Research
Between signals and decisions.
Two connected lines of inquiry: how small models can make decisions you can check, and how neural signals can become useful on the devices that record them.
01 / Machine intelligence
Decisions, not just words.
What should happen next, and how sure are we?
Software that routes a request, checks a condition, or picks an action needs a decision, not a paragraph. A probability over a few well-defined options is easier to test, threshold, and audit than generated prose.
Manchego V3 is that idea as a 4B model: typed questions, the SemIf prompt, and an option-code readout in one forward pass. It starts again from Qwen3.5-4B, with a fresh rank-16 adaptation and versioned evaluation.
What the evidence shows
- Public comparisons improve. On the project’s matched-runtime runs, V3 scores 0.857 versus V2.1’s 0.805 on 231 public JevBench decisions, and 0.753 versus 0.732 on Public8. These are project-run comparisons, not official JevBench composite scores.
- The prompt is part of the model. V3 trained under SemIf with 2–16 options and a nonempty state. The larger-menu fallback was not its training prompt.
- Formats remain separate numeric artifacts. The 8-bit and bf16 V3 builds report matching rounded accuracy on their small conversion checks. The 4-bit build uses less memory and scores lower on both checks.
Still open
- Held-out task-type and sealed unseen-task improvements over V2.1 are not statistically resolved.
- V3 is overconfident on unfamiliar task definitions; its serving map deliberately sharpens yes/no answers.
- Fresh sealed draws of the project’s computed families favor V2.1: 0.795 versus V3’s 0.762.
02 / Neural intelligence
Closer to the neural signal.
A signal is a beginning, not an answer.
EEG is noisy, sparse, and personal. Our EEG work packages foundation models for Apple platforms and keeps three jobs apart: reconstructing a signal, representing it, and decoding a constrained intent from it.
ZUNA and ZUNA1.1 are Zyphra’s reconstruction models, converted here to Core ML. CBraMod is a criss-cross attention backbone: one package extracts its embeddings; another is fine-tuned for cue-paced left/right motor imagery on a 14-channel consumer layout.
What the evidence shows
- Ports can be faithful, and should be proven so. The motor-imagery classifier matches PyTorch on 210 of 210 real trials, with a worst logit error of 7.0 × 10⁻⁶.
- Precision limits are model-specific. Original ZUNA runs in fp16; ZUNA1.1’s bf16-trained decoder overflows it, so every ZUNA1.1 profile is fp32.
- Converters have bugs too. The ZUNA1.1 port found an epsilon-inflation bug in coremltools, reported upstream with a proposed fix.
Still open
- Live, asynchronous decoding on a consumer headset: the motor-imagery result is offline and cue-paced.
- The EPOC X layout has no electrodes over the central motor strip, which limits motor imagery from the start.
Research, with boundaries
Our EEG artifacts are not medical devices or clinically validated tools, and we do not claim unrestricted thought decoding. The public model cards remain the source of truth for capabilities, provenance, licenses, and limitations.
Follow the work
Good questions. Measured answers.
New models, their evidence, and the occasional correction. Model cards come first; this site follows.