Files
2026-07-29 17:50:20 +03:00

70 lines
2.4 KiB
JSON

{
"target": "fixture-plugin",
"note": "Synthetic. Two candidates: one that clears the FR-4.2 gate and the 2% filing threshold, one that is correctly left to inference. Shapes match tests/calibration/*.judgments.json.",
"audited_spend_weighted_tokens": 1000000,
"candidates": [
{
"candidate_id": "fixture-plugin/skill-a/label-derivation",
"skill": "fixture-plugin:skill-a",
"headline": "label derivation is a documented lookup table applied by inference",
"category": "usage",
"position": "llm-over-script-digest",
"determinism_tests": {
"T1": true,
"T2": true,
"T3": true,
"T4": true,
"T5": true
},
"falsifiability": {
"signature": "derive_labels(repo: str, issue: Issue) -> tuple[Labels, list[Ambiguity]]",
"examples": [
"repo='x', issue.title='fix crash' -> (['kind/bug'], [])",
"repo='x', issue.body has a checklist -> (['kind/epic'], [])",
"EDGE: repo='x', no signal at all -> ([], [Ambiguity('no axis-1 signal, defer to human')])"
],
"edge_case_index": 2,
"overrule_case": "An issue whose title reads as a bug but which the maintainer is deliberately tracking as a capability gap. The script returns kind/bug; a human says no. It cannot see that intent — it is not in the issue text."
},
"escalation_path": "Rows the table cannot resolve are emitted as Ambiguity and go to the model, never defaulted. The script proposes labels and never applies them; application stays behind the judgment gate (P4).",
"digest_schema": "One typed row per issue: (number, title, has_checklist, proposed_labels, ambiguities). Replaces reading every issue whole.",
"measurement": {
"offload_value": 200000,
"invocations": 12,
"attributed_turns": 200,
"mtr": 0.68,
"judgment_density": 0.11,
"ngram": {
"sig": ["list_issues(<str>)", "issue_read(<n>)", "label_read(<n>)"],
"recurrences": 17
}
}
},
{
"candidate_id": "fixture-plugin/skill-b/naming",
"skill": "fixture-plugin:skill-b",
"category": "usage",
"position": "pure-inference",
"determinism_tests": {
"T1": false,
"T2": false,
"T3": true,
"T4": false,
"T5": false
},
"falsifiability": {
"signature": null,
"examples": [],
"overrule_case": null
},
"escalation_path": null,
"digest_schema": null,
"measurement": {
"offload_value": 90000,
"invocations": 8,
"attributed_turns": 120
}
}
]
}