What was run
- Evaluator: frozen Phase 5 MAIN (
run_phase5_main), balanced preset - Pack A: 7 industrial actuator CSVs (valve, press, turbine*, choke, boiler, damper, RO*)
- Pack B: 12 adversarial Tier-1 CSVs (independent generator — not complimentary by design)
- Total: 19 files · 0 pure losses under the evaluator’s shaping rule
*03/07 remapped: measured feedback as filter input when setpoints were constant.
Scoreboard
| Pack | Clear | Partial | Mixed | Loss |
|---|---|---|---|---|
| Seven industrial | 5 | 1 | 1 | 0 |
| Adversarial Tier-1 | 2 | 6 | 4 | 0 |
| Combined | 7 | 7 | 5 | 0 |
Shaping score = wear + churn + rate-stress + osc31 vs best EMA (≥3 clear, 2 partial, 1 mixed, 0 loss).
Identity that never went away
On these batteries, GSRF Practical’s MAD vs raw is consistently higher than best EMA. That is the spring. If your KPI is hug-the-raw, start with EMA/Kalman — see When GSRF lost.
Real plant first look (DAMADICS Lublin)
Official Lublin Sugar Factory dumps (https://iair.mchtr.pw.edu.pl/Damadics), 3 actuators × 3 days (~259k samples each), CV→demand / X→measured, same frozen Phase 5 MAIN evaluator:
- Evaporator juice level valve — partial_win (2/4)
- Juice flow valve — partial_win (2/4)
- Boiler water level valve — practical_loses_to_best_EMA (0/4)
Tennessee Eastman Mode-1 XMVs (simulated plant): three channels tested — all loss under the same rule. SWaT not run (auth). Real plant is harder than synthetic 0-loss batteries — we publish that.
Multi-day Lublin pack (25×3) is locked in Workbench — see Evidence industrial. Still open:, ISDB stiction, SWaT, partner plant.
Verdict for buyers
Use GSRF when mid-band ring / peak chase / command shaping around a normal is the costly failure — and validate on your traces (gsrf-bench). Do not buy on “always wins actuators.”
Primary ledger: Evidence §5. Citations: DAMADICS paper + dataset · eval package · GitHub. Full workbench reports local only until partner data lands.