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 is that idea as a 4B model: typed questions (choice, noul, score), a published prompt contract, and a readout from the option-code logits in one forward pass. The work is as much about measurement as training: sealed sets opened once, paired intervals, and every format measured separately.
What the evidence shows
- Calibration can be repaired without a temperature. v2 was overconfident on unfamiliar tasks; a short second stage on human-written task definitions took calibration error on 25 sealed tasks from 0.153 to 0.040.
- Near the top of the open 4B field, not ahead of it. 3= of 28 on JevBench’s public decisions, where one JevBench-targeted model is resolved ahead; 5th on held-out task types, with none resolved ahead; 15th on Public8, where the seven models resolved ahead each trained on, or cannot be cleared of, its source data.
- Precision is a design choice. 8-bit MLX stays within one decision of bf16; 4-bit costs 7.5 points on held-out task rows.
Still open
- JevBench’s hard tier stays near 0.6 for every configuration measured here, while a JevBench-targeted 4B fine-tune reaches 0.739.
- Long policy documents, multi-hop lookups, and judging between drafts are the weak families.
- On unseen task definitions, no accuracy gain over the untrained base has been shown.
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.