From supervised experiment to usable product
A laboratory can ask a participant to repeat known movements while a technician recalibrates a decoder. An everyday assistive device cannot depend on that ritual every morning. It has to interpret today's changed neural signals without today's labels already being available.
Few-Shot Unsupervised is the benchmark category that tests that missing capability. VERITY's result does not make a clinical claim; it demonstrates a practical path for removing one of the barriers between a controlled BCI experiment and a marketable neural prosthetic.
The product test is simple: can the interface calibrate to the person today, without asking the person to help retrain it first?
What is different
The system does not hide everything inside one aggregate score. It measures whether the evidence for each prediction is strong enough to trust, and reports a confidence posture for every timestep.
The operator sees the predicted movement and how strongly the evidence supports it. Not just “the arm moved here”, but “the arm moved here, and this is how strongly the evidence supports that prediction.”
A system that always produces an answer is not a verification system.
Public leaderboard snapshot
FALCON H1 · Human 7-DOF Reach and Grasp · public EvalAI leaderboard. Snapshot observed 24 September 2026, not the current standing. Sorted by held-out R² mean (↑). Ranks 5 and 6 sit between OMEN and ABACUS.
| Rank on 24 Sep | Team | Held out R² | Held in R² | Out σ | Category |
|---|---|---|---|---|---|
| 1 | Credasis (VERITY BCI FSU, CPU-only) | 0.68 | 0.77 | 0.10 | FSU |
| 2 | NDT2 Multi Oracle (FALCON baseline) | 0.63 | 0.68 | 0.08 | Oracle |
| 3 | TaskFuse FSS | 0.60 | 0.67 | 0.11 | FSS |
| 4 | OMEN FSS | 0.60 | 0.68 | 0.09 | FSS |
| 7 | ABACUS (FSU) | 0.57 | 0.66 | 0.10 | FSU |
Current standing, Oct. 29, 2026: VERITY is #1 on FALCON H1. Held-out R² is 0.71, ahead of the NDT2 Multi Oracle baseline at 0.63. Held-in R² is 0.77. The entry is Few-Shot Unsupervised. EvalAI normalized latency for this submission is 1.59. That figure is container time on the evaluator, not on-device speed. Rankings change.
What FALCON measures
FALCON (Few-shot Algorithms for Consistent Neural decoding) is a community benchmark for intracortical brain-computer interfaces. Decoders trained on early sessions often fail on new days when neural signals drift. FALCON evaluates how well adaptation methods restore performance with only a small amount of calibration data on held-out days.
The H1 task uses 176-channel Utah-array recordings with 7 degrees-of-freedom movement decoding. Few-Shot Unsupervised adaptation uses each session's neural activity without movement labels.
Documented limits
| Limitation | Detail |
|---|---|
| One participant | FALCON H1 is a single human implant, not a multi-person clinical trial |
| Not a medical device | Public-benchmark result. Not an FDA clearance or a clinical claim |
| Board latency | Normalized latency 1.59 is EvalAI container time, not on-device inference |
We report every session, including the ones where performance is weakest.
Specifications
| Leaderboard rank | #1 overall, Oct. 29, 2026 |
| Held-out / held-in R² | 0.71 / 0.77 |
| Held-out R² std | 0.08 |
| EvalAI normalized latency | 1.59 |
| Hardware | CPU only |
| On-device footprint | Under 1 MB |
| Category | Few-Shot Unsupervised |
| On-device inference | Under 6 ms / timestep |
| Neural bin size | 20 ms (50 Hz) |
| Channels / DOF | 176 / 7 |
| Air-gapped | Yes |
| Posture | Per-timestep confidence |
Same philosophy, new surface
Quantum verification, network anomaly detection, and intracortical decoding: one commitment across domains. Reproducible evidence, transparent limits, and calibration without forcing new labels on the user.