check(APRProof) → (ProofStatus, APRProof, [remainder], [undecided_obligation]).
The prover is slow because at every step it searches for applicable rules and calls SMT to decide which branches are feasible; re-establishing a proof (after a semantics change, or to trust one from elsewhere) re-pays all of it, though the finished proof already records which rules fire and where it branches.
The Checker instead treats the KCFG as a certificate and validates every element by its kind: edges are re-applied by the Haskell backend (force-apply the recorded rules, no search), splits and covers are discharged by SMT in pyk.
Because the fired rule is known, there is no rule-search SMT and the taken branch is feasible by its concrete witness; only the irreducible SMT remains — split coverage (parent ∧ ¬⋁cᵢ UNSAT) and cover subsumption (implies) — and even that moves off the critical path, since the rule sequence is fixed: rewrite streams ahead while SMT runs concurrently.
All soundness lives here.
Goal. ≥~2× faster than the prover, or reusing a proof / adding concolic execution is not worth it — measured once a first APRProofChecker exists, not before.
What is checked, and how
- Edge
source --[τ]--> target, a rule sequence τ = [r₁, …, rₙ] — Haskell backend.
guided_execute force-applies τ to source and returns the resulting end state target'; it is agnostic to any previously recorded target.
If a step branches before τ is exhausted, the edge hides an uncovered case and the check emits a remainder.
- Split
parent into branches with conditions c₁, …, cₖ — SMT.
The branches must be exhaustive and non-vacuous: parent ∧ ¬(c₁ ∨ … ∨ cₖ) is UNSAT (nothing falls through the cracks) and each parent ∧ cᵢ is SAT.
A satisfiable parent ∧ ¬⋁cᵢ is a coverage hole, and its model is a remainder seed.
- Cover
node ⟹ target (or an ancestor, under a substitution) — SMT, implies.
This closes the final target and loops; a BMC depth bound is the backstop when no cover is found.
- Leaf — it must subsume the target (success) or be a known terminal/bounded node; a stuck leaf that subsumes neither is a
Failed counterexample.
Any of these SMT checks may return unknown/timeout → an UndecidedObligation; it never silently passes.
Layers
- Haskell backend —
guided_execute(S, τ) (new RPC).
Force-apply rules from τ to S (no rule search), folding steps into a single Edge; cheap simplification discharges trivial guards.
At the first step whose guard it cannot cheaply discharge, it solves C ∧ ¬guard in place: UNSAT → fold and continue (no round-trip); SAT → return the Edge (depth, end state S'), the taken branch's guard (the Split condition), and the remainder's model (the ¬guard seed); unknown → return an obligation; the trace end also returns.
It uses SMT for these remainder checks but never for rule search, and can overlap rewriting along the fixed τ with the remainder solve.
Checking the taken branch's feasibility is an optional parameter — skip it when a concrete witness already establishes it (explore mode), enable it when re-validating a proof without one.
- pyk — the driver.
Drive guided_execute and record its Edges, Splits, and remainder seeds; check aggregate split coverage (parent ∧ ¬⋁cᵢ) and cover subsumption via implies; maintain the KCFG and accumulate the result tuple and the monotone ProofStatus.
Reuses the existing implies, get_model, and simplify RPCs.
Outputs
ProofStatus — Passed / Pending / Failed.
- updated
APRProof — the validated (and, when building, grown) KCFG.
[remainder] — coverage holes still open.
Each is remainder = (depth, model, obligation): depth locates the node; model is a concrete seed for the uncovered sub-region (set when the residual is SAT); obligation is set instead when the solver could not decide.
[undecided_obligation] —
UndecidedObligation:
source_node_id : int # node whose check spawned the query (referenced, not owned)
depth : int # rewrite steps from source_node_id to the query
kind : edge | split-coverage | cover
rule_id : str | None
query : Pattern # the term the solver could not decide
result : unknown | timeout
Passed ⟺ remainders == [] and undecided_obligations == [] and every leaf subsumes the target.
check(APRProof) → (ProofStatus, APRProof, [remainder], [undecided_obligation]).The prover is slow because at every step it searches for applicable rules and calls SMT to decide which branches are feasible; re-establishing a proof (after a semantics change, or to trust one from elsewhere) re-pays all of it, though the finished proof already records which rules fire and where it branches.
The Checker instead treats the KCFG as a certificate and validates every element by its kind: edges are re-applied by the Haskell backend (force-apply the recorded rules, no search), splits and covers are discharged by SMT in pyk.
Because the fired rule is known, there is no rule-search SMT and the taken branch is feasible by its concrete witness; only the irreducible SMT remains — split coverage (
parent ∧ ¬⋁cᵢUNSAT) and cover subsumption (implies) — and even that moves off the critical path, since the rule sequence is fixed: rewrite streams ahead while SMT runs concurrently.All soundness lives here.
What is checked, and how
source --[τ]--> target, a rule sequenceτ = [r₁, …, rₙ]— Haskell backend.guided_executeforce-appliesτtosourceand returns the resulting end statetarget'; it is agnostic to any previously recorded target.If a step branches before
τis exhausted, the edge hides an uncovered case and the check emits a remainder.parentinto branches with conditionsc₁, …, cₖ— SMT.The branches must be exhaustive and non-vacuous:
parent ∧ ¬(c₁ ∨ … ∨ cₖ)is UNSAT (nothing falls through the cracks) and eachparent ∧ cᵢis SAT.A satisfiable
parent ∧ ¬⋁cᵢis a coverage hole, and its model is a remainder seed.node ⟹ target(or an ancestor, under a substitution) — SMT,implies.This closes the final target and loops; a BMC depth bound is the backstop when no cover is found.
Failedcounterexample.Any of these SMT checks may return unknown/timeout → an
UndecidedObligation; it never silently passes.Layers
guided_execute(S, τ)(new RPC).Force-apply rules from
τtoS(no rule search), folding steps into a single Edge; cheap simplification discharges trivial guards.At the first step whose guard it cannot cheaply discharge, it solves
C ∧ ¬guardin place: UNSAT → fold and continue (no round-trip); SAT → return the Edge (depth, end stateS'), the taken branch'sguard(the Split condition), and the remainder'smodel(the¬guardseed); unknown → return an obligation; the trace end also returns.It uses SMT for these remainder checks but never for rule search, and can overlap rewriting along the fixed
τwith the remainder solve.Checking the taken branch's feasibility is an optional parameter — skip it when a concrete witness already establishes it (explore mode), enable it when re-validating a proof without one.
Drive
guided_executeand record its Edges, Splits, and remainder seeds; check aggregate split coverage (parent ∧ ¬⋁cᵢ) and cover subsumption viaimplies; maintain the KCFG and accumulate the result tuple and the monotoneProofStatus.Reuses the existing
implies,get_model, andsimplifyRPCs.Outputs
ProofStatus—Passed/Pending/Failed.APRProof— the validated (and, when building, grown) KCFG.[remainder]— coverage holes still open.Each is
remainder = (depth, model, obligation):depthlocates the node;modelis a concrete seed for the uncovered sub-region (set when the residual is SAT);obligationis set instead when the solver could not decide.[undecided_obligation]—Passed ⟺ remainders == [] and undecided_obligations == [] and every leaf subsumes the target.