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
|
#!/usr/bin/env python3
"""Audit and select the frozen S0 training-only stability margin."""
import argparse
import json
import math
import os
import statistics
MARGINS = (
("0p001", 0.001),
("0p003", 0.003),
("0p01", 0.01),
("0p03", 0.03),
)
SPLIT_HASH = "8328b206a97c420e49e54e3eca4abe3274c4756b084355784ea3fb8059e4515b"
def numeric_leaves(value):
if isinstance(value, bool) or value is None:
return
if isinstance(value, (int, float)):
yield float(value)
elif isinstance(value, dict):
for child in value.values():
yield from numeric_leaves(child)
elif isinstance(value, (list, tuple)):
for child in value:
yield from numeric_leaves(child)
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--input_dir", default="results/kp_innovation_stability_grid")
parser.add_argument(
"--out", default="results/kp_innovation_stability_gate.json")
args = parser.parse_args()
records = {}
source_commits = set()
for label, margin in MARGINS:
path = os.path.join(args.input_dir, f"margin_{label}.json")
with open(path) as handle:
record = json.load(handle)
records[label] = record
expected = {
"protocol": "kp_mixed_traffic_nonfinite_diagnosis_v1",
"scope": "training_only_no_validation_or_test_evaluation",
"rule": "innovation",
"predictor_mode": "closed_form",
"predictor_every": 0,
"max_steps": 352,
"stability_margin": margin,
"validation_evaluations": 0,
"test_evaluations": 0,
}
for key, value in expected.items():
if record.get(key) != value:
raise ValueError(f"S0 {label} {key} drift")
if record["provenance"]["git_tracked_dirty"]:
raise ValueError(f"tracked-dirty S0 {label}")
source_commits.add(record["provenance"]["git_commit"])
if record["split"]["validation_index_sha256"] != SPLIT_HASH:
raise ValueError(f"S0 {label} split drift")
if len(source_commits) != 1:
raise ValueError("S0 candidates must share one source revision")
metrics = {}
eligible = []
checks = {}
for label, margin in MARGINS:
record = records[label]
trajectory = record["trajectory"]
if len(trajectory) != 352:
raise ValueError(f"S0 {label} trajectory is incomplete")
all_finite = all(math.isfinite(value) for value in
numeric_leaves(record))
state_finite = all(
value["all_finite"]
for row in trajectory
for value in row["parameter_state"].values())
losses = [float(row["batch_loss"]) for row in trajectory]
ratios = [float(row["teaching_rms"])
/ max(float(row["instruction_rms"]), 1e-30)
for row in trajectory]
forward_weight_max = max(float(
row["parameter_state"]["forward_weight"][
"max_abs_over_finite_tensors"]) for row in trajectory)
feedback_weight_max = max(float(
row["parameter_state"]["feedback_weight"][
"max_abs_over_finite_tensors"]) for row in trajectory)
forward_momentum_max = max(float(
row["parameter_state"]["forward_momentum"][
"max_abs_over_finite_tensors"]) for row in trajectory)
feedback_momentum_max = max(float(
row["parameter_state"]["feedback_momentum"][
"max_abs_over_finite_tensors"]) for row in trajectory)
fit = record["predictor_warmup"]["closed_form_fit"]
ratio_errors = [abs(float(value) - 4.0) for value in
record["traffic_calibration"][
"realized_traffic_instruction_rms_ratio"]]
candidate_checks = {
"record_trajectory_and_state_finite": (
all_finite and state_finite
and record["first_any_nonfinite_step"] is None
and record["first_training_failure_step"] is None),
"maximum_batch_loss_at_most_10": max(losses) <= 10.0,
"final_32_mean_loss_at_most_2p5": (
statistics.mean(losses[-32:]) <= 2.5),
"maximum_used_instruction_rms_ratio_at_most_2": (
max(ratios) <= 2.0),
"forward_and_feedback_weight_max_at_most_10": (
forward_weight_max <= 10.0 and feedback_weight_max <= 10.0),
"forward_and_feedback_momentum_max_at_most_50": (
forward_momentum_max <= 50.0
and feedback_momentum_max <= 50.0),
"one_sided_closed_form_certificate": (
int(fit["observations"]) == 64
and float(fit["max_positive_residual_soma_slope"]) <= 1e-7),
"predictor_frozen_during_task": all(
row["predictor_updated"] is False for row in trajectory),
"traffic_ratio_calibrated": max(ratio_errors) <= 1e-5,
"task_loader_state_restored": (
record["predictor_warmup"][
"task_loader_state_restored"] is True),
"no_held_out_evaluations": (
record["validation_evaluations"] == 0
and record["test_evaluations"] == 0),
}
passed = all(candidate_checks.values())
if passed:
eligible.append((margin, label))
checks[label] = candidate_checks
metrics[label] = {
"margin": margin,
"eligible": passed,
"maximum_batch_loss": max(losses),
"final_32_mean_loss": statistics.mean(losses[-32:]),
"maximum_used_instruction_rms_ratio": max(ratios),
"maximum_forward_weight": forward_weight_max,
"maximum_feedback_weight": feedback_weight_max,
"maximum_forward_momentum": forward_momentum_max,
"maximum_feedback_momentum": feedback_momentum_max,
"max_positive_residual_soma_slope": float(
fit["max_positive_residual_soma_slope"]),
"min_residual_soma_slope": float(
fit["min_residual_soma_slope"]),
"maximum_initial_traffic_ratio_error": max(ratio_errors),
}
selected = min(eligible)[1] if eligible else None
status = "passed" if selected is not None else "failed_no_eligible_margin"
output = {
"protocol": "kp_stability_margin_training_prefix_v1",
"status": status,
"selected": selected,
"checks": checks,
"metrics": metrics,
"source_commit": next(iter(source_commits)),
"validation_evaluations": 0,
"test_evaluations": 0,
"review_score_before": 5,
"review_score_after": 5,
"score_change_rule": (
"training-only stability selection cannot change the paper score"),
}
os.makedirs(os.path.dirname(os.path.abspath(args.out)), exist_ok=True)
with open(args.out, "w") as handle:
json.dump(output, handle, indent=2, sort_keys=True)
handle.write("\n")
print(json.dumps(output, indent=2))
if __name__ == "__main__":
main()
|
