Frozen evaluators
Phase 5 MAIN, Phase 6.x oscillation, excellence hunt, characterization v4, delay/alarm probes — same code path that wrote the public claim ladder.
Audit Workbench · local · offline · GABS-aligned
The evaluation system behind every public claim.
Most filter vendors ship demos. We ship a local characterization system: frozen evaluators, pre-declared rules, multi-method overlays, industrial batteries, and red lights that stay in the manual.
Trial / NDA workbench access: Request Audit Pack · optional later: gsrf@boonmind.io
What the Workbench is
A desktop/local Python lab (not a cloud SaaS) for Gradient-Stabilized Recursive Filtering. Agents and engineers load data and run packs — they do not invent metrics by hand.
Pre-declared rules
Win criteria locked before runs. Thresholds are not relaxed after seeing scores. Failures stay on disk as JSON + Markdown reports.
Full artifact trail
metrics_summary.csv · DECISION_SUMMARY.json · all-methods PNGs · audit logs · executive summaries — per pack, per file.
What it has run (evaluation inventory)
Characterization & identity
- Step / frequency / gain plateau packs
- Excellence hunt multi-family (peak vs MAD)
- Mechanism ablations (k_return as product identity)
- Adaptive x* regime probes
- Delay & fixed-FA residual red lights
Industrial & real-public batteries
- Seven SCADA-style actuator files (+ remaps)
- 12-file adversarial Tier-1 pack (anti-complimentary)
- DAMADICS Lublin real plant valves (partial/loss published)
- Tennessee Eastman XMVs (simulated plant)
- VitalDB / phase mocks in research archive
Public surface: Evidence + Research notes. Full multi-GB experiment tree stays local / NDA — the site shows claim-safe slices and gallery plots.
Gallery — workbench multi-method overlays
Raw vs EMA family vs GSRF Practical / Reference. Captions claim-safe.
The claim ladder
Marketing may only climb what the Workbench locked. If it isn’t on the ladder, it doesn’t ship.
| Rung | Safe public wording | Unsafe wording we refuse |
|---|---|---|
| Identity | Soft thermostat / attractor stabilizer in log-space | “Universal optimal filter” / “better EMA everywhere” |
| Oscillation | Mid-band ring kill on locked packs (~+35% osc31 / ~−70% spectral sweet band) | “Always reduces oscillation on any signal” |
| Peaks | Peak compression toward normal (~38%; 49/52) | “Zero overshoot guarantee on all plants” |
| Industrial actuators | Regime-bounded synthetic batteries + real DAMADICS first look with partial/loss | “GSRF wins industrial actuators” without qualifiers |
| Tracking / delay / FA | Documented red lights (MAD, delay, fixed-FA residual) | “Best tracker” / “drop-in delayed command filter” / “alarm detector” |
| QEC path (BP preprocessing research — not a quantum computer product) | Separate ladder under NDA | Mixing LER heroes into industrial homepage |
Questions people actually ask
Snippet-ready answers. Deep numbers live on Evidence.
What is the GSRF Audit Workbench?
A local offline characterization lab: frozen evaluators, pre-declared rules, multi-method overlays, industrial batteries, DAMADICS staging. Agents load data and run packs — they do not invent metrics.
Is it a cloud SaaS?
No. Desktop/local Python. Public path: Try + gsrf-bench. Full Workbench = trial/commercial.
What industrial data has it run?
Seven synthetic SCADA-style actuators (Phase 5 MAIN), adversarial Tier-1 pack, DAMADICS Lublin real plant (2/3 partial, 1/3 loss published), TE simulated XMVs. Not “wins industrial actuators” without qualifiers.
How do public claims relate to the Workbench?
Marketing only climbs the claim ladder. If it is not locked, it does not ship. See Evidence and methodology.
Public eval vs full Workbench
Public (now)
- Try snippet — notebook math
- gsrf-bench — local CSV vs EMA metrics
- Evidence + Research — claim-safe slices
Full Workbench (trial / commercial)
- GUI + batch runners + frozen Phase packs
- Presets, NDA methodology, multi-GB experiment tree
- Partner / real-plant staging support