] = {}
for label, (left, right) in comparisons.items():
differences = arrays[left] - arrays[right]
boot = differences[bootstrap_indices].mean(axis=1)
observed = float(differences.mean())
permutation = (differences[None, :] * signs).mean(axis=1)
results[label] = {
"definition": f"mean token NLL({left}) - mean token NLL({right}); lower is better",
"mean_delta_nll": observed,
"paired_percentile_bootstrap_95_ci": [
float(np.quantile(boot, 0.025)),
float(np.quantile(boot, 0.975)),
],
"paired_sign_flip_two_sided_p": float(
(np.count_nonzero(np.abs(permutation) >= abs(observed)) + 1) / (samples + 1)
),
"block_count": len(expected),
}
qh_base_upper = results["qh024_minus_base"]["paired_percentile_bootstrap_95_ci"][1]
svd_base_upper = results["svd512_minus_base"]["paired_percentile_bootstrap_95_ci"][1]
causal_upper = results["qh024_active_minus_own_zero"]["paired_percentile_bootstrap_95_ci"][1]
quantum_upper = results["qh024_minus_cc024"]["paired_percentile_bootstrap_95_ci"][1]
reduction = int(payloads["qh024"]["injection"]["net_model_parameter_reduction"])
gates = {
"svd_compression_noninferior_to_base": bool(svd_base_upper <= margin),
"qh024_compression_noninferior_to_base": bool(qh_base_upper <= margin and reduction > 0),
"qh024_causal_active_better_than_own_zero": bool(causal_upper < 0.0),
"qh024_quantum_better_than_equal_parameter_classical": bool(quantum_upper < 0.0),
"net_model_parameter_reduction_positive": bool(reduction > 0),
}
gates["quantum_specific_promotion"] = bool(
gates["qh024_compression_noninferior_to_base"]
and gates["qh024_causal_active_better_than_own_zero"]
and gates["qh024_quantum_better_than_equal_parameter_classical"]
)
if gates["quantum_specific_promotion"]:
verdict = "promote_quantum_specific_replacement"
elif gates["svd_compression_noninferior_to_base"]:
verdict = "retain_svd_compression_route_reject_qh024_quantum_residual"
else:
verdict = "reject_qh024_and_svd_compression_route"
output = {
"status": "ok",
"analysis": "QH-024 paired locked evaluation",
"verdict": verdict,
"metric": "mean token negative log likelihood; lower is better",
"lock": str(LOCK_PATH),
"lock_sha256": sha256(LOCK_PATH),
"input_sha256": {name: sha256(path) for name, path in FILES.items()},
"bootstrap_samples": samples,
"sign_flip_samples": samples,
"seed": seed,
"compression_noninferiority_margin_nll": margin,
"test_rows_inspected_or_used_for_fitness": 0,
"point_metrics": {
name: {
"mean_token_nll": float(values.mean()),
"token_perplexity": float(math.exp(values.mean())),
}
for name, values in arrays.items()
},
"comparisons": results,
"gates": gates,
"parameter_audit": {
"base_model_parameters": payloads["qh024"]["injection"]["base_model_parameters"],
"deployed_model_parameters": payloads["qh024"]["injection"]["deployed_model_parameters"],
"net_model_parameter_reduction": reduction,
"trainable_parameter_count_qh024": payloads["qh024"]["injection"]["trainable_parameter_count"],
"trainable_parameter_count_cc024": payloads["cc024"]["injection"]["trainable_parameter_count"],
},
"interpretation_limits": [
"The exact 10-qubit circuit is classically dequantizable at this scale.",
"A compression noninferiority pass is not evidence that the quantum residual caused the result.",
"The quantum-specific claim requires both active-vs-zero and QH-vs-equal-parameter-classical gates.",
"This formal evaluation uses 32 locked WikiText validation blocks and no test rows.",
],
"created_at": datetime.now(timezone.utc).isoformat(),
}
OUTPUT.write_text(json.dumps(output, indent=2), encoding="utf-8")
print(json.dumps(output, indent=2))
if __name__ == "__main__":
main()
|