Source lean/Determinize/Proof/RewardModel/Replay.lean
import Determinize.Finite.Reward.Graphimport Determinize.Proof.FiniteModel.Replayimport Determinize.Spec.RewardModel.Modelnamespace Determinize.Finite.Rewardopen Spec.FiniteModel Checkingopen Spec.Paper (Core)def Candidate.row (c : Candidate) (i : Fin c.states.size) : Row :=  c.rows[i.val]?.getD ⟨.rejected, []⟩def Candidate.RowClaim (c : Candidate) (i : Fin c.states.size)    (action : Except Failure Step) : Prop :=  (∀ e ∈ (c.row i).edges, e.target < c.states.size ∧ 0 < e.probability) ∧  ((c.row i).edges.map Edge.probability).sum = 1 ∧  match action with  | .error _ => False  | .ok (.returned b) => (c.row i).kind = .returned b  | .ok .rejected => (c.row i).kind = .rejected  | .ok (.next _ outcomes) =>    (c.row i).kind = .transient ∧    (∀ o ∈ outcomes, 0 ≤ o.probability) ∧    ((c.row i).edges.map fun e ↦      (e.probability, c.states[e.target]?.getD .rejected, e.reward)) =    ((outcomes.filter fun o ↦ 0 < o.probability).map fun o ↦      (o.probability, o.state, o.reward))instance (c : Candidate) (i : Fin c.states.size) (action : Except Failure Step) :    Decidable (c.RowClaim i action) := by  unfold Candidate.RowClaim  split <;> infer_instancedef Candidate.RowValid (c : Candidate) (i : Fin c.states.size) : Prop :=  c.RowClaim i (Reward.step c.states[i])instance (c : Candidate) (i : Fin c.states.size) : Decidable (c.RowValid i) :=  inferInstanceAs (Decidable (c.RowClaim i _))/-- This checks local normalized-machine replay, not source-semantic correspondence. -/def Candidate.ReplayValid (c : Candidate) (source : Core) (subject : Subject) : Prop :=  c.rows.size = c.states.size ∧  c.initial < c.states.size ∧  c.states[c.initial]? = some (Finite.initialState source subject) ∧  Proof.FiniteModel.Binding.Scoped 0 source ∧  (∀ i : Fin c.states.size, Reward.normalize c.states[i] = (0, c.states[i])) ∧  (∀ i : Fin c.states.size, c.RowValid i)instance (c : Candidate) (source : Core) (subject : Subject) :    Decidable (c.ReplayValid source subject) := inferInstanceAs (Decidable (_ ∧ _))abbrev Candidate.modelEdges (c : Candidate) {source : Core} {subject : Subject}    (valid : c.ReplayValid source subject) (i : Fin c.states.size) :    List (Spec.RewardModel.Edge c.states.size) :=  (c.row i).edges.attach.map fun ⟨e, he⟩ ↦    ⟨⟨e.target, ((valid.2.2.2.2.2 i).1 e he).1⟩, e.probability, e.reward⟩abbrev Candidate.toModel (c : Candidate) {source : Core} {subject : Subject}    (valid : c.ReplayValid source subject) : Spec.RewardModel.Model where  size := c.states.size  initial := ⟨c.initial, valid.2.1⟩  kind := fun i ↦ (c.row i).kind  edges := c.modelEdges valid  nonnegative := by    intro i e he    obtain ⟨⟨original, member⟩, _, rfl⟩ := List.mem_map.mp he    exact le_of_lt ((valid.2.2.2.2.2 i).1 original member).2  normalized := by    intro i    simpa [Candidate.modelEdges, List.map_map] using (valid.2.2.2.2.2 i).2.1end Determinize.Finite.Reward