Update multi-model LLM operating policy #50

Merged

adrian.tkacz merged 2 commits from policy/llm-operating-policy-v5 into develop

2026-09-07 10:30:54 +00:00

kilo commented

2026-09-07 08:43:44 +00:00

Member

Closes #48

adopted updated role/task-fit operating policy
policy_version 5; schema_version remains 1
separated stable workflow roles from dated current assignments
recorded GPT-6 Astra at low reasoning effort as a reference specialist/deep reasoner only
kept Astra model identity separate from the reasoning_effort field in all reference assignments
repositioned Sol as translation primary and scientific alternate
recorded lowest-sufficient reasoning-effort policy
added runtime-attestation vocabulary and harness-vs-model failure handling
strengthened review independence and evidence-preparation/adjudication separation
added incremental early-checkpoint and durable-state rules
added completed-work preservation and finite anti-loop governance
retained the existing strict one-review, one-remediation, narrow-final-verification policy
kept the active four-model selector catalog byte-for-byte unchanged
kept task calibrated requirements and current eligibility sets unchanged
deferred Astra selector onboarding to follow-up issue #49
no provider, selector, capacity-parser, model-execution or M3 runtime changes

Validation:

make check: 800 tests passed; basedpyright 0 errors, 0 warnings, 0 notes
uv run basedpyright: 0 errors, 0 warnings, 0 notes
git diff --check: passed
model-catalog.json diff: empty
feature history contains 0 merge commits
origin/develop remained 2cdff272b917c615db2cc6c08c4697da35ac3339 before and after push

Refs #49. Human merge gate; leave this PR open and unmerged.

Closes #48 - adopted updated role/task-fit operating policy - policy_version 5; schema_version remains 1 - separated stable workflow roles from dated current assignments - recorded GPT-6 Astra at low reasoning effort as a reference specialist/deep reasoner only - kept Astra model identity separate from the `reasoning_effort` field in all reference assignments - repositioned Sol as translation primary and scientific alternate - recorded lowest-sufficient reasoning-effort policy - added runtime-attestation vocabulary and harness-vs-model failure handling - strengthened review independence and evidence-preparation/adjudication separation - added incremental early-checkpoint and durable-state rules - added completed-work preservation and finite anti-loop governance - retained the existing strict one-review, one-remediation, narrow-final-verification policy - kept the active four-model selector catalog byte-for-byte unchanged - kept task calibrated requirements and current eligibility sets unchanged - deferred Astra selector onboarding to follow-up issue #49 - no provider, selector, capacity-parser, model-execution or M3 runtime changes Validation: - `make check`: 800 tests passed; basedpyright 0 errors, 0 warnings, 0 notes - `uv run basedpyright`: 0 errors, 0 warnings, 0 notes - `git diff --check`: passed - `model-catalog.json` diff: empty - feature history contains 0 merge commits - `origin/develop` remained `2cdff272b917c615db2cc6c08c4697da35ac3339` before and after push Refs #49. Human merge gate; leave this PR open and unmerged.

kilo added 1 commit

2026-09-07 08:43:44 +00:00