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 adapt 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 independent views of the neural signal agree, preserves results for each degree of freedom, and reports whether each prediction has enough support to be trusted.
The operator sees which movements deviate, by how much, and whether the measurements agree. 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 position
FALCON H1 · Human 7-DOF Reach and Grasp · public EvalAI leaderboard, observed June 2026. Sorted by held-out R² mean (↑).
| Rank | Team | Held out R² | Held in R² | Out σ | Category |
|---|---|---|---|---|---|
| 1–6 | FALCON baselines (Oracle / FSS, deep nets) | 0.44–0.63 | n/a | n/a | Not FSU |
| 7 | LM Meyer (RNN, zero-shot) | 0.40 | 0.62 | 0.14 | ZS |
| 8 | Credasis AI (VERITY FSU) | 0.29 | 0.30 | 0.07 | FSU · #1 |
| 9 | SPINT (Gradient-Free FSU) | 0.29 | 0.47 | 0.15 | FSU |
#1 in Few-Shot Unsupervised: tied with SPINT on held-out R² (0.29), VERITY ranks higher on consistency: held-out variance 0.07 vs 0.15, and near-zero drop from held-in (0.30) to held-out (0.29) where SPINT falls 0.47 → 0.29. Ranks 1–6 are organizer Oracle/FSS baselines (they see labels or supervised calibration); rank 7 is a zero-shot RNN, a different category.
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. No trial structure is exploited. Calibration adapts from each session's own held-out data, under 1 ms per timestep after warm-up.
Documented limits
| Limitation | Detail |
|---|---|
| DOF 5 (hand open/close) | R² = 0.21 at 31.6% score weight, primary bottleneck |
| Session drift | Gain correction ranges 0.2×–1.8× across sessions |
| Temporal state | Per-timestep only, no accumulated state across bins |
We report every session, including the ones where performance is weakest.
Specifications
| Model size | 0.2 MB |
| Hardware | CPU only |
| Per-timestep latency | ~7 ms |
| Normalized latency | ~0.40 |
| Neural bin size | 20 ms (50 Hz) |
| Air-gapped | Yes |
| Deep learning at inference | None |
| Calibration data used | 100% of held-out calib |
| Operator parameters | Sensitivity |
| Reproducibility | MD5-tracked |
| Channels / DOF | 176 / 7 |
| Thresholds | Single geometric constant |
Cross-domain validation
The same architectural framework operates across three validated domains: one geometry, one constant, three surfaces.
| Domain | Key result | Calibration |
|---|---|---|
| Network security | F1 = 0.957 (DDoS), 15 classes | 30 s, zero attack labels |
| Neural signals (BCI) | #1 FSU · R² 0.29 (FALCON H1) | Per-session, calibration-emergent |
| Quantum computing | 4×+ routing improvement | Live IBM QPU validation |
What we report honestly
Supported today
#1 in the Few-Shot Unsupervised category (#8 overall) on the public FALCON H1 EvalAI leaderboard, held-out R² 0.29. CPU-only inference, no deep-learning runtime, model artifact 0.2 MB, per-timestep confidence posture. Near-zero degradation from held-in (0.30) to held-out (0.29). Calibration from neural data only, no behavior labels on new sessions.
Limits we disclose
Held-out R² ties SPINT at 0.29. We rank #1 of the FSU entries on consistency, not on a wide margin. A zero-shot RNN (different category) scores 0.40. DOF 5 (hand open/close) remains the primary bottleneck at R² = 0.21. Per-timestep only, no cross-bin state.
Same engine, new surface
Quantum verification, network anomaly detection, and now intracortical decoding: one measurement philosophy across domains. The encoder changes; the commitment to reproducible evidence, transparent limits, and label-free calibration does not.