Independent scholarship · Open accessThe monthly digest

Editorially reviewed

Replay Certificates for Cloud-Trained Branching Agents

We specify replay certificates for branching-agent training and prove prefix stability, translation refinement, score-drift bounds, and retry-safe checkpoint publication. A Lean compiler and Torch auditor regenerate 16 CPU episodes, reproduce two optimizer checkpoint transitions, and detect altered scores and rewards.

Anonymous← All research

Main result

Branching agent training needs a record of which decisions were sampled, which latent cache slices were consumed, and which optimizer version accepted the result. A replay certificate keeps the transition log, typed numerical plan, action ledger, behavior profile, and update membership evidence. Its output states separately whether structural validation, numerical replay, seeded regeneration, and checkpoint transition verification passed.

Methods and results

An accepted append-only event prefix preserves its emitted node descriptors and sampled occurrence IDs through every legal continuation, including rewinds. The proof uses composition of the deterministic transition list and the fact that only append can extend the permanent node strip. A port-preserving translation witness relates optimized numerical gates to their reference inputs, visibility, and action sites. Under smooth gate equality with live dependencies, topological induction gives equality of forward maps, and the reverse maps agree by differentiation.

For a recorded behavior log-probability error of magnitude at most $\delta$, its importance-ratio error is multiplicative in $[e^{-\delta},e^\delta]$. Whole-trace likelihood ratios accumulate these bounds across action occurrences. The paper bounds the resulting change to clipped gains and distinguishes that check from full-policy or pullback equivalence. It also proves exact-once contribution reduction by logical-key union and gives a conditional commit protocol with a linearizable version head.

The Lean 4 compiler and Torch auditor were exercised on two complete CPU training steps with a hybrid model. The auditor regenerated 16 seeded episodes and reproduced both checkpoint transitions, including model weights and optimizer state. Eight of 69 sampled actions were outside final private paths. A changed behavior probability passed structural checking and failed numerical replay. A changed reward passed numerical replay and failed seeded regeneration. The evidence is for the reference workload.

Artifacts and scope

The full manuscript, Lean source, Torch reference, auditor, mutation tests, event records, and both toy checkpoints are included. The archive contains a cloud job template. Distributed commit and cross-host reverse reduction are proved under explicit contracts. Networked rank execution and CUDA qualification remain target-cluster experiments.

The compiler and synthetic workload are shared with a separate mathematical manuscript extending Octopus. The contributions here concern the audit protocol, refinement and recovery proofs, exact checkpoint replay, and validation evidence.

Authorship

Prepared for anonymous review. The responsible author or group is to be confirmed in the submission record. An AI assistant contributed to drafting, implementation, and artifact preparation.

References

Associated artifacts

Discussion

Open a careful, constructive conversation about this note.

Comments appear after editorial approval.