Source lean/Determinize/Finite/Reward/Normalize.lean
import Determinize.Finite.Machinenamespace Determinize.Finite.Reward/-- A numeric check remains even when the extracted translation is zero. -/def guard : Frame := .right .add (.number 0)/-- Split off only the outermost, already evaluated additive context. -/def splitStack : List Frame → List Frame × List Rat  | [] => ([], [])  | frame :: stack =>    let (inner, offsets) := splitStack stack    match inner, frame with    | [], .right .add (.number c) => ([], c :: offsets)    | _, _ => (frame :: inner, offsets)def normalizeStack (stack : List Frame) : Rat × List Frame :=  let (inner, offsets) := splitStack stack  (offsets.sum, inner ++ if offsets.isEmpty then [] else [guard])def normalize : State → Rat × State  | .eval expression environment stack =>    let (reward, stack) := normalizeStack stack    (reward, .eval expression environment stack)  | .deliver value stack =>    let (reward, stack) := normalizeStack stack    (reward, .deliver value stack)  | .rejected => (0, .rejected)structure Outcome where  probability : Rat  state : State  reward : Ratderiving Reprinductive Step where  | next (evidence : Evidence) (successors : List Outcome)  | returned (value : Rat)  | rejectedderiving Repr/-- One CEK transition, followed by outer-context normalization. -/def step (state : State) : Except Failure Step := do  match ← Finite.step state with  | .returned value => return .returned value  | .rejected => return .rejected  | .next evidence successors =>    return .next evidence (successors.map fun (p, state) ↦      let (r, state) := normalize state      ⟨p, state, r⟩)end Determinize.Finite.Reward