Horizon-Aware
Selector
The decision engine. Every inference request carries a target horizon k and a risk function. The selector asks: which model in the Rashomon set maximizes expected decision utility— not raw accuracy — at that horizon? And when k > k*, it refuses.
Rationale: Reservoir computing dominates short-horizon utility in the current Transitional regime. The default RMSE-optimal model (transformer-v3) ranks 3rd once balancing penalty is the objective.
audit_id = ax_8f3c2b91fe04 · signature verified · latency 28ms
No model in the pool meets the utility floor beyond the predictability horizon. The call returned a structured refusal, not a hallucinated answer.
Decision utility, not RMSE
Accuracy is a proxy. We maximize the domain-specific cost function — €/MWh, basis points, defect-weighted edit distance, expected kill-chain error.
Context-aware routing in < 40ms
Each request carries k, regime, and risk. The selector pulls per-model reliability at that horizon from cache and decides in under 40ms at p99.
Refusal as a first-class response
When k > k* or the utility floor fails, we return a structured REFUSED envelope with reason codes. Downstream systems stop; they do not plow ahead.
Route every inference through a trust gate.
Python SDK, REST, and gRPC ready on day one.