How HALO-Vox turns eight channels of EEG into invasive-class communication speed.
Conventional brain-computer interfaces use each brain signal to select a single character or word. Every message is spelled out one symbol at a time, and effort scales directly with how long the message is. HALO-Vox uses each brain signal to do something fundamentally different. Instead of selecting a letter, it narrows down the space of possible meanings the user is trying to express. Each round of interaction reduces uncertainty about intent, not about spelling. The result is a system whose effort depends on how ambiguous the thought is, not how long the sentence is. A five-word message and a twenty-five-word message take roughly the same amount of effort to communicate.
A language model generates the set of sentences the user might mean given the current context, then clusters them by semantic similarity into a small number of distinct intent groups.
Those intent groups are partitioned across the screen into options sized so that each brain signal carries the maximum possible amount of information. The partition is computed using the Blahut-Arimoto algorithm to operate at the channel's true capacity.
The user attends to the option containing their intent. Their evoked EEG response, decoded by an 8-channel cVEP decoder, updates a Bayesian belief distribution over all candidate intents. After a few rounds, one intent dominates the posterior.
Before the system speaks anything aloud, a sequential statistical gate certifies that the false-accept probability falls below a pre-specified bound (α). Until that bound is met, the system continues to refine. Across 111 trials, zero false outputs were ever spoken.
HALO-Vox is grounded in six laws of intent resolution derived from first principles of information theory. Where traditional BCIs are bounded by Shannon channel capacity over a symbol alphabet, intent-amplifying systems are bounded by the entropy of the intent distribution itself, H(Z) — a quantity orders of magnitude smaller than the entropy of natural language at the character level. Posterior matching, originally developed by Shayevitz and Feder for capacity-achieving feedback communication, provides the mathematical machinery for selecting each query optimally. Wald's sequential probability ratio test provides the safety guarantee. Together, these tools yield a system whose verified communication rate scales not with hardware quality but with the inverse of intent uncertainty.
verified communication rate (vs. 3.18 WPM baseline)
intent amplification factor over conventional decoding
speedup in time per message (242s → 10.7s)
false outputs across 111 IRB-approved trials
total hardware cost
consumer EEG, no surgery
Cohen's d across all five primary endpoints
Bayesian evidence for the double dissociation
The most important finding is not that HALO-Vox is faster. It is that HALO-Vox and conventional BCIs operate under fundamentally different computational regimes. In conventional systems, effort scales with message length (β ≈ 0.72) and is statistically independent of intent uncertainty. In HALO-Vox, the relationship inverts: effort scales with intent entropy (β ≈ 0.45) and is statistically independent of message length, confirmed by TOST equivalence testing. Neither system tracks the other's variable. This is the signature of a paradigm shift, not an incremental improvement.
The six laws of intent resolution are not specific to brain-computer interfaces. They apply to any bandwidth-limited human-machine communication channel: EMG-controlled prosthetic limbs, surgical robotics, wheelchair controllers, and assistive devices for the elderly. Every one of these domains faces the same bottleneck conventional BCIs face — exhaustive symbol-level decoding through a constrained channel — and every one is predicted by the theory to admit the same ceiling break. HALO-Vox is the first proof point for a much larger class of systems.