rl-poc-results-publish

name

rl-poc-results-publish

description

Publish or refresh the Vibe22 RL PoC results pack from finished research-long SITE_ROOT artifacts only. Use when writing docs/results, readiness wording, Dec billing-floor disclosure, validation-leader language, figures, 10-slide outlines, provenance counters, or draft PRs for reporting. Never trains, never launches EnergyPlus, never rewrites historical manifests, never invents process-launch counts.

RL PoC results publication (artifacts only)

Derive public claims from finished campaign exports. Prefer scripts/vibe22_publish_rl_poc_results.py over hand-editing scorecards.

When this skill applies

• Building or regenerating docs/results/**
• Explaining PRIMARY / SECONDARY research-long outcomes
• Correcting readiness language (“17/17 ready” is wrong)
• Disclosing December billing-floor gaps
• Drafting slides/figures/PRs about the PoC without re-running E+

Forbidden

• Retrain PPO/DQN or start LIVE EnergyPlus for “better” publication numbers
• Rewrite $SITE_ROOT/reports/eplus_gym/rl/**/campaign_manifest.json or historical eval.json
• Offline “corrected” scoreboards when Dec 1–14 arm facility series are missing
• Inventing actual_energyplus_process_launches / n_process_starts
• Pretending figure 4 has timestep facility/zone MAT traces
• Claiming exhaustive DQN discrete LIVE screen completed (NOT_RUN)
• Comparing absolute $ between flat and illustrative-TOU tariffs
• Saying “winner” in public markdown/figures (use validation leader)
• Claiming verified savings vs actual BAS or continuous 68/74 baseline

Authoritative inputs

Experiment	$SITE_ROOT/reports/eplus_gym/rl/ run root
PRIMARY FLAT_PLUS_DEMAND	research_long_flat_plus_demand_20260820T132506Z
SECONDARY ILLUSTRATIVE_TOU_PLUS_DEMAND	research_long_illustrative_tou_plus_demand_20260820T210304Z

Also cite (do not replace) compact leaders under docs/audits/figures/vibe22_primary_flat_plus_demand/leaders.json and docs/audits/figures/vibe22_secondary_illustrative_tou_plus_demand/leaders.json when present.

Contracts: observed_bas_incumbent_v2, school_calendar_v2, reward readiness via readiness_all_six + school_windows.

Required honesty labels (every public artifact)

• SIMULATION-ONLY RL RESEARCH
• NOT VALIDATED FOR OPERATIONAL DSM
• NO PRISTINE LOCKED TEST
• A04 IS NOT A TRANSIENT-VALIDATED PHYSICS CHAMPION
• TOU TARIFF IS ILLUSTRATIVE
• CURRENT CAMPAIGNS USED OBSERVED_BAS_INCUMBENT_V2
• NO BACNET COMMAND AUTHORITY

Readiness wording (P1)

Recompute from schedule_proof.school_occupancy_window.school_occupied:

• Validation window has 5 checked school days (Dec 15–19) and 12 unchecked non-school days (weekends + winter break Dec 20–31)
• Publish fields: checked_school_days, ready_checked_school_days, readiness_rate_checked_school_days, unchecked_non_school_days, all_validation_rows
• Canonical sentence: Ready on N/5 checked school days; 12 non-school days were not subject to the school-start readiness gate.

December billing floor (P3)

Exact disclosure when Dec 15 opening_mtd_kw = 0 for all arms:

Validation demand-cost accounting initialized the December billing floor at zero and may overstate incremental candidate demand charges.

Keep original scores. No corrected scoreboard without arm-specific Dec 1–14 series.

Counters (do not conflate)

Counter	Meaning
RL transitions	e.g. 8192 × model seeds
Validation arm-days	eval rows (11 × 17 = 187)
Train / validation calendar days	44 / 17
Actual E+ process launches	null unless recorded in manifest

Figures

Write under docs/results/figures/ with PNG + SVG + CSV sidecar + provenance:

1. cost_decomposition_by_tariff — stacked energy vs demand; annotate checked-school readiness rates
2. peak_and_readiness_tradeoff — peak vs cost; mark validation leaders + incumbent; note leaders raised peak
3. representative_daily_control_plan — schedule_proof on shared school day (prefer 2025-12-15)
4. representative_day_outcomes — daily aggregates only; label AGGREGATE_FROM_EVAL_JSON_NOT_TIMESTEP_REPLAY

No reward-as-cost plots.

Publish command

python scripts/vibe22_publish_rl_poc_results.py --site-root $env:SITE_ROOT
python -m pytest tests/test_vibe22_rl_poc_results_publish.py -q

Outputs: docs/results/vibe22_rl_poc_results.{md,json}, vibe22_rl_poc_arm_scorecard.csv, vibe22_rl_poc_provenance.json, vibe22_rl_poc_10_slide_outline.md, vibe22_rl_poc_exhaustive_discrete_screen.json (status updated by grid pack), baseline_evidence_resolution.json.

For LIVE discrete grid vs RL leaders use skill grid-search-comparator / scripts/vibe22_grid_search.py → docs/results/grid_search/.

For Dec–Jan frozen-policy replay vs utility: skill two-month-policy-replay → docs/results/two_month_policy_replay/.

For identical-state nightly compute (15/30 min): skill nightly-grid-compute → docs/results/nightly_grid_compute/ (do not conflate with Dec grid_search/).

Git: commit derived pack + publisher code only — not raw SITE_ROOT eplus_out or multi-MB untrimmed dumps. Cite SITE_ROOT run paths inside provenance JSON. docs/results/**/*.csv is allowlisted in root .gitignore.

Spec: ../../vibe22_agent_spec/RESULTS_PUBLICATION.md.