nightly-grid-compute

name	nightly-grid-compute
description	Identical-state nightly A04 EnergyPlus grid-search compute benchmark on a single cold school day. Measures wall/CPU/memory vs candidate budgets (10/25/50/100/exhaustive), parallelism, anytime regret, and RL compute comparator. No RL training. No BACnet. Use when running or publishing docs/results/nightly_grid_compute. Distinct from the Dec 17-day fixed-policy grid-search-comparator pack.

Nightly A04 grid-search compute benchmark

Claim: SIMULATION-ONLY RESEARCH · RETROSPECTIVE WEATHER BENCHMARK · NO BACNET.

Not the Dec 15–31 exhaustive fixed-policy screen (skills/grid-search-comparator). This experiment proves identical-lookback branching for one target day.

Hard rules

• Do not train PPO/DQN
• Do not issue BACnet commands
• Do not modify A04 physics / create A05
• Do not soften 15/30 minute deadlines after seeing results
• Selection = lowest-cost fully-ready candidate (not RL training reward)
• Keep flat / TOU / dynamic tariff rankings separate

Required wording:

Grid search and RL share the same EnergyPlus trajectories, tariff accounting, and readiness criteria. RL trains on a shaped numerical reward, while grid search selects the lowest-cost fully-ready candidate.

Contract / CLI

• Contract: contracts/nightly_grid_compute_v1.json
• Primary day: 2026-01-26 (lookback 2026-01-25)
• Deadlines: 15 min target / 30 min hard

py -3.12 scripts/vibe22_nightly_grid_compute.py --site-root $env:SITE_ROOT --stage freeze
py -3.12 scripts/vibe22_nightly_grid_compute.py --site-root $env:SITE_ROOT --site-run-dir <DIR> --stage all --resume
py -3.12 scripts/vibe22_nightly_grid_compute.py --site-root $env:SITE_ROOT --site-run-dir <DIR> --stage publish --resume

Stages: freeze → baseline → micro → pilot (×3 + 1/2/4 workers) → budgets → winner determinism ×3 → publish.

Finished pack (2026-08-22)

docs/results/nightly_grid_compute/

Fact	Value
Verdict	NIGHTLY_GRID_FEASIBLE_WITHIN_15_MIN
Recommended budget	25
Unique candidates	130
Exhaustive wall	~321 s
Winner (FLAT/TOU/DYNAMIC)	discrete_114
Midnight proof	Δ = 0°F (tol 0.05)
Determinism	OK (3 matching trajectory hashes)
BACnet	0

Spec

../../vibe22_agent_spec/NIGHTLY_GRID_COMPUTE.md

Related

• Historical Dec screen: ../grid-search-comparator/SKILL.md
• Two-month replay: ../two-month-policy-replay/SKILL.md