Recursion dominates ethproofs e2e around 50% or more compared with base-layer. See benchmark ceno-reth.

Ceno recursion sticks to in-house fork openvm v1.4.1 which is quite old and prover performance degrades compared with the multivariate version (ceno base layer).

OpenVM is WIP upgrading to multivariate develop-v2. To boost ceno recursion performance, we upgrade recursion to openvm v2. Once completed, we can plan to swap the recursion backend proving to ceno as well.

Implementation lives in ceno_recursion_v2/. Detailed plan: ceno_recursion_v2/PLAN.md.

Architecture

The v2 recursive verifier is a multi-AIR STARK circuit (25+ AIRs). Each protocol step is its own module with typed buses. The openvm v2 prover generates the proof; Ceno owns the AIR constraints and trace generation.

proof-shape → GKR tower → main sumcheck → batch constraint → jagged reduction → basefold

Phases

Phase 0: Integration stack (current priority)

Build end-to-end test: read proofs → preflight → tracegen → prove → verify.
All modules registered (placeholder traces OK). Log proving time + proof size.

• Enable all modules in system orchestration
• Every AIR registered and contributes a trace
• Integration test passes with 1-proof and 2-proof inputs
• Performance logging (proving time, proof size, per-AIR trace heights)

Phase 1: Complete existing AIR wiring

Make AIR traces real (not placeholder), module by module.

• Proof shape tracegen from preflight metadata
• Re-enable batch constraint (fix preflight/tracegen alignment)
• Wire VmPvsAir buses (public_values_bus, cached_commit_bus)
• ZKVM type bridges (convert_proof_from_zkvm, convert_vk_from_zkvm)
• Inner circuit tracegen (VerifierPvsAir, DeferralPvsAir)

Phase 2: Jagged Reduction + Basefold

PCS verification pipeline — critical path for soundness.

• Write Basefold module spec
• Implement 7 Jagged AIRs per spec
• Implement Basefold AIR(s)

Phase 3: Multi-proof aggregation

• Self-recursive prover (RecursionVk ↔ local VK bridge)
• Binary tree aggregation: leaf → internal → root
• Cross-shard connector chaining + EC sum in VmPvsAir

AIR inventory (25+)

Base layer changes

Reduce dynamism for recursion implementation:

• make num_instance size fixed as size 2 array, constrain num_instance[1] if possible
• make zkvm proof per chip to be 1 instead of vector
• main-sumcheck chip various degree => make different degree support for sumcheck