| name | grid-search-comparator |
| description | LIVE EnergyPlus fixed-policy discrete grid search vs finished PPO/DQN validation leaders on Lakeside A04 (Dec 15–31). Use when running micro-gate/pilot/exhaustive screens, tariff re-scoring, or publishing docs/results/grid_search. Not RL training. Not mega Phase 10 scaffold. No BACnet. For identical-state nightly compute (single day, 15/30 min deadlines) use skill nightly-grid-compute instead. |
Grid-search comparator (A04 discrete v3)
Claim: SIMULATION-ONLY retrospective validation comparison.
May produce a GRID VALIDATION LEADER, never an operational winner.
Do not
- Start PPO/DQN training
- Call
eplus_gym/mega/grid_search.py a completed experiment (scaffold-only)
- Fake daily adaptive branching inside this Dec multi-day screen
(
NOT_RUN_NO_IDENTICAL_STATE_BRANCHING_CONTRACT still applies here)
- Invent historical RL process-launch counts
- Compare absolute
$ across flat vs TOU
- Soften readiness to include non-school auto-pass
- Commit raw
eplus_out trees
- Conflate this pack with
docs/results/nightly_grid_compute/ (that is the
identical-state nightly compute benchmark)
Contract
contracts/grid_search_experiment_v1.json
CLI
python scripts/vibe22_grid_search.py --site-root $env:SITE_ROOT freeze-check
python scripts/vibe22_grid_search.py --site-root $env:SITE_ROOT micro-gate
python scripts/vibe22_grid_search.py --site-root $env:SITE_ROOT pilot
python scripts/vibe22_grid_search.py --site-root $env:SITE_ROOT fixed-policy-screen --pilot-json <pilot.json> --exhaustive
python scripts/vibe22_grid_search.py publish --screen-root <screen_root>
Ladder: micro-gate → pilot → exhaustive if projected wall ≤ 6h else preregistered bounded subset labeled BOUNDED_GRID_SCREEN_NOT_EXHAUSTIVE.
Pack
docs/results/grid_search/ — summary, scorecard, verdict, compute comparison, figures.
Nightly identical-state compute (separate)
For one-day identical-lookback branching + 15/30 min feasibility, use:
Spec (publication honesty): ../../vibe22_agent_spec/RESULTS_PUBLICATION.md
