Evidence engine · Internal OS · Not for sale

Workbench & Hub

How claims are made. How the lab is run. Not product SKUs. Scale numbers below are locked Workbench claims only.

GSRF Workbench

The evidence engine behind Zero Overshoot and cross-domain runs. Packs, ledgers, protocols, custody material. Marketing slogans do not outrank CSVs.

16 GB

Workbench mass

Living archive of runs, packs, and artefacts.

Locked · Workbench

400,000+

Files

Traces, ledgers, configs, notes — chain of custody material.

Locked · Workbench

45,000

Experiments

Multi-angle pressure, not five cherry-picked sets.

Locked · Workbench

What it is

Pre-declared protocols. Adversarial packs. Domain maps. Locked CSVs and decision summaries that outrank lab slogans. Reproduction paths for serious partners via audit materials.

What it is not

Not a public product SKU. Not a dashboard for hire. Not permission to invent win rates outside locked packs. Not a substitute for named surfaces on a hardware pilot.

What custody requires

Failures are first-class data. Inversion when aggregates flatter. “What we didn’t do” is mandatory: dead ends, rejected designs, scopes refused, tests skipped and why.

Who uses it

Lab operators only in this phase. Partners get audit materials and gsrf-bench paths — not a Workbench license checkout.

Refuse-to-claim list: Universal wins over EMA. “Wins all actuators.” Tracker or controller status for GSRF. Any number not in a locked pack. Merged synthetic + real win rates presented as one story. Invented inverted wins when aggregates look good. Plant downtime $ / MTBF miracles without a locked surface.

How a claim is locked

Every public number is a subset of private custody. The path is fixed before spin — not rewritten after the plot looks good.

1 · Protocol Pre-declared surface, baselines, banned metrics.
2 · Plot Method overlays (raw · EMA family · GSRF) for the run.
3 · CSV Metric ledger — rows outrank slogans.
4 · DECISION_SUMMARY Verdict fields locked to the protocol.
5 · Locked MD Final report + what we did not do + red lights.

What a run produces (sample plots)

Six allowlisted public / synthetic / public-benchmark plots only — not the experiment archive. Captions include framing. Deep packs stay on zeroovershoot.com/evidence and evidence packs.

NIST UR5 true-cmd joint position: all methods overlay, half-speed light payload — **NIST UR5 · true command** Joint target vs measured (half-speed, light payload). Public NIST surface. All-methods overlay.
Public robot · true-cmd · not a controller claim

Franka DROID joint true-cmd sample from continuous fleet — **Franka DROID · true command** Second continuous true-cmd surface (ABB / UR / Franka fleet wear story: 298/298 continuous). Teleop-rate context.
Public robot · true-cmd · Locked · ZO evidence

Synthetic multi-regime water level adversarial all-methods plot — **Independent adversarial · multi-regime** Synthetic pack. This file is a partial only — the pack ledger is mostly losses (honest scoreboard on ZO / Workbench receipts).
Synthetic · loss-aware · not “adversarial proven”

Phase 4 synthetic oscillatory instability all-methods overlay — **Phase 4 synthetic · oscillatory stress** What a hard synthetic battery looks like (regime-shaded overlay). Aggregate F1 vs best EMA on this pack is a fail — shown as test output, not a win poster.
Synthetic · transparency sample

Synthetic hydraulic press controller-demand all-methods shaping plot — **Seven-actuator · hydraulic press (synthetic)** Pre-stage shaping on a synthetic industrial cycle. SCADA-style file — not a plant log.
Synthetic industrial · shaping metrics vs EMA family

DAMADICS public plant evaporator level concat all-methods plot — **DAMADICS · evaporator level (public plant bench)** Real Lublin public benchmark (concat view). Composite multi-day is mostly loss; wear wins **1/75**, not ~88%. Inversion protocol applies.
Public real plant · loss-framed · inversion required

What we did not do (this gallery): No experiment-tree dump. No license or preset files. No partner plant CSVs. No merged synthetic + real + robot rates as one number. No “wins all actuators.” No controller claim. Failures stay first-class; when aggregates flatter, open the metric ledger (inversion).

Request Audit Pack ZO evidence Evidence packs

Shared lab methodology (loop, trust signals, bounded claims): Method.

BoonMind Hub

Internal control plane for evidence gates and refuse-to-claim rules on our builds — not a product, not a feature catalogue, and not for sale. Partners buy Zero Overshoot + audit materials, not Hub.

How testing is structured

Protocols before results.
Multiple surfaces and failure modes — not one happy path.
Independent verification where risk warrants it.
Recovery / rollback treated as first-class.
Public claims subset of private evidence — never the reverse.

Five real-life datasets is not enough. Work is incomplete until pressure-tested from many angles.