brandonmusic commited on
Commit
b492969
·
verified ·
1 Parent(s): 6dbc203

Add matched-window TR3 4bpw cache KLD and reference comparison

Browse files
README.md CHANGED
@@ -141,6 +141,30 @@ FP8. The prior text and linked receipts are retained below without silently
141
  changing either number. Do not use that disputed historical label as a matched
142
  cache comparison. New reference results are recorded separately above.
143
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
144
  ## Run the model
145
 
146
  Requires the custom runtime, four supported GPUs and NVIDIA Container Toolkit.
 
141
  changing either number. Do not use that disputed historical label as a matched
142
  cache comparison. New reference results are recorded separately above.
143
 
144
+ ## Matched-window comparison with TR3 4bpw
145
+
146
+ | Cache | TrellisMX reference KLD | TR3 / EXL3 4bpw KLD |
147
+ | --- | ---: | ---: |
148
+ | FP8 KV | 0.0319451732 | 0.0281899278 |
149
+ | NVFP4 MLA KV | 0.0354562238 | 0.0304785381 |
150
+
151
+ Lower is better on this panel. These runs use the same BF16 teacher, 32 previously
152
+ opened conditional-fit windows, exact token histories and 2,046 true-decode rows
153
+ per window. Both use MTP off, one sequence and CPU FP64 KL(teacher || student).
154
+ The TR3 and EXL3 Hub names identify the same uniform 4bpw checkpoint; this is not
155
+ the stock NVFP4-weight model. FP8 and NVFP4 in the table label the MLA KV cache.
156
+
157
+ This is a matched-data **system comparison**. TR3 uses its compatible r10
158
+ TP4/EP4/DCP4 runtime; TrellisMX uses r27 TP4/DCP4 with EP off. Expert math,
159
+ nonrouted-weight policy and cache layout also differ. The result does not isolate
160
+ the codec or establish general answer quality. One server per cache mode, FP8
161
+ then NVFP4; window intervals do not measure server-run variability.
162
+
163
+ [Per-window scores, paired intervals and audit](results/kld-tr3-20260909/README.md).
164
+ The first TR3 startup failed before any measured windows because kernel metadata
165
+ was unavailable in the shared compiler cache. The retained retry used process-
166
+ separated compiler caches on tmpfs; model weights and kernel math were unchanged.
167
+
168
  ## Run the model
169
 
170
  Requires the custom runtime, four supported GPUs and NVIDIA Container Toolkit.
results/kld-tr3-20260909/README.md ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Matched-data TR3 4bpw and TrellisMX reference KLD
2
+
3
+ | MLA KV cache | TrellisMX reference | TR3 / EXL3 4bpw | TR3 minus reference |
4
+ | --- | ---: | ---: | ---: |
5
+ | fp8 | 0.0319451732 | 0.0281899278 | -0.0037552454 |
6
+ | nvfp4 | 0.0354562238 | 0.0304785381 | -0.0049776857 |
7
+
8
+ Lower KL(teacher || student) is better on these measured inputs. The32windows,
9
+ teacher hashes, exact token histories,2046true-decode rows/window and FP64 scorer
10
+ match the previous reference measurement. Equal mean of32windowmeans; BCa95,
11
+ 20000resamples, seed20260902. Per-window values and paired intervals in comparison.json.
12
+ FP8 then NVFP4, one server preparation per arm; all prefix-hit counters zero.
13
+
14
+ TR3 means the owner's uniform4bpw GLM5.3 Flash checkpoint. Its TR3 and EXL3 Hub names
15
+ resolve to the same revision/manifest, not two different comparison models.
16
+ No stock NVFP4-weight model was substituted; FP8/NVFP4 here label KV caches.
17
+
18
+ This is a matched-data system comparison, not an isolated codec ablation:
19
+ TR3 uses its compatible Jovian r10 TP4/EP4/DCP4 runtime and source-native nonrouted
20
+ weights; TrellisMX uses r27 TP4/EPoff/DCP4 and its carrier. Cache block layout and
21
+ kernel implementations differ. Both use MTPoff,maxseq1,batch4096,maxlen1M,GMU.97,
22
+ 4x300W. TR3 loader requiresEP4 for this checkpoint atTP4; checkpoint unmodified.
23
+ Capture seam CPU tests preserve pre-mask logits and forced histories; immutable
24
+ image/source and fresh full checkpoint verification hashes accompany the results.
25
+
26
+ Already-opened conditional-fit development data, not untouched-final qualification
27
+ or independent replication. Window intervals do not quantify server-run variability,
28
+ and these short windows do not establish long-context or general answer quality.
29
+ Raw logits were retired after verified hashes and durable per-row scores/receipts.
30
+ Retained numerical scores, geometry, metadata, startup contract and zero-prefix-hit
31
+ receipts were audited; no independent recapture is claimed.
results/kld-tr3-20260909/audit.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "passed",
3
+ "arms": [
4
+ {
5
+ "arm": "fp8",
6
+ "windows": 32,
7
+ "prediction_rows": 65504,
8
+ "true_decode_rows": 65472,
9
+ "all_receipt_hashes_and_masks_pass": true,
10
+ "all_prefix_hit_counters_zero": true,
11
+ "raw_logits_retired_after_scoring": true,
12
+ "source_runtime_audit_sha256": "f0ecea1e32f5abd1a3ce6c016ddd904e26a0077e51666235f51a19326904e4c8",
13
+ "source_startup_sha256": "f3a2774cb50c34e38e73653c3ee820322148f40a4c99c811f28ed435a2eb4fde"
14
+ },
15
+ {
16
+ "arm": "nvfp4",
17
+ "windows": 32,
18
+ "prediction_rows": 65504,
19
+ "true_decode_rows": 65472,
20
+ "all_receipt_hashes_and_masks_pass": true,
21
+ "all_prefix_hit_counters_zero": true,
22
+ "raw_logits_retired_after_scoring": true,
23
+ "source_runtime_audit_sha256": "f0596d465f77d7bf0d39bf3ec55e251422b58aa741434c8dce2963fbdc94d8cd",
24
+ "source_startup_sha256": "cf960ddc018b9db95a713249f403e9b9571c1a596c1b5c46c606abc43e0c408e"
25
+ }
26
+ ],
27
+ "producing_analysis_sha256": "1427d9c5aadacf929ad0237792f99c275185e5fb74fde2574bfb229bca5b6b0a",
28
+ "verifier_sha256": "f602e38ce3a96423120ed32ad5c4ae500f1a02f319dd3344b29911f6b47bd7cc",
29
+ "claim": "Same-operator receipt/score audit; not independent replication."
30
+ }
results/kld-tr3-20260909/comparison.json ADDED
@@ -0,0 +1,1126 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": "glm53.tr3-r10-kv-cf32.v1",
3
+ "evidence": "controlled comparison on already-opened conditional-fit; one prepared run per KV mode, fixed order; not final qualification",
4
+ "arms": {
5
+ "fp8": {
6
+ "mean_true_decode_kld": 0.028189927801987147,
7
+ "window_bca95": [
8
+ 0.024202161460927674,
9
+ 0.03335995274457346
10
+ ],
11
+ "per_window": [
12
+ {
13
+ "window_id": "conditional-fit-0056",
14
+ "domain": "axis1_general",
15
+ "prediction_rows": 2047,
16
+ "true_decode_rows": 2046,
17
+ "true_decode_mean_kld": 0.017427964501784732,
18
+ "including_prefill_mean_kld": 0.017875558300170118,
19
+ "prefill_row_kld": 0.9336524697966719,
20
+ "raw_sha256": "b3fd9a2f4cdc037cf620ddf8c3618e9e672510e59deb3843efc785ee0f7340d8",
21
+ "score_sha256": "629c695ecc4a87d63fd1bfcf604c4abe59e2a2e45957f00e094d4b5470d21489",
22
+ "teacher_sha256": "55a70a330f39c5f23294c23d270a1ddc53031def99a42206e2706d4d0bb4466b",
23
+ "token_sha256": "219d8cd8873c6cf41f52782297cc5cfa9c4ed9666f2bc30d84cee9caa68ea744",
24
+ "raw_retirement": "after durable score and verified raw hash",
25
+ "raw_bytes": 1268157440
26
+ },
27
+ {
28
+ "window_id": "conditional-fit-0008",
29
+ "domain": "axis1_general",
30
+ "prediction_rows": 2047,
31
+ "true_decode_rows": 2046,
32
+ "true_decode_mean_kld": 0.051913750730159566,
33
+ "including_prefill_mean_kld": 0.05265334565792964,
34
+ "prefill_row_kld": 1.5658645678754968,
35
+ "raw_sha256": "5242ebf8026d2d37f5dd39bc39bcbb9c15c1403c8a93eb3a139bef8a39c32312",
36
+ "score_sha256": "d64bc1b949dfd4ee1f82701a4744cdc2c10697e8797548200fff05f40f042035",
37
+ "teacher_sha256": "040641247b2e060035d89c2e7345e029d7c142983585e7c5d63dc818486f650d",
38
+ "token_sha256": "581d72e3e78b15c02d9e73f6bc742260cb57cdcd74337f4419c8f90cbc7c9c11",
39
+ "raw_retirement": "after durable score and verified raw hash",
40
+ "raw_bytes": 1268157440
41
+ },
42
+ {
43
+ "window_id": "conditional-fit-0100",
44
+ "domain": "axis1_general",
45
+ "prediction_rows": 2047,
46
+ "true_decode_rows": 2046,
47
+ "true_decode_mean_kld": 0.039491981933763715,
48
+ "including_prefill_mean_kld": 0.04021472286668824,
49
+ "prefill_row_kld": 1.5189426716302628,
50
+ "raw_sha256": "a0369bd3021e54fa316bc58a417fbbdbc13e759bb63200228a07d3b428a7d3c1",
51
+ "score_sha256": "1900eb85047c84f1d0443c53d28948b89f190d4af5ef688d348fe360edd0c117",
52
+ "teacher_sha256": "f62991383b3ef5e8db0ed71442b17c428972e58da1c868c6aeb0f3fa143c865b",
53
+ "token_sha256": "7f008ec99068f373c6e15dd059ee977fb70e6dc832f46e009f0f299d48d19892",
54
+ "raw_retirement": "after durable score and verified raw hash",
55
+ "raw_bytes": 1268157440
56
+ },
57
+ {
58
+ "window_id": "conditional-fit-0094",
59
+ "domain": "axis1_general",
60
+ "prediction_rows": 2047,
61
+ "true_decode_rows": 2046,
62
+ "true_decode_mean_kld": 0.015535803605103407,
63
+ "including_prefill_mean_kld": 0.015712373665840943,
64
+ "prefill_row_kld": 0.37697471793484083,
65
+ "raw_sha256": "29b8c79257dc5c579b678bfb55a17abe8284dd6cad1b8f23af69ab6ab4ff58c0",
66
+ "score_sha256": "0a460dffb702b478bb580dd992a839f9d6d9b35ba9e0230cb5e1a44fd0057a31",
67
+ "teacher_sha256": "656911b8fb34d7f42e984dd21311caf9ef56db0b7618b414cad4089562d93cba",
68
+ "token_sha256": "6b6a9b21c49a9984cc8c5ea90e0bda6db5761d039c7cdd335dc65e30444e017b",
69
+ "raw_retirement": "after durable score and verified raw hash",
70
+ "raw_bytes": 1268157440
71
+ },
72
+ {
73
+ "window_id": "conditional-fit-0083",
74
+ "domain": "axis2_legal",
75
+ "prediction_rows": 2047,
76
+ "true_decode_rows": 2046,
77
+ "true_decode_mean_kld": 0.029388792659834843,
78
+ "including_prefill_mean_kld": 0.0362794436846717,
79
+ "prefill_row_kld": 14.134551440500875,
80
+ "raw_sha256": "6fe417498a67db9089a448ff7de2bece595c0aa050c3a9f0585ce0d2d0b6fa3e",
81
+ "score_sha256": "410e458afb3c884a5beea1c2b9db17e39f5b22b43ea06f7abb0cb33c875b5608",
82
+ "teacher_sha256": "66a7870a1f13c8b5d647c1681fca005bb0f55a6e502e94f282da673a6bed3f1b",
83
+ "token_sha256": "3df9d2872815e061ff8d32a9a19da0d11d80ac67070cccd28579b24ec3079552",
84
+ "raw_retirement": "after durable score and verified raw hash",
85
+ "raw_bytes": 1268157440
86
+ },
87
+ {
88
+ "window_id": "conditional-fit-0021",
89
+ "domain": "axis2_legal",
90
+ "prediction_rows": 2047,
91
+ "true_decode_rows": 2046,
92
+ "true_decode_mean_kld": 0.0284748160560963,
93
+ "including_prefill_mean_kld": 0.029186429441588693,
94
+ "prefill_row_kld": 1.4851474161590292,
95
+ "raw_sha256": "e5a1e26519c2a8076c70a975feab3cdcf78a89062a5d0384349d6b7309c12219",
96
+ "score_sha256": "7c98b790d0de2c415f8cbee4d0ba75933327020f9cc001c40f43e138cd2ed80e",
97
+ "teacher_sha256": "ebf76910c4db8f6955ad03f2ec7352e2f8aaf15e0c939d28f9a7d78f6bd7451b",
98
+ "token_sha256": "dbd9f2539cb3e97127ae760a04ca5b3456b0b13d6ce8698f88b215cc68b8f1a9",
99
+ "raw_retirement": "after durable score and verified raw hash",
100
+ "raw_bytes": 1268157440
101
+ },
102
+ {
103
+ "window_id": "conditional-fit-0098",
104
+ "domain": "axis2_legal",
105
+ "prediction_rows": 2047,
106
+ "true_decode_rows": 2046,
107
+ "true_decode_mean_kld": 0.0346264939607325,
108
+ "including_prefill_mean_kld": 0.036578765311046334,
109
+ "prefill_row_kld": 4.030925948053122,
110
+ "raw_sha256": "d5f7e952890fe0379e31fe759e93fe3860c8055f33e67cf446ae3e4359594b73",
111
+ "score_sha256": "86b6a9c5890a66f41269cd72cd28f781f56a4c60be73d1adbf4e87a0e17cb2b2",
112
+ "teacher_sha256": "16e3d9524100836c020535635f7722a4c93ce0fdddc2ad0781f25bda77e8bcc9",
113
+ "token_sha256": "749bda3291a0323d286b0018b7dc3e0f469de8afd5d194370cde357b7b315c7d",
114
+ "raw_retirement": "after durable score and verified raw hash",
115
+ "raw_bytes": 1268157440
116
+ },
117
+ {
118
+ "window_id": "conditional-fit-0080",
119
+ "domain": "axis2_legal",
120
+ "prediction_rows": 2047,
121
+ "true_decode_rows": 2046,
122
+ "true_decode_mean_kld": 0.02895811228055347,
123
+ "including_prefill_mean_kld": 0.03093290251338079,
124
+ "prefill_row_kld": 4.071353718878079,
125
+ "raw_sha256": "80c7f344db5c7832fd7fddb0ec9a06dc91c70cbcfb00e464923c700ed8ea2b93",
126
+ "score_sha256": "2d5a035aac780c72195b34cfe00040c0b17dd0e364ed04fce62a5b831e869445",
127
+ "teacher_sha256": "e9c1a77490a5db63d410297bd8febe06adf4df2a6f4e16552db187738c9289d2",
128
+ "token_sha256": "a31d43f38710eb305c684501a2b02909a4e82c43101c07e76e4a3ca4edb4d0dd",
129
+ "raw_retirement": "after durable score and verified raw hash",
130
+ "raw_bytes": 1268157440
131
+ },
132
+ {
133
+ "window_id": "conditional-fit-0081",
134
+ "domain": "axis3_code_agentic",
135
+ "prediction_rows": 2047,
136
+ "true_decode_rows": 2046,
137
+ "true_decode_mean_kld": 0.052266882698564825,
138
+ "including_prefill_mean_kld": 0.05347097920326377,
139
+ "prefill_row_kld": 2.5170524278172977,
140
+ "raw_sha256": "3c30ca32b24360b07f0a88aca0120b7cd773213bbca703a10e75b036550e422f",
141
+ "score_sha256": "ada8103098f904fe3c2fb6da66cc341be1e026d4a2a2fe970e702ed1a9c4b428",
142
+ "teacher_sha256": "5ae07760d91030d1eba925b76ea463309baa49c2e7d5f5e14409a23ced7fd20d",
143
+ "token_sha256": "350628488b94d59f3e4f3e99a418913544ea1fe7eacee14a642b996d9c614d16",
144
+ "raw_retirement": "after durable score and verified raw hash",
145
+ "raw_bytes": 1268157440
146
+ },
147
+ {
148
+ "window_id": "conditional-fit-0123",
149
+ "domain": "axis3_code_agentic",
150
+ "prediction_rows": 2047,
151
+ "true_decode_rows": 2046,
152
+ "true_decode_mean_kld": 0.03932676693569891,
153
+ "including_prefill_mean_kld": 0.040414307470727294,
154
+ "prefill_row_kld": 2.2655222421387964,
155
+ "raw_sha256": "0bcc6d599e857bc3b62f1ac24b104183e19a17ec34e3dcc902431cc69bc79cca",
156
+ "score_sha256": "f2224c4dac79d6edbe23e4f44fbd72c0822ad7e679181f255bae3c1e8ed75ad1",
157
+ "teacher_sha256": "4a24b5b750ba98c06d9bc69e19334aff89e7bd4f9429592598af5d62de7ecb42",
158
+ "token_sha256": "aab84ee426f91621d123aa68bd53f75e220fbaf8cb270db1feab1529a62ddbf0",
159
+ "raw_retirement": "after durable score and verified raw hash",
160
+ "raw_bytes": 1268157440
161
+ },
162
+ {
163
+ "window_id": "conditional-fit-0090",
164
+ "domain": "axis3_code_agentic",
165
+ "prediction_rows": 2047,
166
+ "true_decode_rows": 2046,
167
+ "true_decode_mean_kld": 0.034422243044485955,
168
+ "including_prefill_mean_kld": 0.03574033212388724,
169
+ "prefill_row_kld": 2.7325505885789205,
170
+ "raw_sha256": "3ac4ed68f867c402e552b070ac57442bd522e27cfacdabf3bf1fdbd0c96d798c",
171
+ "score_sha256": "e960b4edefc921bd1bb8a0b45e11ce897025db2c54770f8b78a03b304612d386",
172
+ "teacher_sha256": "db2a5fe33c9e457a22ebeb31a7073c1dac666b877503e9b4e90a8d6abf7b8a32",
173
+ "token_sha256": "a06e76993db717afee847fcc4d86f10b52dd23cdce5dae916f32d238d08d886c",
174
+ "raw_retirement": "after durable score and verified raw hash",
175
+ "raw_bytes": 1268157440
176
+ },
177
+ {
178
+ "window_id": "conditional-fit-0062",
179
+ "domain": "axis3_code_agentic",
180
+ "prediction_rows": 2047,
181
+ "true_decode_rows": 2046,
182
+ "true_decode_mean_kld": 0.020455155344797665,
183
+ "including_prefill_mean_kld": 0.020865469467274723,
184
+ "prefill_row_kld": 0.860368164055328,
185
+ "raw_sha256": "824f8c491ef4c9801831c10051477fa763c510d9886cdec790504dd422b36a38",
186
+ "score_sha256": "37c80b3e1f079d7e8eca4392ad90b008c42641e4348aa0c6ff350a4740d088dc",
187
+ "teacher_sha256": "d6f7883a533625c7c75119fe7403b48bf81885b1de0b311a1a0571949f7f231c",
188
+ "token_sha256": "7c6d2ef2f4b7c7091a517f6ccd778009f9ee144165c9aecf2f8f28aea7b3753f",
189
+ "raw_retirement": "after durable score and verified raw hash",
190
+ "raw_bytes": 1268157440
191
+ },
192
+ {
193
+ "window_id": "conditional-fit-0047",
194
+ "domain": "axis4_reasoning_termination",
195
+ "prediction_rows": 2047,
196
+ "true_decode_rows": 2046,
197
+ "true_decode_mean_kld": 0.027885252605622038,
198
+ "including_prefill_mean_kld": 0.03487315957624962,
199
+ "prefill_row_kld": 14.332130821480268,
200
+ "raw_sha256": "f464eb0b3d6a51bf1395d53cfdc02238c2737d50c8e277a97b80cedc81ddb2cc",
201
+ "score_sha256": "e7a823fdcbbdf35a6528132ac2c8c053cd326326e978b22da70e35eb0018cf7d",
202
+ "teacher_sha256": "8513cdcb9940786c7c42bf14d2f5f2cd452791e4ec6036c06a1f973eb969018d",
203
+ "token_sha256": "89730ee6826d4302c352156d5df766c51622a8b6681b8052d4ca6e7d481177fe",
204
+ "raw_retirement": "after durable score and verified raw hash",
205
+ "raw_bytes": 1268157440
206
+ },
207
+ {
208
+ "window_id": "conditional-fit-0051",
209
+ "domain": "axis4_reasoning_termination",
210
+ "prediction_rows": 2047,
211
+ "true_decode_rows": 2046,
212
+ "true_decode_mean_kld": 0.017317721864985175,
213
+ "including_prefill_mean_kld": 0.024214269846732066,
214
+ "prefill_row_kld": 14.134551440500875,
215
+ "raw_sha256": "78478b098be41de36580b22123c19413b17dce2a4aed154d0c1b79cc26a9263d",
216
+ "score_sha256": "d954cb12e752711f8cb0e9dd3424a83f843de27fb4d84a99d54b053da6e4dbc1",
217
+ "teacher_sha256": "177477c19b54dc0b0230e19202d48ab71d52f2b1a37fe383511bb64f0aae2fcf",
218
+ "token_sha256": "c79269b16560e7301549593443c9394007e7a957ad149c46224e7ad271dc6a18",
219
+ "raw_retirement": "after durable score and verified raw hash",
220
+ "raw_bytes": 1268157440
221
+ },
222
+ {
223
+ "window_id": "conditional-fit-0055",
224
+ "domain": "axis4_reasoning_termination",
225
+ "prediction_rows": 2047,
226
+ "true_decode_rows": 2046,
227
+ "true_decode_mean_kld": 0.017814014293265207,
228
+ "including_prefill_mean_kld": 0.01839697269156558,
229
+ "prefill_row_kld": 1.211129855614137,
230
+ "raw_sha256": "fff58b329c6a1a3cf889a1edc7d8448c890ed0fe7d2f73c0ea8a13b975e78093",
231
+ "score_sha256": "8c3dfdfcfeb29b57a8013b6cea2597593ce06f5e3356f0673917a8728fda26cd",
232
+ "teacher_sha256": "7679c828eae4bf08f17598d044904693fc90c05b9d3f189dddeeb846100e43a7",
233
+ "token_sha256": "feec9ac508172dd275ec6452a03493acab5ffdff9bb57161317c77f24daf90f2",
234
+ "raw_retirement": "after durable score and verified raw hash",
235
+ "raw_bytes": 1268157440
236
+ },
237
+ {
238
+ "window_id": "conditional-fit-0031",
239
+ "domain": "axis4_reasoning_termination",
240
+ "prediction_rows": 2047,
241
+ "true_decode_rows": 2046,
242
+ "true_decode_mean_kld": 0.019452131991742394,
243
+ "including_prefill_mean_kld": 0.026444158708639576,
244
+ "prefill_row_kld": 14.332130821480268,
245
+ "raw_sha256": "08a1aaf540ebe77228ecb2ecde7ce637c92884266ef5bf4e212908592a90e857",
246
+ "score_sha256": "0cb570aa8af25ddd54dacb8d2f253ba63e06b2f3e86e500cb0e333f6aa9cb121",
247
+ "teacher_sha256": "e04a6c04a951efb5a178927e91457535cc67745ce5c81e785cb6d963aac6ce83",
248
+ "token_sha256": "2cff3482b4212537d0c35b88b8a0f479502c8905952430c973d25b1ab0bb5818",
249
+ "raw_retirement": "after durable score and verified raw hash",
250
+ "raw_bytes": 1268157440
251
+ },
252
+ {
253
+ "window_id": "conditional-fit-0003",
254
+ "domain": "axis4_reasoning_termination",
255
+ "prediction_rows": 2047,
256
+ "true_decode_rows": 2046,
257
+ "true_decode_mean_kld": 0.02716569936922329,
258
+ "including_prefill_mean_kld": 0.029435662037889526,
259
+ "prefill_row_kld": 4.67377928212901,
260
+ "raw_sha256": "6643b88ddd06bec1ebbb5602b2bedf4a4158c5828774b8ef715bca74dd133a10",
261
+ "score_sha256": "79e9920a561a1fc211209654fa6d71e703180842fe511793076c38ae3a3edc92",
262
+ "teacher_sha256": "7677c89cc3dee8e2a64e3c7947add39475f828921a3a8fe2e92b379b3ed8320b",
263
+ "token_sha256": "d416537c6e8273747bf025e97287a0bb6e3ff047bfee65cf54e4dfe287569b93",
264
+ "raw_retirement": "after durable score and verified raw hash",
265
+ "raw_bytes": 1268157440
266
+ },
267
+ {
268
+ "window_id": "conditional-fit-0009",
269
+ "domain": "axis2_legal",
270
+ "prediction_rows": 2047,
271
+ "true_decode_rows": 2046,
272
+ "true_decode_mean_kld": 0.02991776464625218,
273
+ "including_prefill_mean_kld": 0.029992131082785946,
274
+ "prefill_row_kld": 0.18214586023087223,
275
+ "raw_sha256": "19098c2e85ca77cccb297ffedb2631f0f8cd34674001a036c941fcb9699f9ef1",
276
+ "score_sha256": "1b80d3d4ae89247649f7d5aef8fc3c8c1dd5336ed98fbde5523331d66554a7c4",
277
+ "teacher_sha256": "00701ad8bf4a4a5eedcedd4eafd956830c9ff1c790b15074ca721a3462abf930",
278
+ "token_sha256": "fc1166707f5938a914faf364dab717b0de3ea0e8451ed1a3fb399469e659b105",
279
+ "raw_retirement": "after durable score and verified raw hash",
280
+ "raw_bytes": 1268157440
281
+ },
282
+ {
283
+ "window_id": "conditional-fit-0010",
284
+ "domain": "axis3_code_agentic",
285
+ "prediction_rows": 2047,
286
+ "true_decode_rows": 2046,
287
+ "true_decode_mean_kld": 0.016131585306488436,
288
+ "including_prefill_mean_kld": 0.016976009517953087,
289
+ "prefill_row_kld": 1.744667946174626,
290
+ "raw_sha256": "1f5d1aef5bb9a5d6319c859c5e36a93af0db774af05a01fd5419107b63c6f116",
291
+ "score_sha256": "21ef43a97e3f94010195645753fb3ce605c09e5368a72009dee91def0dd3c390",
292
+ "teacher_sha256": "f9bf4519890b70d972dfdf9760181418eb7c55b6237c0108f01d44df89943278",
293
+ "token_sha256": "ddd9d769a76eba74bc8e680ff25736a3b9146371801d1d2b29e4dc48f6d256a2",
294
+ "raw_retirement": "after durable score and verified raw hash",
295
+ "raw_bytes": 1268157440
296
+ },
297
+ {
298
+ "window_id": "conditional-fit-0011",
299
+ "domain": "axis4_reasoning_termination",
300
+ "prediction_rows": 2047,
301
+ "true_decode_rows": 2046,
302
+ "true_decode_mean_kld": 0.020680451065987218,
303
+ "including_prefill_mean_kld": 0.027671877724714272,
304
+ "prefill_row_kld": 14.332130821480268,
305
+ "raw_sha256": "2effaea31efc163f3e08c9dad434a9e92de6b022be1bab59f86e314b002f08c6",
306
+ "score_sha256": "c3b22b58a1d88ff064ecf1a84094251263f8d22ba3c27a7bbb2b03b5a91e5cb4",
307
+ "teacher_sha256": "9f37c4ead38bb80a5d85f825b156e5b493b6d3f1ca9dabfad06b0416f37db286",
308
+ "token_sha256": "c4b96acfa47944be3485b20eab76eb16d59913c9fedb5001a537335a3ee13614",
309
+ "raw_retirement": "after durable score and verified raw hash",
310
+ "raw_bytes": 1268157440
311
+ },
312
+ {
313
+ "window_id": "conditional-fit-0030",
314
+ "domain": "axis3_code_agentic",
315
+ "prediction_rows": 2047,
316
+ "true_decode_rows": 2046,
317
+ "true_decode_mean_kld": 0.03895077932312888,
318
+ "including_prefill_mean_kld": 0.04040829292766928,
319
+ "prefill_row_kld": 3.0224811278173216,
320
+ "raw_sha256": "7d5ba89d1d3ff0279f2607d4df7c2b3aff458c80d6110c94c16b3f347a91ecda",
321
+ "score_sha256": "565e800158334826fa5b60d429082d16030da4e92596466756442510492fd852",
322
+ "teacher_sha256": "90c08d120cfac926855e15c106dba1e242099bcd0ae6da7655bfaa54d5ce2420",
323
+ "token_sha256": "d644a79da2ba01cefc5f2020eab9e8662d45ba3025377844e4bf4d3f1d36802d",
324
+ "raw_retirement": "after durable score and verified raw hash",
325
+ "raw_bytes": 1268157440
326
+ },
327
+ {
328
+ "window_id": "conditional-fit-0032",
329
+ "domain": "axis1_general",
330
+ "prediction_rows": 2047,
331
+ "true_decode_rows": 2046,
332
+ "true_decode_mean_kld": 0.01767228690329637,
333
+ "including_prefill_mean_kld": 0.018596997273520655,
334
+ "prefill_row_kld": 1.9105544147523985,
335
+ "raw_sha256": "0ea92010e573b624dd75419e9973c119ab716e7ce71d0bca25f4aa1a77715528",
336
+ "score_sha256": "e29bcfc3f80b04ce3dbe703ee061c43e61d2e41368ecb5d9631321561a35764f",
337
+ "teacher_sha256": "1aa70ccf8e224c3a58834e2232a3a2477768db29ffeca39f5b27c17ad8c40c2b",
338
+ "token_sha256": "88745c333ead140a6fb05929031613429cbd42e25881d92add968a94a5d2d930",
339
+ "raw_retirement": "after durable score and verified raw hash",
340
+ "raw_bytes": 1268157440
341
+ },
342
+ {
343
+ "window_id": "conditional-fit-0035",
344
+ "domain": "axis4_reasoning_termination",
345
+ "prediction_rows": 2047,
346
+ "true_decode_rows": 2046,
347
+ "true_decode_mean_kld": 0.017841740154740375,
348
+ "including_prefill_mean_kld": 0.020141260127286724,
349
+ "prefill_row_kld": 4.724959123957103,
350
+ "raw_sha256": "5a7c0ba7a40f804c93d3fbe21bcb4454c9502fb75950c0058ba5625423d236c8",
351
+ "score_sha256": "f6659f2a052304cbaaf0cedb1a896448e8cb320547c713d23e1335d902e5f5a2",
352
+ "teacher_sha256": "4e958dce74e093eb62ba5ac210fe0a4e50f12f1875de678769700cde206b429d",
353
+ "token_sha256": "49f702136da39db9e762b6815a1952c0eba0a1fb04af22fb28a27bb63e2a41dc",
354
+ "raw_retirement": "after durable score and verified raw hash",
355
+ "raw_bytes": 1268157440
356
+ },
357
+ {
358
+ "window_id": "conditional-fit-0040",
359
+ "domain": "axis1_general",
360
+ "prediction_rows": 2047,
361
+ "true_decode_rows": 2046,
362
+ "true_decode_mean_kld": 0.021052493606303763,
363
+ "including_prefill_mean_kld": 0.021580772401606545,
364
+ "prefill_row_kld": 1.1024391875910946,
365
+ "raw_sha256": "fc7a5373ee49eb1bb5a3b7c44a32ce2cfd74570bf93ac6006609d9c7be5fa36d",
366
+ "score_sha256": "1975bbd320349e00e79ddf2384a4037c0554cb25d7c4f6aa945e6f3e1ee28d67",
367
+ "teacher_sha256": "f4c96798fc9eb5564d9aec52779cc0ce88bc55312dcd7663240938d711c8a90d",
368
+ "token_sha256": "50f3bc08ab37c551a05364eb48f0d726742f19331e79114dd5a9fb3c42ce6d4f",
369
+ "raw_retirement": "after durable score and verified raw hash",
370
+ "raw_bytes": 1268157440
371
+ },
372
+ {
373
+ "window_id": "conditional-fit-0041",
374
+ "domain": "axis2_legal",
375
+ "prediction_rows": 2047,
376
+ "true_decode_rows": 2046,
377
+ "true_decode_mean_kld": 0.05927521059021604,
378
+ "including_prefill_mean_kld": 0.059798391996830465,
379
+ "prefill_row_kld": 1.130227549929964,
380
+ "raw_sha256": "84231f0db55357de11d2fbb626e31d1ba1ad11b723532291a5bdb8eb6454fee8",
381
+ "score_sha256": "1dd61ce4c3c8167ec5ac0465d51e34ca6717a13692822b7b7fdf32ee1cfcfa03",
382
+ "teacher_sha256": "c2bbe4ea31532baceff966423c6ce6e042aebb5a2eb1de45e382a7d3df6f6827",
383
+ "token_sha256": "8acf2614f0f787c185af112f21bcdf247c9b6ae76d292ada28f476e2a7b0241d",
384
+ "raw_retirement": "after durable score and verified raw hash",
385
+ "raw_bytes": 1268157440
386
+ },
387
+ {
388
+ "window_id": "conditional-fit-0046",
389
+ "domain": "axis3_code_agentic",
390
+ "prediction_rows": 2047,
391
+ "true_decode_rows": 2046,
392
+ "true_decode_mean_kld": 0.05896017735987805,
393
+ "including_prefill_mean_kld": 0.05920087226615672,
394
+ "prefill_row_kld": 0.5516626505123096,
395
+ "raw_sha256": "2ddc9b5cdbe487273c6777c34ca8fe26f7716c77bbacff4ad573fc1a814f3d48",
396
+ "score_sha256": "a50c693d5782c36572153176604e283f019c36df15031274c7807a72cf9c01b3",
397
+ "teacher_sha256": "3caa5c58fb0b0e80129d551b39867fd52fefd93830060bb5655862fdd0e3b8ce",
398
+ "token_sha256": "3f2df1c090abf4473163a078c168fec33ab98ed6415985b92449caa861c04444",
399
+ "raw_retirement": "after durable score and verified raw hash",
400
+ "raw_bytes": 1268157440
401
+ },
402
+ {
403
+ "window_id": "conditional-fit-0060",
404
+ "domain": "axis1_general",
405
+ "prediction_rows": 2047,
406
+ "true_decode_rows": 2046,
407
+ "true_decode_mean_kld": 0.017536468901318984,
408
+ "including_prefill_mean_kld": 0.01880800919234263,
409
+ "prefill_row_kld": 2.6203794446267246,
410
+ "raw_sha256": "d3b10f3205de708bbf3846a1d7181f776f147cc9726491923b2dcb26613f2c24",
411
+ "score_sha256": "6ab3fb95b1f219ac7724586806f3380d8263d2ff29ddafade2a84392f2b85c99",
412
+ "teacher_sha256": "d9e6344664b92527ee58d3bca038702676b357e7052ade03ac4bc9c70908cc34",
413
+ "token_sha256": "04ffe38aafcd9407b2846d59e8d797bee04b97f00683f8bc74d475a2d6b8592b",
414
+ "raw_retirement": "after durable score and verified raw hash",
415
+ "raw_bytes": 1268157440
416
+ },
417
+ {
418
+ "window_id": "conditional-fit-0063",
419
+ "domain": "axis4_reasoning_termination",
420
+ "prediction_rows": 2047,
421
+ "true_decode_rows": 2046,
422
+ "true_decode_mean_kld": 0.022349021824846185,
423
+ "including_prefill_mean_kld": 0.022571380045245257,
424
+ "prefill_row_kld": 0.47751629898174547,
425
+ "raw_sha256": "6e485d96484b8bce12cd5be5f251986a24087272f3d448c74ac6547f356bea37",
426
+ "score_sha256": "a1b56be7afcd4c43d829a1e8ea30271035d48e915c04ba74e337c2fd12336dd5",
427
+ "teacher_sha256": "a9bb59659437c1df5b3f3ef0b54eda61402842a2597b76b0124f58b59ecf3ab4",
428
+ "token_sha256": "38003e2d26db850dd8ff39f2e95fa3d1be9a34838c70d9f7e87056b7ef50b67d",
429
+ "raw_retirement": "after durable score and verified raw hash",
430
+ "raw_bytes": 1268157440
431
+ },
432
+ {
433
+ "window_id": "conditional-fit-0074",
434
+ "domain": "axis2_legal",
435
+ "prediction_rows": 2047,
436
+ "true_decode_rows": 2046,
437
+ "true_decode_mean_kld": 0.026004243654354962,
438
+ "including_prefill_mean_kld": 0.026482970726161778,
439
+ "prefill_row_kld": 1.0059585596428908,
440
+ "raw_sha256": "fc3a11b48cff4d3da57912972bba65140339c038d1ee74580cec14a3472efc4b",
441
+ "score_sha256": "6273db0cc03541162097c5911869fc2a5376c734a62f36c23bb85e9fc30900bf",
442
+ "teacher_sha256": "440341730d063ae27bbc15bcbfc879d9d16b3b25691d97b8cfd60be8f32f2313",
443
+ "token_sha256": "6f6618094ffce3f44b50f94cc533d338a3bfbbda7432890cd5ad2c13efd50129",
444
+ "raw_retirement": "after durable score and verified raw hash",
445
+ "raw_bytes": 1268157440
446
+ },
447
+ {
448
+ "window_id": "conditional-fit-0099",
449
+ "domain": "axis3_code_agentic",
450
+ "prediction_rows": 2047,
451
+ "true_decode_rows": 2046,
452
+ "true_decode_mean_kld": 0.012071783373105112,
453
+ "including_prefill_mean_kld": 0.01209520758653113,
454
+ "prefill_row_kld": 0.06002114825616887,
455
+ "raw_sha256": "b058724c3a3db126a6257067f93ed860b1210706894fa8e47fde2a4567d5638d",
456
+ "score_sha256": "a3845cb8f5ae202a29c727e46dcfc3ef14839d4b29f1d4cc8c4e84bcd7b25368",
457
+ "teacher_sha256": "da8dea840dd57564f0b3675872eaf297631b38a6d8dabcde7f8b54db06623e9a",
458
+ "token_sha256": "07e37956eb606026bdd7479ac5516275e92c0ee834c3c0439a3f300f9099608d",
459
+ "raw_retirement": "after durable score and verified raw hash",
460
+ "raw_bytes": 1268157440
461
+ },
462
+ {
463
+ "window_id": "conditional-fit-0113",
464
+ "domain": "axis2_legal",
465
+ "prediction_rows": 2047,
466
+ "true_decode_rows": 2046,
467
+ "true_decode_mean_kld": 0.031512954836832044,
468
+ "including_prefill_mean_kld": 0.03204969865465966,
469
+ "prefill_row_kld": 1.130227549929964,
470
+ "raw_sha256": "a1d485c18de7d34f6f481af8457a5cfaf9a99f966e30b42ceb8f55c451cd2833",
471
+ "score_sha256": "51130d113ad231c69f8a5bd8182d13aef01224ffc5021f39ef3bfd34229e9fe3",
472
+ "teacher_sha256": "6dc5f38f3436d6833b0ab68ce34e0f7ee5cde80359c26602526d8de512fd3207",
473
+ "token_sha256": "742ab50951388e712d20dab49b12ad7eb31490cfe5091a2fbc76f31ed8fa1b09",
474
+ "raw_retirement": "after durable score and verified raw hash",
475
+ "raw_bytes": 1268157440
476
+ },
477
+ {
478
+ "window_id": "conditional-fit-0118",
479
+ "domain": "axis1_general",
480
+ "prediction_rows": 2047,
481
+ "true_decode_rows": 2046,
482
+ "true_decode_mean_kld": 0.010197144240426125,
483
+ "including_prefill_mean_kld": 0.01043432648845951,
484
+ "prefill_row_kld": 0.49570920596476387,
485
+ "raw_sha256": "0158d84bd1af408c659df3de244c9b1be7248703c994b440f263268e2b6c1b21",
486
+ "score_sha256": "41fd3b38d432de5b97802a65f8ee6172a4ad7f8dea34db44c914fd57a6b748f7",
487
+ "teacher_sha256": "e296287593aa0dc89b695a6a082a224a85904757c4a3cefa69729394d1469f89",
488
+ "token_sha256": "2e03fdfcf4d54fc3174ee4620d5efd9f2e5c10e2e6e0b1217577f08392a09749",
489
+ "raw_retirement": "after durable score and verified raw hash",
490
+ "raw_bytes": 1268157440
491
+ }
492
+ ],
493
+ "domains": {
494
+ "axis1_general": 0.02385348680276958,
495
+ "axis2_legal": 0.033519798585609044,
496
+ "axis3_code_agentic": 0.03407317167326848,
497
+ "axis4_reasoning_termination": 0.021313254146301485
498
+ },
499
+ "correctness_profile_engine_capacity_tokens": [
500
+ 24942148
501
+ ]
502
+ },
503
+ "nvfp4": {
504
+ "mean_true_decode_kld": 0.03047853810548577,
505
+ "window_bca95": [
506
+ 0.026326096738464547,
507
+ 0.03570441626879746
508
+ ],
509
+ "per_window": [
510
+ {
511
+ "window_id": "conditional-fit-0056",
512
+ "domain": "axis1_general",
513
+ "prediction_rows": 2047,
514
+ "true_decode_rows": 2046,
515
+ "true_decode_mean_kld": 0.017746517936888917,
516
+ "including_prefill_mean_kld": 0.01851573129320549,
517
+ "prefill_row_kld": 1.5923262583169164,
518
+ "raw_sha256": "6491e15ab1d95ca20655d9adb49e8bbd01be21e3a173c4023c3e69afb6025d82",
519
+ "score_sha256": "64aa90f615fbd6627f63149354d7caaa23208d128919047c6858c1580ef94b14",
520
+ "teacher_sha256": "55a70a330f39c5f23294c23d270a1ddc53031def99a42206e2706d4d0bb4466b",
521
+ "token_sha256": "219d8cd8873c6cf41f52782297cc5cfa9c4ed9666f2bc30d84cee9caa68ea744",
522
+ "raw_retirement": "after durable score and verified raw hash",
523
+ "raw_bytes": 1268157440
524
+ },
525
+ {
526
+ "window_id": "conditional-fit-0008",
527
+ "domain": "axis1_general",
528
+ "prediction_rows": 2047,
529
+ "true_decode_rows": 2046,
530
+ "true_decode_mean_kld": 0.05626682254319565,
531
+ "including_prefill_mean_kld": 0.05686490218529273,
532
+ "prefill_row_kld": 1.2805358499159034,
533
+ "raw_sha256": "d6a7c0a473761ce011d8f3211bd51b1c0747af28bbfc715a5ef309d5819d4fdb",
534
+ "score_sha256": "bdae67c585d91399a114f68a6c036388b847c90aab64ba4103473f553796ba2c",
535
+ "teacher_sha256": "040641247b2e060035d89c2e7345e029d7c142983585e7c5d63dc818486f650d",
536
+ "token_sha256": "581d72e3e78b15c02d9e73f6bc742260cb57cdcd74337f4419c8f90cbc7c9c11",
537
+ "raw_retirement": "after durable score and verified raw hash",
538
+ "raw_bytes": 1268157440
539
+ },
540
+ {
541
+ "window_id": "conditional-fit-0100",
542
+ "domain": "axis1_general",
543
+ "prediction_rows": 2047,
544
+ "true_decode_rows": 2046,
545
+ "true_decode_mean_kld": 0.021613741186795158,
546
+ "including_prefill_mean_kld": 0.022109190690777637,
547
+ "prefill_row_kld": 1.035798875838937,
548
+ "raw_sha256": "6910f0d1e0bd8436885fc4bd17c9ab0523f6bcf91256cfd60657d30d065b8924",
549
+ "score_sha256": "c35cea7ec4bc16ca8447852f03057b18967360cd8e3861176c9999c99974e561",
550
+ "teacher_sha256": "f62991383b3ef5e8db0ed71442b17c428972e58da1c868c6aeb0f3fa143c865b",
551
+ "token_sha256": "7f008ec99068f373c6e15dd059ee977fb70e6dc832f46e009f0f299d48d19892",
552
+ "raw_retirement": "after durable score and verified raw hash",
553
+ "raw_bytes": 1268157440
554
+ },
555
+ {
556
+ "window_id": "conditional-fit-0094",
557
+ "domain": "axis1_general",
558
+ "prediction_rows": 2047,
559
+ "true_decode_rows": 2046,
560
+ "true_decode_mean_kld": 0.01800379452343463,
561
+ "including_prefill_mean_kld": 0.018188032876869008,
562
+ "prefill_row_kld": 0.3951397040035959,
563
+ "raw_sha256": "6d8119d885b4af13eb38861ac7c0404b97ad009f003d73f52701b75e96238f6c",
564
+ "score_sha256": "cf897a078082fb7f2ac0c5a4ae117954abbde098c4fec8fa0b984d51a427fe7d",
565
+ "teacher_sha256": "656911b8fb34d7f42e984dd21311caf9ef56db0b7618b414cad4089562d93cba",
566
+ "token_sha256": "6b6a9b21c49a9984cc8c5ea90e0bda6db5761d039c7cdd335dc65e30444e017b",
567
+ "raw_retirement": "after durable score and verified raw hash",
568
+ "raw_bytes": 1268157440
569
+ },
570
+ {
571
+ "window_id": "conditional-fit-0083",
572
+ "domain": "axis2_legal",
573
+ "prediction_rows": 2047,
574
+ "true_decode_rows": 2046,
575
+ "true_decode_mean_kld": 0.03224313772811532,
576
+ "including_prefill_mean_kld": 0.03880860459782181,
577
+ "prefill_row_kld": 13.471753820017287,
578
+ "raw_sha256": "2b54dd1dfeb5ed9b02d8feea06bd6087679c5f0bd6d85a6d39622f34e4121fd3",
579
+ "score_sha256": "aedb9e1bb7b73ef8ab3010bc3895abcecbe7cab8c51d2c395650367e30fb9bc2",
580
+ "teacher_sha256": "66a7870a1f13c8b5d647c1681fca005bb0f55a6e502e94f282da673a6bed3f1b",
581
+ "token_sha256": "3df9d2872815e061ff8d32a9a19da0d11d80ac67070cccd28579b24ec3079552",
582
+ "raw_retirement": "after durable score and verified raw hash",
583
+ "raw_bytes": 1268157440
584
+ },
585
+ {
586
+ "window_id": "conditional-fit-0021",
587
+ "domain": "axis2_legal",
588
+ "prediction_rows": 2047,
589
+ "true_decode_rows": 2046,
590
+ "true_decode_mean_kld": 0.032370031971847645,
591
+ "including_prefill_mean_kld": 0.03375160585353929,
592
+ "prefill_row_kld": 2.86045176779466,
593
+ "raw_sha256": "0400319d0b4c5692904b52ba45ecb8b2a6d6553ed5209ca32ff9e232960a2919",
594
+ "score_sha256": "527fba07ed59b9f9f88428b63067cf137dce3f34ca8f1fc0fad13505af326cae",
595
+ "teacher_sha256": "ebf76910c4db8f6955ad03f2ec7352e2f8aaf15e0c939d28f9a7d78f6bd7451b",
596
+ "token_sha256": "dbd9f2539cb3e97127ae760a04ca5b3456b0b13d6ce8698f88b215cc68b8f1a9",
597
+ "raw_retirement": "after durable score and verified raw hash",
598
+ "raw_bytes": 1268157440
599
+ },
600
+ {
601
+ "window_id": "conditional-fit-0098",
602
+ "domain": "axis2_legal",
603
+ "prediction_rows": 2047,
604
+ "true_decode_rows": 2046,
605
+ "true_decode_mean_kld": 0.032517442948549406,
606
+ "including_prefill_mean_kld": 0.03278446891938593,
607
+ "prefill_row_kld": 0.5791196052509112,
608
+ "raw_sha256": "5b87b06965c10b3004c8c48b8c82ba2d4d0d7506adac90cc41cab4d2b55d95d8",
609
+ "score_sha256": "42f6d2de57b6c31bc259495cb8c5072051fe270ed8b90cbe4800754b890a8619",
610
+ "teacher_sha256": "16e3d9524100836c020535635f7722a4c93ce0fdddc2ad0781f25bda77e8bcc9",
611
+ "token_sha256": "749bda3291a0323d286b0018b7dc3e0f469de8afd5d194370cde357b7b315c7d",
612
+ "raw_retirement": "after durable score and verified raw hash",
613
+ "raw_bytes": 1268157440
614
+ },
615
+ {
616
+ "window_id": "conditional-fit-0080",
617
+ "domain": "axis2_legal",
618
+ "prediction_rows": 2047,
619
+ "true_decode_rows": 2046,
620
+ "true_decode_mean_kld": 0.0299181930475973,
621
+ "including_prefill_mean_kld": 0.03007133969215208,
622
+ "prefill_row_kld": 0.343409374451233,
623
+ "raw_sha256": "7f21b6b28f274de8e67659eb9e891aa55977b248ce887843e4a8320529f24e26",
624
+ "score_sha256": "707ef9169740952d4b6a78edbc0c4d2d55264ec15d2b4e67dfaeaced842e348a",
625
+ "teacher_sha256": "e9c1a77490a5db63d410297bd8febe06adf4df2a6f4e16552db187738c9289d2",
626
+ "token_sha256": "a31d43f38710eb305c684501a2b02909a4e82c43101c07e76e4a3ca4edb4d0dd",
627
+ "raw_retirement": "after durable score and verified raw hash",
628
+ "raw_bytes": 1268157440
629
+ },
630
+ {
631
+ "window_id": "conditional-fit-0081",
632
+ "domain": "axis3_code_agentic",
633
+ "prediction_rows": 2047,
634
+ "true_decode_rows": 2046,
635
+ "true_decode_mean_kld": 0.052740365545818985,
636
+ "including_prefill_mean_kld": 0.05405242932673105,
637
+ "prefill_row_kld": 2.738534925072807,
638
+ "raw_sha256": "13a6c84c0f11b7412ffa714550b4463e8be2eff681d1fafcfe25083303fb8477",
639
+ "score_sha256": "0ceb0ca2908082543ad211a12e492327dc6f17f080f522e5272d924cf9393079",
640
+ "teacher_sha256": "5ae07760d91030d1eba925b76ea463309baa49c2e7d5f5e14409a23ced7fd20d",
641
+ "token_sha256": "350628488b94d59f3e4f3e99a418913544ea1fe7eacee14a642b996d9c614d16",
642
+ "raw_retirement": "after durable score and verified raw hash",
643
+ "raw_bytes": 1268157440
644
+ },
645
+ {
646
+ "window_id": "conditional-fit-0123",
647
+ "domain": "axis3_code_agentic",
648
+ "prediction_rows": 2047,
649
+ "true_decode_rows": 2046,
650
+ "true_decode_mean_kld": 0.048590810367257806,
651
+ "including_prefill_mean_kld": 0.049930734905846354,
652
+ "prefill_row_kld": 2.791416340858005,
653
+ "raw_sha256": "d8087072915e90cfc6726fe31673e5eeef694105e21f3ec1ab216b863cdd0cae",
654
+ "score_sha256": "9f9437e38f0ee4b314666c69891eb17110f06d6a4f37eb73dda3cc8042691b68",
655
+ "teacher_sha256": "4a24b5b750ba98c06d9bc69e19334aff89e7bd4f9429592598af5d62de7ecb42",
656
+ "token_sha256": "aab84ee426f91621d123aa68bd53f75e220fbaf8cb270db1feab1529a62ddbf0",
657
+ "raw_retirement": "after durable score and verified raw hash",
658
+ "raw_bytes": 1268157440
659
+ },
660
+ {
661
+ "window_id": "conditional-fit-0090",
662
+ "domain": "axis3_code_agentic",
663
+ "prediction_rows": 2047,
664
+ "true_decode_rows": 2046,
665
+ "true_decode_mean_kld": 0.037893750945486235,
666
+ "including_prefill_mean_kld": 0.03874853429194332,
667
+ "prefill_row_kld": 1.7876352611431368,
668
+ "raw_sha256": "636db52da0da40e9962c2bd078a261d488324dc7b340ae68a788ea226f6b2dc3",
669
+ "score_sha256": "629cc66fb44cd5d225e3886f3cd60d801b779cba228a00bb0ba2559a6e302912",
670
+ "teacher_sha256": "db2a5fe33c9e457a22ebeb31a7073c1dac666b877503e9b4e90a8d6abf7b8a32",
671
+ "token_sha256": "a06e76993db717afee847fcc4d86f10b52dd23cdce5dae916f32d238d08d886c",
672
+ "raw_retirement": "after durable score and verified raw hash",
673
+ "raw_bytes": 1268157440
674
+ },
675
+ {
676
+ "window_id": "conditional-fit-0062",
677
+ "domain": "axis3_code_agentic",
678
+ "prediction_rows": 2047,
679
+ "true_decode_rows": 2046,
680
+ "true_decode_mean_kld": 0.024132677263350163,
681
+ "including_prefill_mean_kld": 0.02473193056836959,
682
+ "prefill_row_kld": 1.250804192638108,
683
+ "raw_sha256": "004aedbb025d73e028f6b65854e3dad2edee3e30d4087802a9bd63aa61df4cf6",
684
+ "score_sha256": "f1fb466ed49b458271027e76e7dc648e381084251d5f2a5237ff3f2936322bf6",
685
+ "teacher_sha256": "d6f7883a533625c7c75119fe7403b48bf81885b1de0b311a1a0571949f7f231c",
686
+ "token_sha256": "7c6d2ef2f4b7c7091a517f6ccd778009f9ee144165c9aecf2f8f28aea7b3753f",
687
+ "raw_retirement": "after durable score and verified raw hash",
688
+ "raw_bytes": 1268157440
689
+ },
690
+ {
691
+ "window_id": "conditional-fit-0047",
692
+ "domain": "axis4_reasoning_termination",
693
+ "prediction_rows": 2047,
694
+ "true_decode_rows": 2046,
695
+ "true_decode_mean_kld": 0.03135213041927493,
696
+ "including_prefill_mean_kld": 0.03828158589133786,
697
+ "prefill_row_kld": 14.21594748173208,
698
+ "raw_sha256": "c8d2fdf91b1e4355f428f72f0a0181dd355aaa5ea2ec212386e50727278b4b4c",
699
+ "score_sha256": "e3c6958be8754e7abc39e03c85157a5884cbcc8763f5b914d529e249dcf8c286",
700
+ "teacher_sha256": "8513cdcb9940786c7c42bf14d2f5f2cd452791e4ec6036c06a1f973eb969018d",
701
+ "token_sha256": "89730ee6826d4302c352156d5df766c51622a8b6681b8052d4ca6e7d481177fe",
702
+ "raw_retirement": "after durable score and verified raw hash",
703
+ "raw_bytes": 1268157440
704
+ },
705
+ {
706
+ "window_id": "conditional-fit-0051",
707
+ "domain": "axis4_reasoning_termination",
708
+ "prediction_rows": 2047,
709
+ "true_decode_rows": 2046,
710
+ "true_decode_mean_kld": 0.01865448654587657,
711
+ "including_prefill_mean_kld": 0.02522659174053774,
712
+ "prefill_row_kld": 13.471753820017287,
713
+ "raw_sha256": "b559fb1c0eeff31756672abec66a163b89a7da3a6795d9f44ffe4b311e565ee2",
714
+ "score_sha256": "84e0bd93f12917045bf701bd4fa6d40035540160a58ee0431637b89f03df2bf3",
715
+ "teacher_sha256": "177477c19b54dc0b0230e19202d48ab71d52f2b1a37fe383511bb64f0aae2fcf",
716
+ "token_sha256": "c79269b16560e7301549593443c9394007e7a957ad149c46224e7ad271dc6a18",
717
+ "raw_retirement": "after durable score and verified raw hash",
718
+ "raw_bytes": 1268157440
719
+ },
720
+ {
721
+ "window_id": "conditional-fit-0055",
722
+ "domain": "axis4_reasoning_termination",
723
+ "prediction_rows": 2047,
724
+ "true_decode_rows": 2046,
725
+ "true_decode_mean_kld": 0.022170981387968584,
726
+ "including_prefill_mean_kld": 0.02258329765519538,
727
+ "prefill_row_kld": 0.8661823804012292,
728
+ "raw_sha256": "9719e9768a9925f58093732d9714d5fbb3aea468c7050acd3cd0792cbd881251",
729
+ "score_sha256": "d0e2d9144c32318129ee466ed25bb0edc602b923509ab6d156ec5409401acf4e",
730
+ "teacher_sha256": "7679c828eae4bf08f17598d044904693fc90c05b9d3f189dddeeb846100e43a7",
731
+ "token_sha256": "feec9ac508172dd275ec6452a03493acab5ffdff9bb57161317c77f24daf90f2",
732
+ "raw_retirement": "after durable score and verified raw hash",
733
+ "raw_bytes": 1268157440
734
+ },
735
+ {
736
+ "window_id": "conditional-fit-0031",
737
+ "domain": "axis4_reasoning_termination",
738
+ "prediction_rows": 2047,
739
+ "true_decode_rows": 2046,
740
+ "true_decode_mean_kld": 0.019741844191964442,
741
+ "including_prefill_mean_kld": 0.026676971518559515,
742
+ "prefill_row_kld": 14.21594748173208,
743
+ "raw_sha256": "96fb2126b8e205c3c495ce22f4dc874a92356c05659907c90260cf3d9f51ffb3",
744
+ "score_sha256": "702d5542b4a85c14800d50162427b556f45f2132b941c1f73fdebd4b4230ed0a",
745
+ "teacher_sha256": "e04a6c04a951efb5a178927e91457535cc67745ce5c81e785cb6d963aac6ce83",
746
+ "token_sha256": "2cff3482b4212537d0c35b88b8a0f479502c8905952430c973d25b1ab0bb5818",
747
+ "raw_retirement": "after durable score and verified raw hash",
748
+ "raw_bytes": 1268157440
749
+ },
750
+ {
751
+ "window_id": "conditional-fit-0003",
752
+ "domain": "axis4_reasoning_termination",
753
+ "prediction_rows": 2047,
754
+ "true_decode_rows": 2046,
755
+ "true_decode_mean_kld": 0.029583912479471508,
756
+ "including_prefill_mean_kld": 0.0327787433732799,
757
+ "prefill_row_kld": 6.5694027521052485,
758
+ "raw_sha256": "954b19fced659109f00fe7982d9b2ab923526b4b802f0321caee26a8f65a4238",
759
+ "score_sha256": "7e805292cc17606d83742f49c1df34de2ff5b26a67ca41d6c7f61c368ce40c09",
760
+ "teacher_sha256": "7677c89cc3dee8e2a64e3c7947add39475f828921a3a8fe2e92b379b3ed8320b",
761
+ "token_sha256": "d416537c6e8273747bf025e97287a0bb6e3ff047bfee65cf54e4dfe287569b93",
762
+ "raw_retirement": "after durable score and verified raw hash",
763
+ "raw_bytes": 1268157440
764
+ },
765
+ {
766
+ "window_id": "conditional-fit-0009",
767
+ "domain": "axis2_legal",
768
+ "prediction_rows": 2047,
769
+ "true_decode_rows": 2046,
770
+ "true_decode_mean_kld": 0.03265315307692232,
771
+ "including_prefill_mean_kld": 0.03271898198293517,
772
+ "prefill_row_kld": 0.1674049236852354,
773
+ "raw_sha256": "787798fb04be0d8e67bbe4e8a093239f92911fcb11a3b84e0b7457209fc8cd7d",
774
+ "score_sha256": "6776a73abb4f04e9618ded353b69528bf994727853f7e7380ceb61c495374e0b",
775
+ "teacher_sha256": "00701ad8bf4a4a5eedcedd4eafd956830c9ff1c790b15074ca721a3462abf930",
776
+ "token_sha256": "fc1166707f5938a914faf364dab717b0de3ea0e8451ed1a3fb399469e659b105",
777
+ "raw_retirement": "after durable score and verified raw hash",
778
+ "raw_bytes": 1268157440
779
+ },
780
+ {
781
+ "window_id": "conditional-fit-0010",
782
+ "domain": "axis3_code_agentic",
783
+ "prediction_rows": 2047,
784
+ "true_decode_rows": 2046,
785
+ "true_decode_mean_kld": 0.0243919341540783,
786
+ "including_prefill_mean_kld": 0.025417884511426313,
787
+ "prefill_row_kld": 2.1245123156454757,
788
+ "raw_sha256": "d6ff9c1db93a38e19149e7161a3cd894cb0c659585d308e23665f79ab2cb2977",
789
+ "score_sha256": "8be897b402e76673cc9572a59aafb69db975f71fde51f9ef4cd260169866b56c",
790
+ "teacher_sha256": "f9bf4519890b70d972dfdf9760181418eb7c55b6237c0108f01d44df89943278",
791
+ "token_sha256": "ddd9d769a76eba74bc8e680ff25736a3b9146371801d1d2b29e4dc48f6d256a2",
792
+ "raw_retirement": "after durable score and verified raw hash",
793
+ "raw_bytes": 1268157440
794
+ },
795
+ {
796
+ "window_id": "conditional-fit-0011",
797
+ "domain": "axis4_reasoning_termination",
798
+ "prediction_rows": 2047,
799
+ "true_decode_rows": 2046,
800
+ "true_decode_mean_kld": 0.019764568124150814,
801
+ "including_prefill_mean_kld": 0.026699684349655417,
802
+ "prefill_row_kld": 14.21594748173208,
803
+ "raw_sha256": "d3710139d24820e349c860fb21deb8992818333747babc4f5902bed9ff70d5db",
804
+ "score_sha256": "5c4f3ab0d6939d026f3942c4ce4edfeeae221dcd41948f8011778fdda9ce056e",
805
+ "teacher_sha256": "9f37c4ead38bb80a5d85f825b156e5b493b6d3f1ca9dabfad06b0416f37db286",
806
+ "token_sha256": "c4b96acfa47944be3485b20eab76eb16d59913c9fedb5001a537335a3ee13614",
807
+ "raw_retirement": "after durable score and verified raw hash",
808
+ "raw_bytes": 1268157440
809
+ },
810
+ {
811
+ "window_id": "conditional-fit-0030",
812
+ "domain": "axis3_code_agentic",
813
+ "prediction_rows": 2047,
814
+ "true_decode_rows": 2046,
815
+ "true_decode_mean_kld": 0.03899826510937706,
816
+ "including_prefill_mean_kld": 0.040525864524546235,
817
+ "prefill_row_kld": 3.165994267960693,
818
+ "raw_sha256": "274aff4d828b06b7bb9f9616cddf8804fc2dca5e3d740f71f1f08e5015e25a7a",
819
+ "score_sha256": "f0913b4b4c3bcaa83a6a6e57a1131a29d193337d8458a7546db5d7fb08cf638e",
820
+ "teacher_sha256": "90c08d120cfac926855e15c106dba1e242099bcd0ae6da7655bfaa54d5ce2420",
821
+ "token_sha256": "d644a79da2ba01cefc5f2020eab9e8662d45ba3025377844e4bf4d3f1d36802d",
822
+ "raw_retirement": "after durable score and verified raw hash",
823
+ "raw_bytes": 1268157440
824
+ },
825
+ {
826
+ "window_id": "conditional-fit-0032",
827
+ "domain": "axis1_general",
828
+ "prediction_rows": 2047,
829
+ "true_decode_rows": 2046,
830
+ "true_decode_mean_kld": 0.021563674524295112,
831
+ "including_prefill_mean_kld": 0.02261755095926481,
832
+ "prefill_row_kld": 2.178848736907266,
833
+ "raw_sha256": "a5cd3e32913df74d47b0a876dfc95e949362e5b31040b425cf1deebe86fa9d9c",
834
+ "score_sha256": "4b0052992b6a59a7fb5c78e8096483b1a0588ad54c610126722fcd898d08e528",
835
+ "teacher_sha256": "1aa70ccf8e224c3a58834e2232a3a2477768db29ffeca39f5b27c17ad8c40c2b",
836
+ "token_sha256": "88745c333ead140a6fb05929031613429cbd42e25881d92add968a94a5d2d930",
837
+ "raw_retirement": "after durable score and verified raw hash",
838
+ "raw_bytes": 1268157440
839
+ },
840
+ {
841
+ "window_id": "conditional-fit-0035",
842
+ "domain": "axis4_reasoning_termination",
843
+ "prediction_rows": 2047,
844
+ "true_decode_rows": 2046,
845
+ "true_decode_mean_kld": 0.013848512477088149,
846
+ "including_prefill_mean_kld": 0.01384591365068994,
847
+ "prefill_row_kld": 0.008528714839950826,
848
+ "raw_sha256": "976688ae8f2c4e42609faafc6a2a733bfc0582a4bd27ee68011b107d46ebca29",
849
+ "score_sha256": "f4dc38b6a3e8b6f08b9b515e1d78a25c6c96b8dccd7d9bf805bbef8d94303fe8",
850
+ "teacher_sha256": "4e958dce74e093eb62ba5ac210fe0a4e50f12f1875de678769700cde206b429d",
851
+ "token_sha256": "49f702136da39db9e762b6815a1952c0eba0a1fb04af22fb28a27bb63e2a41dc",
852
+ "raw_retirement": "after durable score and verified raw hash",
853
+ "raw_bytes": 1268157440
854
+ },
855
+ {
856
+ "window_id": "conditional-fit-0040",
857
+ "domain": "axis1_general",
858
+ "prediction_rows": 2047,
859
+ "true_decode_rows": 2046,
860
+ "true_decode_mean_kld": 0.04249810688224839,
861
+ "including_prefill_mean_kld": 0.04340623494150831,
862
+ "prefill_row_kld": 1.9014362441872743,
863
+ "raw_sha256": "e3bc0ea2242eeb3863a83c34299451c3f268119819e745c4349880d1dc6b382c",
864
+ "score_sha256": "1ee893f5cc53a872e454b2d1c59a0fb4c03d847c45492bb1b29ca839bb2a374c",
865
+ "teacher_sha256": "f4c96798fc9eb5564d9aec52779cc0ce88bc55312dcd7663240938d711c8a90d",
866
+ "token_sha256": "50f3bc08ab37c551a05364eb48f0d726742f19331e79114dd5a9fb3c42ce6d4f",
867
+ "raw_retirement": "after durable score and verified raw hash",
868
+ "raw_bytes": 1268157440
869
+ },
870
+ {
871
+ "window_id": "conditional-fit-0041",
872
+ "domain": "axis2_legal",
873
+ "prediction_rows": 2047,
874
+ "true_decode_rows": 2046,
875
+ "true_decode_mean_kld": 0.06821517978445629,
876
+ "including_prefill_mean_kld": 0.0689047479976133,
877
+ "prefill_row_kld": 1.479761312116889,
878
+ "raw_sha256": "5d6ee43966c5f0003561aca12d117c464c6ef93fefdce80b44dc59a5cf344489",
879
+ "score_sha256": "c42329e082770e11b073a244c69b798d306c4fa0032b54eb53ad87d41c2e868c",
880
+ "teacher_sha256": "c2bbe4ea31532baceff966423c6ce6e042aebb5a2eb1de45e382a7d3df6f6827",
881
+ "token_sha256": "8acf2614f0f787c185af112f21bcdf247c9b6ae76d292ada28f476e2a7b0241d",
882
+ "raw_retirement": "after durable score and verified raw hash",
883
+ "raw_bytes": 1268157440
884
+ },
885
+ {
886
+ "window_id": "conditional-fit-0046",
887
+ "domain": "axis3_code_agentic",
888
+ "prediction_rows": 2047,
889
+ "true_decode_rows": 2046,
890
+ "true_decode_mean_kld": 0.051633578140284686,
891
+ "including_prefill_mean_kld": 0.05171196180589413,
892
+ "prefill_row_kld": 0.21208494164282748,
893
+ "raw_sha256": "98edf2f9d7d63f8a9f44764e4bf14372a74e51f8973c6f4567eb99b44adaeece",
894
+ "score_sha256": "8bf86881944a318f90d54907d4f2dac3e8aa06d3aee040494a27278047179821",
895
+ "teacher_sha256": "3caa5c58fb0b0e80129d551b39867fd52fefd93830060bb5655862fdd0e3b8ce",
896
+ "token_sha256": "3f2df1c090abf4473163a078c168fec33ab98ed6415985b92449caa861c04444",
897
+ "raw_retirement": "after durable score and verified raw hash",
898
+ "raw_bytes": 1268157440
899
+ },
900
+ {
901
+ "window_id": "conditional-fit-0060",
902
+ "domain": "axis1_general",
903
+ "prediction_rows": 2047,
904
+ "true_decode_rows": 2046,
905
+ "true_decode_mean_kld": 0.017313252044285136,
906
+ "including_prefill_mean_kld": 0.018362279720832667,
907
+ "prefill_row_kld": 2.1646729059370715,
908
+ "raw_sha256": "7de3e05d55fc294f64c01e69a53f0b810fa6d9bccb86049891c16d1c58403490",
909
+ "score_sha256": "21cf45d738834141c7cbe70e13f2fe49a3964f0121c8624638eb41e296eba8df",
910
+ "teacher_sha256": "d9e6344664b92527ee58d3bca038702676b357e7052ade03ac4bc9c70908cc34",
911
+ "token_sha256": "04ffe38aafcd9407b2846d59e8d797bee04b97f00683f8bc74d475a2d6b8592b",
912
+ "raw_retirement": "after durable score and verified raw hash",
913
+ "raw_bytes": 1268157440
914
+ },
915
+ {
916
+ "window_id": "conditional-fit-0063",
917
+ "domain": "axis4_reasoning_termination",
918
+ "prediction_rows": 2047,
919
+ "true_decode_rows": 2046,
920
+ "true_decode_mean_kld": 0.026396085671889582,
921
+ "including_prefill_mean_kld": 0.026907175741609994,
922
+ "prefill_row_kld": 1.0725974583895699,
923
+ "raw_sha256": "c3aa9cf4ccf4bd98d3f6adec969f1616bf1e23ec5c5309beac44e8b990703734",
924
+ "score_sha256": "d62e911a296524f0a72af50f0b56efa6d213c4b3e9c4cc7c0954a02d57f0ff0d",
925
+ "teacher_sha256": "a9bb59659437c1df5b3f3ef0b54eda61402842a2597b76b0124f58b59ecf3ab4",
926
+ "token_sha256": "38003e2d26db850dd8ff39f2e95fa3d1be9a34838c70d9f7e87056b7ef50b67d",
927
+ "raw_retirement": "after durable score and verified raw hash",
928
+ "raw_bytes": 1268157440
929
+ },
930
+ {
931
+ "window_id": "conditional-fit-0074",
932
+ "domain": "axis2_legal",
933
+ "prediction_rows": 2047,
934
+ "true_decode_rows": 2046,
935
+ "true_decode_mean_kld": 0.029326217586224376,
936
+ "including_prefill_mean_kld": 0.029896648143729315,
937
+ "prefill_row_kld": 1.1969975687988312,
938
+ "raw_sha256": "1aa8113d5b3b2cdf81ad0e3483bfb6902edb505fb34fbf78ade44909cd08b966",
939
+ "score_sha256": "b13021bc3b7bfe0e0aeba1b3233876bd551e3f2ad0b897d5a0b41aa05678eac1",
940
+ "teacher_sha256": "440341730d063ae27bbc15bcbfc879d9d16b3b25691d97b8cfd60be8f32f2313",
941
+ "token_sha256": "6f6618094ffce3f44b50f94cc533d338a3bfbbda7432890cd5ad2c13efd50129",
942
+ "raw_retirement": "after durable score and verified raw hash",
943
+ "raw_bytes": 1268157440
944
+ },
945
+ {
946
+ "window_id": "conditional-fit-0099",
947
+ "domain": "axis3_code_agentic",
948
+ "prediction_rows": 2047,
949
+ "true_decode_rows": 2046,
950
+ "true_decode_mean_kld": 0.014314480651288897,
951
+ "including_prefill_mean_kld": 0.014320539180160559,
952
+ "prefill_row_kld": 0.026716289251575875,
953
+ "raw_sha256": "e3c69ffe17dd14bbb05c009fbd87c809eaa47dcbd0f20f6ffc1abf2ee7b47a4a",
954
+ "score_sha256": "f7d755188f14f4f3a00f988e41cdc1b2fbe4e0038f22ba7fdadea7f40735ee82",
955
+ "teacher_sha256": "da8dea840dd57564f0b3675872eaf297631b38a6d8dabcde7f8b54db06623e9a",
956
+ "token_sha256": "07e37956eb606026bdd7479ac5516275e92c0ee834c3c0439a3f300f9099608d",
957
+ "raw_retirement": "after durable score and verified raw hash",
958
+ "raw_bytes": 1268157440
959
+ },
960
+ {
961
+ "window_id": "conditional-fit-0113",
962
+ "domain": "axis2_legal",
963
+ "prediction_rows": 2047,
964
+ "true_decode_rows": 2046,
965
+ "true_decode_mean_kld": 0.03746853211926064,
966
+ "including_prefill_mean_kld": 0.038173120678126116,
967
+ "prefill_row_kld": 1.479761312116889,
968
+ "raw_sha256": "c95b49deeed5558206431e36a6cf85a980ca5d32cdf1ffeb1d4cf46af759d548",
969
+ "score_sha256": "29912ecf9f7745859d110d893fd7c9e328ca8ac9839774aa43c288bdee286e38",
970
+ "teacher_sha256": "6dc5f38f3436d6833b0ab68ce34e0f7ee5cde80359c26602526d8de512fd3207",
971
+ "token_sha256": "742ab50951388e712d20dab49b12ad7eb31490cfe5091a2fbc76f31ed8fa1b09",
972
+ "raw_retirement": "after durable score and verified raw hash",
973
+ "raw_bytes": 1268157440
974
+ },
975
+ {
976
+ "window_id": "conditional-fit-0118",
977
+ "domain": "axis1_general",
978
+ "prediction_rows": 2047,
979
+ "true_decode_rows": 2046,
980
+ "true_decode_mean_kld": 0.011387037996801677,
981
+ "including_prefill_mean_kld": 0.011681803391236737,
982
+ "prefill_row_kld": 0.6147718004053654,
983
+ "raw_sha256": "0cd1ee02214b3c0bfa1d230020541415fc0e8797ccdb931e3baabeb57174fb81",
984
+ "score_sha256": "a67dd96f3f21ce0deae94aef892654c383a2ff5a4a8c30e034923dc3fe93ceda",
985
+ "teacher_sha256": "e296287593aa0dc89b695a6a082a224a85904757c4a3cefa69729394d1469f89",
986
+ "token_sha256": "2e03fdfcf4d54fc3174ee4620d5efd9f2e5c10e2e6e0b1217577f08392a09749",
987
+ "raw_retirement": "after durable score and verified raw hash",
988
+ "raw_bytes": 1268157440
989
+ }
990
+ ],
991
+ "domains": {
992
+ "axis1_general": 0.025799118454743086,
993
+ "axis2_legal": 0.03683898603287167,
994
+ "axis3_code_agentic": 0.03658698277211777,
995
+ "axis4_reasoning_termination": 0.022689065162210573
996
+ },
997
+ "correctness_profile_engine_capacity_tokens": [
998
+ 39454545
999
+ ]
1000
+ }
1001
+ },
1002
+ "fp8_minus_nvfp4": -0.0022886103034986238,
1003
+ "paired_window_bca95": [
1004
+ -0.004297810897491988,
1005
+ -0.0001245637507589668
1006
+ ],
1007
+ "fp8_lower_windows": 26,
1008
+ "true_decode_rows_per_window": 2046,
1009
+ "windows": 32,
1010
+ "limits": [
1011
+ "TR3 r10 TP4/EP4/DCP4 versus reference r27 TP4/EPoff/DCP4; system comparison, not isolated quantizer attribution.",
1012
+ "MTP off and maxseq1 forced decode; not MTP quality or throughput evidence.",
1013
+ "Fixed FP8 then NVFP4 order; one server preparation per arm, window intervals do not capture server-run variability."
1014
+ ],
1015
+ "receipt_audit": [
1016
+ {
1017
+ "arm": "fp8",
1018
+ "windows": 32,
1019
+ "prediction_rows": 65504,
1020
+ "true_decode_rows": 65472,
1021
+ "all_receipt_hashes_and_masks_pass": true,
1022
+ "all_prefix_hit_counters_zero": true,
1023
+ "raw_logits_retired_after_scoring": true,
1024
+ "source_runtime_audit_sha256": "f0ecea1e32f5abd1a3ce6c016ddd904e26a0077e51666235f51a19326904e4c8",
1025
+ "source_startup_sha256": "f3a2774cb50c34e38e73653c3ee820322148f40a4c99c811f28ed435a2eb4fde"
1026
+ },
1027
+ {
1028
+ "arm": "nvfp4",
1029
+ "windows": 32,
1030
+ "prediction_rows": 65504,
1031
+ "true_decode_rows": 65472,
1032
+ "all_receipt_hashes_and_masks_pass": true,
1033
+ "all_prefix_hit_counters_zero": true,
1034
+ "raw_logits_retired_after_scoring": true,
1035
+ "source_runtime_audit_sha256": "f0596d465f77d7bf0d39bf3ec55e251422b58aa741434c8dce2963fbdc94d8cd",
1036
+ "source_startup_sha256": "cf960ddc018b9db95a713249f403e9b9571c1a596c1b5c46c606abc43e0c408e"
1037
+ }
1038
+ ],
1039
+ "cross_system_comparison": {
1040
+ "fp8": {
1041
+ "tr3_minus_reference": -0.0037552453645590413,
1042
+ "paired_window_bca95": [
1043
+ -0.0074398856791629215,
1044
+ 0.00034559699472292387
1045
+ ],
1046
+ "tr3_lower_windows": 22,
1047
+ "reference_mean": 0.03194517316654619,
1048
+ "tr3_mean": 0.028189927801987147
1049
+ },
1050
+ "nvfp4": {
1051
+ "tr3_minus_reference": -0.004977685666671991,
1052
+ "paired_window_bca95": [
1053
+ -0.010671394642632472,
1054
+ -0.0008185255081647577
1055
+ ],
1056
+ "tr3_lower_windows": 22,
1057
+ "reference_mean": 0.03545622377215776,
1058
+ "tr3_mean": 0.03047853810548577
1059
+ }
1060
+ },
1061
+ "model": "brandonmusic/GLM-5.3-Flash-tr3-4bpw",
1062
+ "model_revision": "aba59d2175e1ee2887ae0ae1300ba848b1deed84",
1063
+ "model_file_verification_sha256": "8a7be5d476e6286429ab169d93eb3e24fe53ce7d6c5f76f5b69c2d9e89bb9ed5",
1064
+ "capture_image_local_id": "sha256:62e069faf47f2d42eae5f2c1677f8730a2f3f93d576301fe1bcc7e55f2fdb673",
1065
+ "capture_source": {
1066
+ "image_id": "sha256:62e069faf47f2d42eae5f2c1677f8730a2f3f93d576301fe1bcc7e55f2fdb673",
1067
+ "base_image_id": "sha256:69e72a61008e6892b517b45e036c5df69a3851ea39698631a8f87e64adcec456",
1068
+ "hook_files": {
1069
+ "p8_decode_capture/v2_hook.py": "89e3528d2a54ee8ad90b60419bd86b559cdc8df10127d01f4c9486409007102e",
1070
+ "p8_decode_capture/processor.py": "8fb2f0d9c62f5b9b4d5abf32ac6b20be8b1f93052ddd42c36f4bf1f1a1a2a01d",
1071
+ "p8_decode_capture/__init__.py": "267c4f5a565687a18b224d04c1f5c702149ea8ac866986dc5922c084077fcb41"
1072
+ },
1073
+ "original_files": {
1074
+ "sampler": "e40830547705a6622fa7c4733791c261c0908245d376514d7ee891e587357cad",
1075
+ "warmup": "c15f6cdbc70680302d7c620b51f67319347c7111326bec22eb5e0b475107c956"
1076
+ },
1077
+ "patched_files": {
1078
+ "sampler": "0b081ae158bd0191ac94dbc8615b439197967a25c6286f4cd86b5efd40e0b8c5",
1079
+ "warmup": "70bc4dd415f4afd2120a4821a00b7e43bf088a142c5ac002142889b4a75abf41"
1080
+ },
1081
+ "preserved_runtime_difference": "r10 gumbel call lacks r27 is_drafting=False; original r10 call retained",
1082
+ "cpu_test": "cpu-test.log PASS; synthetic only, no GPU qualification",
1083
+ "capture_parent_image_id": "sha256:dbd042475c3f939a850fc7db3c88779fe09895ffce5d4fcdb93659f945a85bac",
1084
+ "compiler_cache_only_delta": {
1085
+ "image_id": "sha256:62e069faf47f2d42eae5f2c1677f8730a2f3f93d576301fe1bcc7e55f2fdb673",
1086
+ "parent_capture_image_id": "sha256:dbd042475c3f939a850fc7db3c88779fe09895ffce5d4fcdb93659f945a85bac",
1087
+ "changes": [
1088
+ "Add stock FileCacheManager subclass with PID namespace for normal cache keys; inherited get/put/group semantics",
1089
+ "TRITON_CACHE_DIR=/cache/tr3-triton-cache",
1090
+ "Replace launch /cache bind with --tmpfs /cache:rw,exec,size=16g"
1091
+ ],
1092
+ "hypothesis": "shared cross-rank metadata visibility or rename race on fuseblk; not conclusively reproduced",
1093
+ "observed": [
1094
+ "FileNotFoundError immediately after compiler cache.put and put_group, constructor reads metadata",
1095
+ "Missing JSON exists post-run; no symlinks in paths",
1096
+ "Host cache filesystem fuseblk",
1097
+ "TP1,TP3 failures while TP0 proceeded with runtime planning"
1098
+ ],
1099
+ "cpu_test": "PASS 4 processes each100 same-key put/read/group operations; spawn and fork; distinct PID dirs on16GiB /cache tmpfs",
1100
+ "files": {
1101
+ "Dockerfile": "f480abe402049d699f6289ada6836c54054882e06db5a8b06cc82b9b5aaf0675",
1102
+ "cpu-test.log": "b57cf395d84acae6ab328b86bf01946e4dfaf8764e1937c413c29a1430c70033",
1103
+ "tr3_process_cache.py": "527d66870da0312ed4ddd9a5039424b08183df847837c34314c182aaa58fcc0c",
1104
+ "build.log": "13b7168bf0dc2dcac43ad8f6f80a5054dca1bbc2386d9f0c0a9f8b25a4a1eddd",
1105
+ "test_process_cache.py": "0a82e3c8cbeb71a30a4cb406b77819f7b3ed3e061fdde28c45275d667ba8c6a1"
1106
+ }
1107
+ }
1108
+ },
1109
+ "plan_seal": {
1110
+ "algorithm": "sha256",
1111
+ "created_at": "2026-09-09T21:25:29.904418+00:00",
1112
+ "plan_sha256": "b4f88a3c399ae8e6661d14f0f29ea5625860f1075a88aafded3045c1d5a58252"
1113
+ },
1114
+ "metric_sha256": "752af6740595791f300cf47f15ef81f8ca85245d5282fc600ef26ee57eae41e6",
1115
+ "role_manifest_sha256": "b5d7e4524eb98ddfbd230a5d9a44de0dc5dbeb796c03e859898b5838e4463d14",
1116
+ "teacher_revision": "7c378d5f17dba158c4c803eff27c346dd0615660",
1117
+ "teacher_model_revision": "a6c167b62691b2bac901344b65cb651a70f53e43",
1118
+ "prior_attempts": [
1119
+ {
1120
+ "attempt": 1,
1121
+ "measured_windows": 0,
1122
+ "failure": "Triton metadata unavailable during warmup on shared fuseblk cache; retry isolates process cache on tmpfs."
1123
+ }
1124
+ ],
1125
+ "reference_comparison_sha256": "c70b2892ee69287554c8a8879a4d848b9077df3d9febcdc818c780851de600db"
1126
+ }
runtime-reference-20260909/source-parity-audit.json ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "passed",
3
+ "selected_image": "sha256:ca6b80188dce154b91f49108b7d87792d2ba6328935afc71b44d1c0e6f6a1adf",
4
+ "base_image": "sha256:1c8a10d2b21bd6ed5a7ca4a29bcc3900d29acc3ce42e1d722b9ebaa74357de3f",
5
+ "python_files_compared": {
6
+ "b12x": 416,
7
+ "vllm": 2411
8
+ },
9
+ "overlay_files": 33,
10
+ "claim": "Virtual Python source reconstruction from pinned base plus33overlays; not a fresh image GPU test."
11
+ }