An LLM writes the reasoning layer ahead of time. The application serves precomputed explanations at request time; the decision logic itself is deterministic.
A bounded decision space — model selection, content classification, eligibility, routing — is handled by deterministic logic that already works. But users need to understand why they got the answer they got. Generating that explanation with a live LLM call on every request burns tokens and adds latency for reasoning that rarely changes between inputs of the same class.
Before request time, an LLM is prompted to generate explanations for each region of the decision space. The resulting explanation artifact — a JSON or YAML file mapping decision outcomes to human-readable reasoning — is versioned and served alongside the deterministic result. The decision logic runs without the LLM. The explanation layer is the precomputed artifact.
RightModel maps coding task signals to model tiers using deterministic decision logic. The PAI artifact is the precomputed explanation layer: LLM-authored reasoning for each recommendation, versioned and served alongside the result. Requests served from the artifact cost zero tokens. When a precomputed explanation can’t cover a case — novel input, edge case, user wants deeper reasoning — the user can opt into “deep analysis,” a disclosed-cost live LLM call.
The decision logic is already deterministic. The value the LLM adds is the reasoning layer — the “why,” not the “what.” Explanations change more slowly than request frequency. Coverage gaps in the explanation space are detectable.
The decision itself requires live LLM reasoning. Explanations are highly personalized or context-dependent. The explanation space is too large to precompute meaningfully.