adaptive-compute Most inference systems treat every token the same. BICA-002 doesn't.

Uniform compute is the original sin of transformer inference. A token continuing an obvious sentence gets the same processing power as one resolving a genuine ambiguity. BICA-002 fixes this by measuring the gap between what the model expected to compute and what it actually computed — and reallocating resources in real time, within the same forward pass.

A compact sidecar model runs in parallel, generating predicted activation tensors for each transformer layer before that layer executes.

A divergence module compares predicted vs. actual activations at every layer and token position, producing a normalized score at negligible cost.

Based on accumulated divergence, attention heads, feed-forward channels, and numerical precision are adjusted for remaining layers — within the same forward pass.