This page points at the frozen EVO-001 acceptance matrix, baseline workload, and cost-authorization rules. It does not record live optimization results.
Frozen plan
- Checked-in mirror:
tests/fixtures/evolution/AWR-EVO-001/ - Local working copies (gitignored):
.local/awr-evolution-20260919/evaluation-plan.json,fixture-manifest.json,authorization-requirements.md
Validate:
python3 scripts/evolution/verify_evo_001_plan.py
python3 scripts/check_public_tree.py
What this freeze covers
- Fixed inputs, assertions, output identities, change axes, and result denominators for all ten check groups — sealed before optimization.
- Pre-registered performance / non-degradation thresholds, sample counts, and missing-measurement handling — thresholds must not change after seeing results.
- Independent native and paid authorization fields; simulated / protocol / native / paid evidence layers do not substitute for each other.
- Separate evaluation axes for implementation correctness, performance gain, and full-workflow incremental value.
Explicit non-goals of EVO-001
- Does not call paid models or start native agents.
- Does not start EVO-002, DEC-040, or DEC-041.
- RET-AMD-001: independent R0/R1/R2 matrix for RET-006 is frozen; the old 36-trial proposal/budget does not auto-expand.
Related public context benchmark (orthogonal; not a substitute for full-workflow value): README.md, workflow.md.
