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
|
#!/usr/bin/env python3
"""Audit the frozen five-seed Rain EP layer-state confirmation."""
from __future__ import annotations
import json
import math
from pathlib import Path
import statistics
from rain_ep_bias_c1 import CONDITIONS, RESULT_ROOT, SEEDS
ROOT = Path(__file__).resolve().parents[1]
T95 = 2.131846786326649
def read(path: Path) -> dict:
with path.open(encoding="utf-8") as handle:
return json.load(handle)
def path_for(seed: int, file_mode: str) -> Path:
return RESULT_ROOT / f"rain-ep-c1-s{seed}-{file_mode}.json"
def interval(values: list[float], absolute_mean: bool = False) -> dict:
mean = statistics.fmean(values)
sem = statistics.stdev(values) / math.sqrt(len(values))
center = abs(mean) if absolute_mean else mean
return {
"mean": mean,
"sem": sem,
"lower_95": mean - T95 * sem,
"upper_95": center + T95 * sem,
}
def main() -> None:
records = {}
missing = []
for seed in SEEDS:
for mode, file_mode in CONDITIONS:
path = path_for(seed, file_mode)
if not path.is_file():
missing.append(f"s{seed}:{mode}")
else:
records[(seed, mode)] = read(path)
if missing:
raise RuntimeError("missing C1 cells: " + ", ".join(missing))
rows = []
gain_raw = []
gain_constant = []
deficit_clean = []
difference_oracle = []
difference_noise = []
wall_ratios = []
for seed in SEEDS:
final = {
mode: records[(seed, mode)]["final"]
for mode, _ in CONDITIONS
}
accuracy = {
mode: final[mode]["test_accuracy"] for mode, _ in CONDITIONS
}
gain_raw.append(accuracy["innovation"] - accuracy["raw"])
gain_constant.append(
accuracy["innovation"] - accuracy["constant"])
deficit_clean.append(accuracy["clean"] - accuracy["innovation"])
difference_oracle.append(
accuracy["innovation"] - accuracy["oracle"])
difference_noise.append(
accuracy["same_rms_noise"] - accuracy["clean"])
wall_ratios.append(
final["innovation"]["wall_seconds"] / final["clean"]["wall_seconds"])
rows.append({
"seed": seed,
**{mode: value for mode, value in accuracy.items()},
"innovation_minus_raw": gain_raw[-1],
"innovation_minus_constant": gain_constant[-1],
"clean_minus_innovation": deficit_clean[-1],
"innovation_minus_oracle": difference_oracle[-1],
"noise_minus_clean": difference_noise[-1],
"innovation_wall_over_clean": wall_ratios[-1],
})
clean_mean = statistics.fmean(row["clean"] for row in rows)
required_finite = (
"clean", "same_rms_noise", "constant", "innovation", "oracle")
finite_complete = all(
len(records[(seed, mode)]["metrics"]) == 3
and all(metric["finite"] for metric in records[(seed, mode)]["metrics"])
for seed in SEEDS for mode in required_finite
)
innovation_residual_below_constant = all(
records[(seed, "innovation")]["final"]["corrector"]
["residual_to_clean_state_difference_rms"]
< records[(seed, "constant")]["final"]["corrector"]
["residual_to_clean_state_difference_rms"]
for seed in SEEDS
)
neutral_matched = all(
records[(seed, mode)]["final"]["corrector"]["neutral_observations"]
== 128
for seed in SEEDS for mode in ("constant", "innovation")
)
source_revisions = {
record["sdil"]["revision"] for record in records.values()
}
author_revisions = {
record["author"]["revision"] for record in records.values()
}
protocol_fixed = all(
record["protocol"]["evaluation_split"] == "train_holdout"
and record["protocol"]["data_seed"] == 6100
and record["protocol"]["bias_ratio"] == 0.01
and record["protocol"]["adapter"] == "layer"
and record["protocol"]["autodiff_used_for_learning"] is False
and record["protocol"]["extra_equilibrium_phases_for_predictor"] == 0
for record in records.values()
)
stats = {
"innovation_minus_raw": interval(gain_raw),
"innovation_minus_constant": interval(gain_constant),
"clean_minus_innovation": interval(deficit_clean),
"innovation_minus_oracle": interval(
difference_oracle, absolute_mean=True),
"noise_minus_clean": interval(difference_noise, absolute_mean=True),
"innovation_wall_over_clean": interval(wall_ratios),
}
checks = {
"finite_complete_required_conditions": finite_complete,
"mean_clean_validation_at_least_70": clean_mean >= 0.70,
"raw_damage_at_least_10_points_every_seed": all(
row["clean"] - row["raw"] >= 0.10 for row in rows),
"innovation_beats_raw_every_seed": all(value > 0 for value in gain_raw),
"innovation_raw_lower_95_at_least_10_points": (
stats["innovation_minus_raw"]["lower_95"] >= 0.10),
"innovation_beats_constant_every_seed": all(
value > 0 for value in gain_constant),
"innovation_constant_lower_95_at_least_5_points": (
stats["innovation_minus_constant"]["lower_95"] >= 0.05),
"clean_deficit_upper_95_below_3_points": (
stats["clean_minus_innovation"]["upper_95"] < 0.03),
"oracle_abs_mean_upper_95_below_3_points": (
stats["innovation_minus_oracle"]["upper_95"] < 0.03),
"noise_abs_mean_upper_95_below_3_points": (
stats["noise_minus_clean"]["upper_95"] < 0.03),
"innovation_residual_below_constant_every_seed": (
innovation_residual_below_constant),
"matched_128_neutral_observations": neutral_matched,
"mean_wall_overhead_at_most_15_percent": (
stats["innovation_wall_over_clean"]["mean"] <= 1.15),
"single_sdil_revision": len(source_revisions) == 1,
"single_author_revision": len(author_revisions) == 1,
"protocol_fixed_and_bp_free": protocol_fixed,
}
report = {
"stage": "rain_ep_bias_c1",
"gate": "pass" if all(checks.values()) else "fail",
"checks": checks,
"rows": rows,
"statistics": stats,
"mean_clean_validation_accuracy": clean_mean,
"num_records": len(records),
"sdil_revision": next(iter(source_revisions)),
"author_revision": next(iter(author_revisions)),
"test_policy": "test_set_never_loaded; fixed epoch-3 train holdout",
}
output = RESULT_ROOT.parent / "c1_gate.json"
output.write_text(
json.dumps(report, indent=2, sort_keys=True, allow_nan=False) + "\n")
print(json.dumps(report, indent=2, sort_keys=True, allow_nan=False))
if __name__ == "__main__":
main()
|
