hraness commited on
Commit
359473f
·
verified ·
1 Parent(s): 7a2f52d

Publish reviewed v6 calibration aggregate

Browse files
README.md CHANGED
@@ -19,6 +19,7 @@ This archive contains recorded experiment results from [ALGAL](https://algal.com
19
  | Coding-harness pilot, September 20, 2026 | `studies/coding-harness-pilot-2026-09-20/results.json` | All 12 trials failed their task grading. Eight ALGAL execution receipts verified offline. |
20
  | Synthetic warehouse routes, September 23, 2026 | `studies/application-research-2026-09-23/results.json` | Structured facts scored six of eight exact answers; adding verified query answers scored five of eight. No revision activated. |
21
  | Synthetic record triage, September 28, 2026 | `studies/cumulative-skill-v5-2026-09-28/results.json` | One of three seed searches produced a passing program. The three-corpus comparison remains insufficient. |
 
22
 
23
  The coding pilot used Terminal-Bench 2.0, a four-attempt limit, and the `claude/sonnet/low` selector. Its [method and interpretation](https://github.com/hraness/algal/blob/0c458a95ee04a01f16265dd9825c202b75b929d9/docs/coding-harness-pilot-results.md) describe the small sample, policy selection, and unavailable token and subscription-cost attribution. Receipt verification checks recorded execution integrity; task grading measures task success. The selected policy remained the original policy. This pilot demonstrated no performance benefit.
24
 
@@ -26,6 +27,8 @@ The [warehouse-route study](https://github.com/hraness/algal/blob/0c458a95ee04a0
26
 
27
  The record-triage study used `grok-4.5` through `https://api.x.ai/v1`. Four strategies started from the same generated program that passed validation and had equal resource ceilings. Two additional corpora each exhausted eight seed-generation attempts without a passing program. Their failed attempts remain in the cost totals. Learning scores and the final evaluation of fixed programs are reported separately. ALGAL work units and model-call counts describe recorded execution; provider charges are unavailable. One completed corpus and one session per strategy do not establish a general improvement in later work.
28
 
 
 
29
  ## Files and reuse
30
 
31
  The JSON files preserve the published result records byte for byte, including recorded failures, null usage values, source identities, and limitations. Their schemas differ because the studies answer different questions. Use Python's `json` module or another JSON reader to inspect a study; this archive does not define a combined training split.
 
19
  | Coding-harness pilot, September 20, 2026 | `studies/coding-harness-pilot-2026-09-20/results.json` | All 12 trials failed their task grading. Eight ALGAL execution receipts verified offline. |
20
  | Synthetic warehouse routes, September 23, 2026 | `studies/application-research-2026-09-23/results.json` | Structured facts scored six of eight exact answers; adding verified query answers scored five of eight. No revision activated. |
21
  | Synthetic record triage, September 28, 2026 | `studies/cumulative-skill-v5-2026-09-28/results.json` | One of three seed searches produced a passing program. The three-corpus comparison remains insufficient. |
22
+ | Synthetic record triage calibration, September 28, 2026 | `studies/cumulative-skill-v6-2026-09-28/results.json` | Two of three calibration seeds passed; the preregistered gate stopped the nine-block comparison. |
23
 
24
  The coding pilot used Terminal-Bench 2.0, a four-attempt limit, and the `claude/sonnet/low` selector. Its [method and interpretation](https://github.com/hraness/algal/blob/0c458a95ee04a01f16265dd9825c202b75b929d9/docs/coding-harness-pilot-results.md) describe the small sample, policy selection, and unavailable token and subscription-cost attribution. Receipt verification checks recorded execution integrity; task grading measures task success. The selected policy remained the original policy. This pilot demonstrated no performance benefit.
25
 
 
27
 
28
  The record-triage study used `grok-4.5` through `https://api.x.ai/v1`. Four strategies started from the same generated program that passed validation and had equal resource ceilings. Two additional corpora each exhausted eight seed-generation attempts without a passing program. Their failed attempts remain in the cost totals. Learning scores and the final evaluation of fixed programs are reported separately. ALGAL work units and model-call counts describe recorded execution; provider charges are unavailable. One completed corpus and one session per strategy do not establish a general improvement in later work.
29
 
30
+ The follow-up calibration used three fresh corpora with labeled training examples and training-only feedback. Two seeds passed validation and the third exhausted eight generations, so no confirmatory arm ran. Fixed calibration tests passed 2/8 and 3/8; one recorded transport failure had uncertain external completion and missing usage. Runtime work and charged-attempt totals are reported, while provider billing remains unavailable.
31
+
32
  ## Files and reuse
33
 
34
  The JSON files preserve the published result records byte for byte, including recorded failures, null usage values, source identities, and limitations. Their schemas differ because the studies answer different questions. Use Python's `json` module or another JSON reader to inspect a study; this archive does not define a combined training split.
export-manifest.json CHANGED
@@ -1,5 +1,5 @@
1
  {
2
- "card_sha256": "fd05aae7e929c0940c3d69d809b32ceee3b9773ed67c554bf9d8ff733d1d375c",
3
  "files": [
4
  {
5
  "destination": "studies/coding-harness-pilot-2026-09-20/results.json",
@@ -16,6 +16,11 @@
16
  "sha256": "741387086751b4e7058ebc587ea3f03b6b9383a35221f36764a27e8226cb7a03",
17
  "source": "docs/cumulative-skill-results-2026-09-28.json"
18
  },
 
 
 
 
 
19
  {
20
  "destination": "LICENSE",
21
  "sha256": "74b69bf37c8f340c9c2a54d431a15218738d9c463d0e014fa6a8bb8edce4e539",
@@ -28,6 +33,6 @@
28
  "rights_status": "reviewed-existing-mit",
29
  "schema": "hraness-hf-export-v1",
30
  "scope": "Existing public experiment summaries only. No raw stores, Terminal-Bench task content, private traces, or provider accounting.",
31
- "source_commit": "57c0d8aa0ae2398119b5635edafa370af7d9072a",
32
  "source_repository": "https://github.com/hraness/algal"
33
  }
 
1
  {
2
+ "card_sha256": "409c1c54a823bc6531276aa35db1d3ff77135913f7e54a289cb16c9659e094f0",
3
  "files": [
4
  {
5
  "destination": "studies/coding-harness-pilot-2026-09-20/results.json",
 
16
  "sha256": "741387086751b4e7058ebc587ea3f03b6b9383a35221f36764a27e8226cb7a03",
17
  "source": "docs/cumulative-skill-results-2026-09-28.json"
18
  },
19
+ {
20
+ "destination": "studies/cumulative-skill-v6-2026-09-28/results.json",
21
+ "sha256": "3012fd2c66b379bcf00be126e79728f37e8a79e7655f3de676955182195bd8a2",
22
+ "source": "docs/cumulative-skill-v6-results-2026-09-28.json"
23
+ },
24
  {
25
  "destination": "LICENSE",
26
  "sha256": "74b69bf37c8f340c9c2a54d431a15218738d9c463d0e014fa6a8bb8edce4e539",
 
33
  "rights_status": "reviewed-existing-mit",
34
  "schema": "hraness-hf-export-v1",
35
  "scope": "Existing public experiment summaries only. No raw stores, Terminal-Bench task content, private traces, or provider accounting.",
36
+ "source_commit": "15d4a028847ec6e4645a9ad05cd83018f081b8fe",
37
  "source_repository": "https://github.com/hraness/algal"
38
  }
studies/cumulative-skill-v6-2026-09-28/results.json ADDED
@@ -0,0 +1,633 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "archive": {
3
+ "files": 239,
4
+ "frozenReplay": "passed",
5
+ "path": "experiments/cumulative-skill/results/v6/calibration-2026-09-28/evidence.tar.gz",
6
+ "receiptFiles": 43,
7
+ "sha256": "3d536bf21475745992461730b07a6fea7561759cdae80a5cccce0178159fff79",
8
+ "uncompressedBytes": 19484593
9
+ },
10
+ "bunVersion": "1.3.14",
11
+ "calibration": {
12
+ "decision": {
13
+ "cost": {
14
+ "admittedRuns": 72,
15
+ "effectsMissingUsage": 1,
16
+ "modelCalls": 72,
17
+ "tokensIn": null,
18
+ "tokensOut": null,
19
+ "workUnits": 2441155
20
+ },
21
+ "headroomBlocks": 2,
22
+ "proceed": false,
23
+ "reasons": [
24
+ "a calibration seed did not qualify"
25
+ ]
26
+ },
27
+ "rows": [
28
+ {
29
+ "block": "cal-a",
30
+ "fixedEvaluation": {
31
+ "cost": {
32
+ "admittedRuns": 8,
33
+ "effectsMissingUsage": 1,
34
+ "modelCalls": 8,
35
+ "tokensIn": null,
36
+ "tokensOut": null,
37
+ "workUnits": 163654
38
+ },
39
+ "counts": {
40
+ "complete": 7,
41
+ "labelsCorrect": 75,
42
+ "outputsExact": 0,
43
+ "recordsCorrect": 75,
44
+ "recordsTotal": 96,
45
+ "scorerPassed": 2,
46
+ "summariesExact": 2,
47
+ "tasks": 8
48
+ },
49
+ "head": "sha256:58f13040b248046452b53a700680073e5212585f64033468ce46a283234badcb",
50
+ "observedTasks": 8,
51
+ "outcome": "complete",
52
+ "session": "sha256:ab9394f25e38744357b5a721569d381c2a1f737c107b2a93199610c8629f818a"
53
+ },
54
+ "seed": {
55
+ "account": "sha256:f69465242a33e1baa8a7547266ba8cb2b67b1aea0ef8857e10e2e50c3e526885",
56
+ "attempts": [
57
+ {
58
+ "failureCode": null,
59
+ "generation": 1,
60
+ "manifest": "sha256:ea3396d87f5f94ca0531acf4b765276bb294cef88f0f2cab5c8f442ffcbe8ed4",
61
+ "mode": "generate",
62
+ "outcome": "complete",
63
+ "trainingBatches": {
64
+ "passed": 0,
65
+ "total": 1
66
+ },
67
+ "validationBatches": {
68
+ "passed": 0,
69
+ "total": 1
70
+ }
71
+ },
72
+ {
73
+ "failureCode": null,
74
+ "generation": 2,
75
+ "manifest": "sha256:0540b15434fb55e0140f6d7477ee9228b5fa4b7dbf2d3a5eb10b8b8a90ed897e",
76
+ "mode": "revise",
77
+ "outcome": "complete",
78
+ "trainingBatches": {
79
+ "passed": 0,
80
+ "total": 1
81
+ },
82
+ "validationBatches": {
83
+ "passed": 0,
84
+ "total": 1
85
+ }
86
+ },
87
+ {
88
+ "failureCode": null,
89
+ "generation": 3,
90
+ "manifest": "sha256:05ed6dea82a34625060a872f50645d567ad0137655c71780b889822ff734999f",
91
+ "mode": "revise",
92
+ "outcome": "complete",
93
+ "trainingBatches": {
94
+ "passed": 0,
95
+ "total": 1
96
+ },
97
+ "validationBatches": {
98
+ "passed": 0,
99
+ "total": 1
100
+ }
101
+ },
102
+ {
103
+ "failureCode": null,
104
+ "generation": 4,
105
+ "manifest": "sha256:05ed6dea82a34625060a872f50645d567ad0137655c71780b889822ff734999f",
106
+ "mode": "revise",
107
+ "outcome": "complete",
108
+ "trainingBatches": {
109
+ "passed": 0,
110
+ "total": 1
111
+ },
112
+ "validationBatches": {
113
+ "passed": 0,
114
+ "total": 1
115
+ }
116
+ },
117
+ {
118
+ "failureCode": null,
119
+ "generation": 5,
120
+ "manifest": "sha256:58f13040b248046452b53a700680073e5212585f64033468ce46a283234badcb",
121
+ "mode": "revise",
122
+ "outcome": "complete",
123
+ "trainingBatches": {
124
+ "passed": 1,
125
+ "total": 1
126
+ },
127
+ "validationBatches": {
128
+ "passed": 1,
129
+ "total": 1
130
+ }
131
+ }
132
+ ],
133
+ "cost": {
134
+ "admittedRuns": 20,
135
+ "effectsMissingUsage": 0,
136
+ "modelCalls": 20,
137
+ "tokensIn": 116528,
138
+ "tokensOut": 12413,
139
+ "workUnits": 799170
140
+ },
141
+ "generations": 5,
142
+ "manifest": "sha256:58f13040b248046452b53a700680073e5212585f64033468ce46a283234badcb",
143
+ "qualified": true,
144
+ "termination": "qualified"
145
+ }
146
+ },
147
+ {
148
+ "block": "cal-b",
149
+ "fixedEvaluation": {
150
+ "cost": {
151
+ "admittedRuns": 8,
152
+ "effectsMissingUsage": 0,
153
+ "modelCalls": 8,
154
+ "tokensIn": 28166,
155
+ "tokensOut": 980,
156
+ "workUnits": 180588
157
+ },
158
+ "counts": {
159
+ "complete": 8,
160
+ "labelsCorrect": 80,
161
+ "outputsExact": 2,
162
+ "recordsCorrect": 80,
163
+ "recordsTotal": 96,
164
+ "scorerPassed": 3,
165
+ "summariesExact": 3,
166
+ "tasks": 8
167
+ },
168
+ "head": "sha256:9ec62d06556415a1926a57f518c36c30cd7ecab815380e19ecd40a2bb9869142",
169
+ "observedTasks": 8,
170
+ "outcome": "complete",
171
+ "session": "sha256:b219fa7cbd721c8973b44c4b44d34de837e60ebf814c4de5e8385406c57d52ed"
172
+ },
173
+ "seed": {
174
+ "account": "sha256:c6ab2ad38e24f9b579afc487e4ed8e139d4cc9ca60d0b0c2eece298a269e02d6",
175
+ "attempts": [
176
+ {
177
+ "failureCode": null,
178
+ "generation": 1,
179
+ "manifest": "sha256:9ec62d06556415a1926a57f518c36c30cd7ecab815380e19ecd40a2bb9869142",
180
+ "mode": "generate",
181
+ "outcome": "complete",
182
+ "trainingBatches": {
183
+ "passed": 0,
184
+ "total": 1
185
+ },
186
+ "validationBatches": {
187
+ "passed": 1,
188
+ "total": 1
189
+ }
190
+ }
191
+ ],
192
+ "cost": {
193
+ "admittedRuns": 4,
194
+ "effectsMissingUsage": 0,
195
+ "modelCalls": 4,
196
+ "tokensIn": 17972,
197
+ "tokensOut": 2532,
198
+ "workUnits": 120545
199
+ },
200
+ "generations": 1,
201
+ "manifest": "sha256:9ec62d06556415a1926a57f518c36c30cd7ecab815380e19ecd40a2bb9869142",
202
+ "qualified": true,
203
+ "termination": "qualified"
204
+ }
205
+ },
206
+ {
207
+ "block": "cal-c",
208
+ "fixedEvaluation": null,
209
+ "seed": {
210
+ "account": "sha256:1831f76fc5e2f8c5772a7eca1b24e506a573604c0559ab872954d5a855209520",
211
+ "attempts": [
212
+ {
213
+ "failureCode": null,
214
+ "generation": 1,
215
+ "manifest": "sha256:e53ee023fe26547fc7dd02a1e5633393ef8fb4c2d012c0ed8543953d337df7a0",
216
+ "mode": "generate",
217
+ "outcome": "complete",
218
+ "trainingBatches": {
219
+ "passed": 0,
220
+ "total": 1
221
+ },
222
+ "validationBatches": {
223
+ "passed": 0,
224
+ "total": 1
225
+ }
226
+ },
227
+ {
228
+ "failureCode": null,
229
+ "generation": 2,
230
+ "manifest": "sha256:39993e34022a4f9477a91c85009902b0e9095c909824084f061541cb6fa0d8b2",
231
+ "mode": "revise",
232
+ "outcome": "complete",
233
+ "trainingBatches": {
234
+ "passed": 1,
235
+ "total": 1
236
+ },
237
+ "validationBatches": {
238
+ "passed": 0,
239
+ "total": 1
240
+ }
241
+ },
242
+ {
243
+ "failureCode": null,
244
+ "generation": 3,
245
+ "manifest": "sha256:39993e34022a4f9477a91c85009902b0e9095c909824084f061541cb6fa0d8b2",
246
+ "mode": "revise",
247
+ "outcome": "complete",
248
+ "trainingBatches": {
249
+ "passed": 1,
250
+ "total": 1
251
+ },
252
+ "validationBatches": {
253
+ "passed": 0,
254
+ "total": 1
255
+ }
256
+ },
257
+ {
258
+ "failureCode": null,
259
+ "generation": 4,
260
+ "manifest": "sha256:39993e34022a4f9477a91c85009902b0e9095c909824084f061541cb6fa0d8b2",
261
+ "mode": "revise",
262
+ "outcome": "complete",
263
+ "trainingBatches": {
264
+ "passed": 1,
265
+ "total": 1
266
+ },
267
+ "validationBatches": {
268
+ "passed": 0,
269
+ "total": 1
270
+ }
271
+ },
272
+ {
273
+ "failureCode": null,
274
+ "generation": 5,
275
+ "manifest": "sha256:39993e34022a4f9477a91c85009902b0e9095c909824084f061541cb6fa0d8b2",
276
+ "mode": "revise",
277
+ "outcome": "complete",
278
+ "trainingBatches": {
279
+ "passed": 1,
280
+ "total": 1
281
+ },
282
+ "validationBatches": {
283
+ "passed": 0,
284
+ "total": 1
285
+ }
286
+ },
287
+ {
288
+ "failureCode": null,
289
+ "generation": 6,
290
+ "manifest": "sha256:39993e34022a4f9477a91c85009902b0e9095c909824084f061541cb6fa0d8b2",
291
+ "mode": "revise",
292
+ "outcome": "complete",
293
+ "trainingBatches": {
294
+ "passed": 1,
295
+ "total": 1
296
+ },
297
+ "validationBatches": {
298
+ "passed": 0,
299
+ "total": 1
300
+ }
301
+ },
302
+ {
303
+ "failureCode": null,
304
+ "generation": 7,
305
+ "manifest": "sha256:39993e34022a4f9477a91c85009902b0e9095c909824084f061541cb6fa0d8b2",
306
+ "mode": "revise",
307
+ "outcome": "complete",
308
+ "trainingBatches": {
309
+ "passed": 1,
310
+ "total": 1
311
+ },
312
+ "validationBatches": {
313
+ "passed": 0,
314
+ "total": 1
315
+ }
316
+ },
317
+ {
318
+ "failureCode": null,
319
+ "generation": 8,
320
+ "manifest": "sha256:39993e34022a4f9477a91c85009902b0e9095c909824084f061541cb6fa0d8b2",
321
+ "mode": "revise",
322
+ "outcome": "complete",
323
+ "trainingBatches": {
324
+ "passed": 1,
325
+ "total": 1
326
+ },
327
+ "validationBatches": {
328
+ "passed": 0,
329
+ "total": 1
330
+ }
331
+ }
332
+ ],
333
+ "cost": {
334
+ "admittedRuns": 32,
335
+ "effectsMissingUsage": 0,
336
+ "modelCalls": 32,
337
+ "tokensIn": 198379,
338
+ "tokensOut": 20884,
339
+ "workUnits": 1177198
340
+ },
341
+ "generations": 8,
342
+ "manifest": null,
343
+ "qualified": false,
344
+ "termination": "generation-limit"
345
+ }
346
+ }
347
+ ]
348
+ },
349
+ "confirmatory": {
350
+ "observedBlocks": 0,
351
+ "plannedBlocks": 9,
352
+ "primaryClaim": null
353
+ },
354
+ "costDefinitions": {
355
+ "modelCalls": "Charged executor attempts; no provider billing attestation.",
356
+ "providerDollars": null,
357
+ "tokens": "Recorded usage counted per admitted run, including content-identical receipts. One missing usage makes full totals null.",
358
+ "workUnits": "Runtime work accounting, including every failed or repeated admission."
359
+ },
360
+ "dateTimeZone": "America/Puerto_Rico",
361
+ "diagnosis": {
362
+ "scope": "offline calibration cost, transport-failure, and repeated-request diagnosis; no raw prompts or responses",
363
+ "sha256": "47d7dd6d655727e59b01df1b41e604c83970997c1c71f03a114a5146907bda31"
364
+ },
365
+ "executor": {
366
+ "baseUrl": "https://api.x.ai/v1",
367
+ "model": "grok-4.5"
368
+ },
369
+ "failedSeedRepetition": {
370
+ "block": "cal-c",
371
+ "cacheEvidence": "No cached effects or memoization wrapper; content-identical receipts do not establish cache hits or fewer provider requests.",
372
+ "distinctParsedWriterResponses": 2,
373
+ "distinctWriterRequestDigests": 3,
374
+ "identicalFeedbackGenerations": [
375
+ 3,
376
+ 4,
377
+ 5,
378
+ 6,
379
+ 7,
380
+ 8
381
+ ],
382
+ "interpretation": "Later writer requests repeated already-correct training feedback while validation continued to fail. These attempts are not independent task samples.",
383
+ "trainingActualEqualsExpected": true,
384
+ "writerAdmissions": 8
385
+ },
386
+ "identity": {
387
+ "datasets": "sha256:805cea9df0da8f93ab24212dd54f10708eec42091bc614169ef607da107ac299",
388
+ "freeze": "sha256:0beca4ec2f85482adb8806506fa902702e4b36eb32a5adb166bc487e254d8a27",
389
+ "nativeResults": "sha256:e70b3166628a752021a21a85e5e4937e38581190f7feac545d7e035540372e23",
390
+ "protocol": "sha256:d46e15194616532179b99b7e5ce81ee54df65437fba81c22a89d229bfa6a67a2"
391
+ },
392
+ "limitations": [
393
+ "Two passing calibration seeds and one failed search do not establish reliable seed acquisition.",
394
+ "No optimizer or other confirmatory learning arm ran; there is no optimizer advantage or nine-block repeatability result.",
395
+ "Fresh draws from one shared synthetic template family do not represent separate domains or human-adjudicated support messages.",
396
+ "The v5 and v6 draws differ; this is not a paired causal estimate of the training-feedback change.",
397
+ "Validation selected seeds repeatedly and is not an untouched final test.",
398
+ "One transport-aborted call has uncertain external completion and missing usage; full token totals and provider charges remain unknown."
399
+ ],
400
+ "method": {
401
+ "budgets": {
402
+ "frozen": {
403
+ "attempts": 1024,
404
+ "runs": 512,
405
+ "work": 40000000
406
+ },
407
+ "learning": {
408
+ "attempts": 1024,
409
+ "runs": 512,
410
+ "work": 40000000
411
+ },
412
+ "seed": {
413
+ "attempts": 64,
414
+ "runs": 64,
415
+ "work": 8000000
416
+ }
417
+ },
418
+ "calibrationRule": {
419
+ "headroomMaxPassed": 6,
420
+ "headroomMinimumBlocks": 2,
421
+ "requiredSeeds": 3
422
+ },
423
+ "family": "synthetic record triage",
424
+ "fixedRecordsPerTask": 12,
425
+ "maximumSeedGenerations": 8,
426
+ "plannedConfirmatoryComparison": {
427
+ "aggregate": "strict-win",
428
+ "baseline": "fixed",
429
+ "corpora": [
430
+ "a",
431
+ "b",
432
+ "c"
433
+ ],
434
+ "minimumWinsPerCorpus": 2,
435
+ "primary": "optimizer",
436
+ "repetitions": 3,
437
+ "requireAllBlocks": true,
438
+ "requiredArms": [
439
+ "fixed",
440
+ "optimizer"
441
+ ],
442
+ "secondaryFailures": "report-separately"
443
+ },
444
+ "score": "(4 * exact full-record fraction + exact-summary indicator) / 5",
445
+ "scorerPassAt": 0.9,
446
+ "seedTrainingRecords": 24,
447
+ "seedValidationRecords": 12,
448
+ "taskSeeds": [
449
+ {
450
+ "id": "cal-a",
451
+ "seeds": {
452
+ "acquisition": 61000001,
453
+ "ancestor": 61000004,
454
+ "shift": 61000003,
455
+ "unseen": 61000002
456
+ },
457
+ "stage": "calibration"
458
+ },
459
+ {
460
+ "id": "cal-b",
461
+ "seeds": {
462
+ "acquisition": 61000005,
463
+ "ancestor": 61000008,
464
+ "shift": 61000007,
465
+ "unseen": 61000006
466
+ },
467
+ "stage": "calibration"
468
+ },
469
+ {
470
+ "id": "cal-c",
471
+ "seeds": {
472
+ "acquisition": 61000009,
473
+ "ancestor": 61000012,
474
+ "shift": 61000011,
475
+ "unseen": 61000010
476
+ },
477
+ "stage": "calibration"
478
+ },
479
+ {
480
+ "id": "a-r1",
481
+ "seeds": {
482
+ "acquisition": 61000013,
483
+ "ancestor": 61000016,
484
+ "shift": 61000015,
485
+ "unseen": 61000014
486
+ },
487
+ "stage": "confirmatory"
488
+ },
489
+ {
490
+ "id": "a-r2",
491
+ "seeds": {
492
+ "acquisition": 61000017,
493
+ "ancestor": 61000020,
494
+ "shift": 61000019,
495
+ "unseen": 61000018
496
+ },
497
+ "stage": "confirmatory"
498
+ },
499
+ {
500
+ "id": "a-r3",
501
+ "seeds": {
502
+ "acquisition": 61000021,
503
+ "ancestor": 61000024,
504
+ "shift": 61000023,
505
+ "unseen": 61000022
506
+ },
507
+ "stage": "confirmatory"
508
+ },
509
+ {
510
+ "id": "b-r1",
511
+ "seeds": {
512
+ "acquisition": 61000025,
513
+ "ancestor": 61000028,
514
+ "shift": 61000027,
515
+ "unseen": 61000026
516
+ },
517
+ "stage": "confirmatory"
518
+ },
519
+ {
520
+ "id": "b-r2",
521
+ "seeds": {
522
+ "acquisition": 61000029,
523
+ "ancestor": 61000032,
524
+ "shift": 61000031,
525
+ "unseen": 61000030
526
+ },
527
+ "stage": "confirmatory"
528
+ },
529
+ {
530
+ "id": "b-r3",
531
+ "seeds": {
532
+ "acquisition": 61000033,
533
+ "ancestor": 61000036,
534
+ "shift": 61000035,
535
+ "unseen": 61000034
536
+ },
537
+ "stage": "confirmatory"
538
+ },
539
+ {
540
+ "id": "c-r1",
541
+ "seeds": {
542
+ "acquisition": 61000037,
543
+ "ancestor": 61000040,
544
+ "shift": 61000039,
545
+ "unseen": 61000038
546
+ },
547
+ "stage": "confirmatory"
548
+ },
549
+ {
550
+ "id": "c-r2",
551
+ "seeds": {
552
+ "acquisition": 61000041,
553
+ "ancestor": 61000044,
554
+ "shift": 61000043,
555
+ "unseen": 61000042
556
+ },
557
+ "stage": "confirmatory"
558
+ },
559
+ {
560
+ "id": "c-r3",
561
+ "seeds": {
562
+ "acquisition": 61000045,
563
+ "ancestor": 61000048,
564
+ "shift": 61000047,
565
+ "unseen": 61000046
566
+ },
567
+ "stage": "confirmatory"
568
+ }
569
+ ],
570
+ "validationUse": "Repeated selection of the first passing candidate; validation labels, outputs, and scores excluded from writer requests.",
571
+ "writerFeedback": "Training examples and previous training output beside expected training output only.",
572
+ "wrongRowCountScore": 0
573
+ },
574
+ "partialTokenUsage": {
575
+ "completeTotalsKnown": false,
576
+ "inputTokens": 387413,
577
+ "missingUsageOccurrences": 1,
578
+ "outputTokens": 37679,
579
+ "recordedEffectOccurrencesWithUsage": 71
580
+ },
581
+ "recordedFailure": {
582
+ "block": "cal-a",
583
+ "category": "transport aborted; external completion uncertain",
584
+ "code": "BUDGET_EXHAUSTED",
585
+ "countsAsFailedTask": true,
586
+ "limit": "The record does not establish the exact abort cause, provider completion, or billing.",
587
+ "retried": false,
588
+ "taskId": "shift-03",
589
+ "usageKnown": false
590
+ },
591
+ "schema": "algal.cumulative-skill-public.v6",
592
+ "sourceCommit": "f8ddcf9dc0fb82e638c0a9405d2b0df289b09274",
593
+ "sourceRepository": "https://github.com/hraness/algal",
594
+ "status": "stopped-after-calibration",
595
+ "study": "cumulative-skill-v6-calibration",
596
+ "studyDate": "2026-09-28",
597
+ "totalRealizedCost": {
598
+ "admittedRuns": 72,
599
+ "effectsMissingUsage": 1,
600
+ "modelCalls": 72,
601
+ "tokensIn": null,
602
+ "tokensOut": null,
603
+ "workUnits": 2441155
604
+ },
605
+ "verification": {
606
+ "countScope": "Receipt files per store; copied seed receipts are counted in every store that contains them.",
607
+ "interpretation": "Recorded execution replay is separate from task correctness and provider billing.",
608
+ "offlineReplayPassed": true,
609
+ "receiptFilesChecked": 43,
610
+ "stores": [
611
+ {
612
+ "checkedReceipts": 8,
613
+ "store": "cal-a/frozen-fixed"
614
+ },
615
+ {
616
+ "checkedReceipts": 15,
617
+ "store": "cal-a/seed"
618
+ },
619
+ {
620
+ "checkedReceipts": 8,
621
+ "store": "cal-b/frozen-fixed"
622
+ },
623
+ {
624
+ "checkedReceipts": 4,
625
+ "store": "cal-b/seed"
626
+ },
627
+ {
628
+ "checkedReceipts": 8,
629
+ "store": "cal-c/seed"
630
+ }
631
+ ]
632
+ }
633
+ }