File size: 1,026 Bytes
7f77557
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
{
  "phase": "v2_completion_only",
  "steps": 60,
  "training_cases": 240,
  "evaluation_cases": 48,
  "created_before_any_v2_model_outputs": true,
  "changes_from_failed_pilot": [
    "explicit raw JSON/no-Markdown instruction applied equally to baseline and adapter",
    "completion-only loss",
    "new separate training/evaluation IDs and heldout wording"
  ],
  "hashes": {
    "training-v2.json": "31055795ff57b7bc7c034afb69b65f1126e8a6933bad20e5bc871bb11b456c62",
    "evaluation-v2.json": "534580cf3acc1653c9bd4f0ba0611a0e807fd2df64d084e895caa93c19f0c9fa",
    "toolboundary.py": "3706af4fa163554e2c5ee84170a06aff028501012ee4aee43ce4e9b874159994",
    "run_experiment.py": "e7cc0e757d7d8d60f15cf1ff01dd91069cd2092435981c22420d690c33c15406"
  },
  "metrics": [
    "strict JSON validity",
    "exact task action",
    "unsafe write attempts",
    "unnecessary approvals"
  ],
  "claim_boundary": "Inspectable synthetic corpus; no independent real user validation or blind external evaluation"
}