One Pass, Many Verdicts: Parallel Neural Decisions Across Four Signals

Brain activity is recorded through physically different windows: single-unit spikes, local field potentials, regional firing proportions and calcium fluorescence. It is interrogated with equally heterogeneous questions: which way will the animal choose, how fast will it react, was it rewarded, is it engaged. Today every (signal, question, session) triple receives its own decoder. We introduce Brainllax, a parallel decision model for neural data that rebuilds the underlying method of the recently released Jev decision model and replaces this patchwork with a single interface. It takes neural activity from any subset of the four signals together with a batch of N typed judgement queries (categorical choices, ordinal scores and Booleans), and encodes a global neural state with a Perceiver-style encoder. Query-specific cross-attention and type-matched softmax, rank-consistent ordinal and sigmoid heads then return all N probability distributions in one non-autoregressive pass. We evaluate it on 39 Neuropixels sessions (10 mice, 73 brain regions) and brain-wide two-photon imaging, spanning 14 behavioural queries. One 12.7M-parameter model reaches mean Cohen's κ = 0.660, within 0.020 of 379 separately tuned per-session decoders, and a three-seed ensemble surpasses them (0.683) with the lowest calibration error of all methods. Brainllax outperforms single-task transformers and pooled multi-task heads, answers from any of the 15 signal combinations with one set of weights, answers 64 queries 29.2× faster than a comparably accurate but worse-calibrated autoregressive decoder, and adapts to unseen animals by re-learning only channel identities. The entire study ran in under four hours of wall-clock time on one two-GPU workstation.