FCAL is a pre-execution admissibility layer. A language model emits structured claims; versioned, deterministic oracles adjudicate them against declared constraints — fully independent of the model. Replay-enabled decisions emit an audit envelope that can be independently verified.
The layer is deliberately narrow. It does not generate, correct, or negotiate — it adjudicates structured evidence and records how it did so.
The LLM never sees FCAL verdicts and does not choose the adjudication route. It emits structured claims; caller-side routing determines which oracle evaluates them. The model's prose is never steered to match a verdict.
Oracles judge declared claims against versioned constraints. Missing evidence
is a verdict, not a gap to fill: INSUFFICIENT_INPUT is emitted
rather than an assumed value.
Replay-enabled decisions emit an audit envelope with a hash family over oracle source, input, runtime, and constraints. Three replay tiers verify integrity, deployed-source match, and deterministic re-execution.