Partial-ReNoise Neural Architecture Mutation / results.json
Mechanism confirmed, baseline not beaten
1{
2 "seed": 2761,
3 "rows": [
4 {
5 "gamma": 0.0,
6 "alpha_bar": 1.0,
7 "mean_edit": 0.0,
8 "raw_validity": 1.0,
9 "post_repair_validity": 1.0,
10 "mean_accuracy": 77.1825237632243,
11 "p95_edit": 0.0
12 },
13 {
14 "gamma": 0.05,
15 "alpha_bar": 0.99,
16 "mean_edit": 0.01375,
17 "raw_validity": 0.985,
18 "post_repair_validity": 1.0,
19 "mean_accuracy": 77.20209463277646,
20 "p95_edit": 0.125
21 },
22 {
23 "gamma": 0.15,
24 "alpha_bar": 0.9258820055401665,
25 "mean_edit": 0.060625,
26 "raw_validity": 0.955,
27 "post_repair_validity": 1.0,
28 "mean_accuracy": 77.09487255802566,
29 "p95_edit": 0.25
30 },
31 {
32 "gamma": 0.3,
33 "alpha_bar": 0.7420133348094461,
34 "mean_edit": 0.200625,
35 "raw_validity": 0.85,
36 "post_repair_validity": 1.0,
37 "mean_accuracy": 77.48550950572152,
38 "p95_edit": 0.5
39 },
40 {
41 "gamma": 0.6,
42 "alpha_bar": 0.29991795168457636,
43 "mean_edit": 0.499375,
44 "raw_validity": 0.755,
45 "post_repair_validity": 1.0,
46 "mean_accuracy": 77.1911932121219,
47 "p95_edit": 0.75
48 },
49 {
50 "gamma": 1.0,
51 "alpha_bar": 0.030871261192240658,
52 "mean_edit": 0.701875,
53 "raw_validity": 0.725,
54 "post_repair_validity": 1.0,
55 "mean_accuracy": 77.26988210452124,
56 "p95_edit": 0.8812499999999979
57 }
58 ],
59 "independent": {
60 "mean_edit": 0.723125,
61 "validity": 1.0
62 },
63 "retention_check": [
64 {
65 "gamma": 0.05,
66 "predicted_retention": 0.99,
67 "observed_retention": 0.9899,
68 "absolute_error": 9.999999999998899e-05
69 },
70 {
71 "gamma": 0.15,
72 "predicted_retention": 0.9258820055401662,
73 "observed_retention": 0.9255,
74 "absolute_error": 0.00038200554016620814
75 },
76 {
77 "gamma": 0.3,
78 "predicted_retention": 0.7420133348094462,
79 "observed_retention": 0.7392,
80 "absolute_error": 0.002813334809446233
81 },
82 {
83 "gamma": 0.6,
84 "predicted_retention": 0.2999179516845764,
85 "observed_retention": 0.2977,
86 "absolute_error": 0.0022179516845763914
87 },
88 {
89 "gamma": 1.0,
90 "predicted_retention": 0.030871261192240654,
91 "observed_retention": 0.03045,
92 "absolute_error": 0.0004212611922406531
93 }
94 ],
95 "predictions": {
96 "P1": "categorical q retains each token with alpha_bar; observed retention should match alpha_bar within 0.01",
97 "P2": "edit distance should increase monotonically with gamma",
98 "P3": "gamma=1 mutation edit distance should approach independent-prior edit distance; gap expected small relative to 1 token fraction"
99 },
100 "convergence": {
101 "gamma1_vs_independent_absolute_edit_gap": 0.02124999999999999,
102 "gamma005_vs_independent_absolute_edit_gap": 0.709375
103 }
104}