Sanitized metrics publish through the same channel as the public Lab. Raw engineering result dumps never appear in this room.
Wedge: Fast, private, sim-first physics verification for robot policies and generative world-model outputs — not a public leaderboard, not a sim platform, not a training loop.
Current proof surface: Lift + Stack + PickPlaceCan + Door tier degradation (Pillar 1), checker calibration + CogVideoX static_hover generative failure (Pillar 2). Charts render below (brand theme via @haga/brand).
| Document | Purpose | Status |
|---|---|---|
| Evidence snapshots | Curated numeric narrative of Pillar 1 / 2 results | Ready |
| Methodology report | Public methodology write-up (claim surface) | Ready |
| Technical article (public) | Distribution article — citations + commands | Ready |
| Design-partner pipeline | Outreach + demand-signal log; intake + branded report Ready | Draft (outreach unlocked; demand open) |
Dual-pillar headline (Lab-matched)
| Pillar | Artifact | Headline | Gate vs stress |
|---|---|---|---|
| 1 | Lift | Success 1.00 → 0.26; severe grasp-slip 0.74 | Mild PASS; severe FAIL |
| 1 | Stack | Success 0.96 → 0.20; severe grasp-fail 0.80 | Mild PASS; severe FAIL |
| 1 | PickPlaceCan | Success 1.00 → 0.24; severe grasp-slip 0.76 | Mild PASS; severe FAIL |
| 1 | Door | Success 0.60 → 0.52; fail rate 0.48 | Mild PASS (success-primary); curve is shallow |
| 2 | Checker | Recall 1.000 · FPR 0.000 on labeled set | Calibration before generative claims |
| 2 | CogVideoX | Flag 1.000 via static_hover (n=6, seeds 0–1) vs real 0.000 |
Multi-seed cohort — not Cosmos |
All Pillar 1 rows: n=50 per condition · Wilson 95% CIs · sim-only · scripted OSC baselines. Prefer live charts / JSON if figures diverge.