Predicted first. Measured after.
Cupel records scores and predictions before collecting real-robot outcomes. The public track record shows reviewed quarterly aggregates from consenting customers.
Three numbers, one per product.
Did the refined data help?
The gain in task success from refined hours against the same number of random hours, measured on the buyer's own policy, with a 95% interval and the trial count.
Did the prediction hold?
How closely the world model's checkpoint ranking matched the ranking real robots measured later. Every prediction carries its world-model version.
What happened on the fleet?
Success rate, interventions and late actions by anonymised deployment cohort.
Refine pilot outcomes
No reviewed quarter has been published. A release requires outcomes from at least three consenting pilot customers.
Test accuracy
No reviewed Test quarter has been published. A release requires verified rankable results from at least three consenting customers.
What has to be true, and when.
Each phase of Cupel ends with a test. We publish the result whether it passes or fails. Dates are estimates from November 2026.
| Gate | When | Test | Pass if | Result |
|---|---|---|---|---|
| A | Week 8 | Refined footage against random hours and simple filters on SO-101 | At equal hours, ≥10 pp over random and >0 pp over simple filters | Formal result pending |
| B | Week 16 | A paying customer reproduces the gain on their own training stack | ≥1 customer, ≥2 paid pilots | Formal result pending |
| C | Month 10 | World-model rankings against real-robot rankings, on held-out checkpoints | Rank correlation ≥0.8 across two robots | Formal result pending |
| D | Month 18 | Cupel's runtime against plain lerobot-rollout | Fewer late actions and task success at least as high | Formal result pending |
Why the numbers can't be fitted afterwards.
- Written before the result. A prediction is stored before the real-robot trial runs, so it can't be adjusted to match.
- Append-only. Rows go to a storage bucket with Object Lock. Each file records the hash of the file before it, so a deleted or changed row breaks the chain.
- Versioned. Every row records the score version or world-model version that produced it.
- Aggregates only. Scores, trial outcomes and fleet summaries. Never footage, checkpoints or recordings.