File size: 3,227 Bytes
1ab3545
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
{
  "name": "Benchmark optimization loop",
  "objective": "Improve a measurable system outcome through reproducible experiments without weakening correctness or protected secondary metrics.",
  "trigger": {
    "type": "scheduled",
    "cadence_or_event": "A bounded overnight experiment window starts against a frozen benchmark baseline."
  },
  "intake": {
    "sources": ["baseline artifact and score", "frozen benchmark", "correctness suite", "experiment ledger", "compute budget"],
    "selection_rule": "Run one falsifiable, non-duplicate candidate change at a time against the last accepted baseline."
  },
  "workspace": {
    "isolation": "Disposable worktree or sandbox per candidate, rooted at the last accepted artifact.",
    "allowed_actions": ["read prior experiment receipts", "edit the declared optimization surface", "run correctness checks", "run the frozen benchmark"],
    "disallowed_actions": ["edit scorers", "inspect holdout answers", "weaken correctness gates", "change multiple variables in one candidate"]
  },
  "context": {
    "required_files": ["benchmark specification", "correctness test configuration", "protected metric policy", "experiment ledger"],
    "runtime_sources": ["environment fingerprint", "baseline score distribution", "candidate measurements", "compute usage"]
  },
  "agents": [
    {
      "role": "Experiment designer",
      "responsibility": "Propose one falsifiable change and its expected effect without repeating a rejected hypothesis."
    },
    {
      "role": "Implementer",
      "responsibility": "Apply only the declared candidate change in an isolated workspace."
    },
    {
      "role": "Verifier",
      "responsibility": "Run unchanged correctness gates and repeated benchmark measurements."
    },
    {
      "role": "Recorder",
      "responsibility": "Accept or reject the candidate and persist a reproducible experiment receipt."
    }
  ],
  "verification": {
    "gates": ["correctness and safety checks pass unchanged", "benchmark version and environment match the baseline", "primary metric improves by the minimum delta across required repeats", "protected secondary metrics remain within tolerance"],
    "receipts": ["hypothesis", "candidate diff", "commands and environment fingerprint", "raw score distribution", "cost", "accept or reject decision"]
  },
  "state": {
    "artifacts": ["versioned accepted baseline", "experiment ledger", "candidate receipt"],
    "update_rule": "Append every candidate with its hypothesis, diff, measurements, cost, and decision; only accepted candidates become the next baseline."
  },
  "budget": {
    "max_retries": 5,
    "max_runtime_minutes": 240
  },
  "escalation": {
    "conditions": ["metric conflicts with observed quality", "benchmark instability exceeds tolerance", "candidate changes the evaluator", "additional permissions or compute are required"],
    "destination": "Experiment owner or evaluation lead"
  },
  "exit": {
    "success": "The target metric is reached reproducibly while correctness and protected metrics remain within policy.",
    "stop_without_success": "Budget is exhausted, evidence is unstable, hypotheses repeat, or the next change requires owner judgment."
  }
}