Review and Sources Record
Report: TA-TR-2026-17
Version: 1.0
DOI: 10.5281/zenodo.22950904
Publication posture
- Theoretical/computational preprint; not peer reviewed.
- Human author of record and responsible depositor: Hongju Liu.
- Substantial AI assistance disclosed.
- The paper is non-amending adjacent research and is not independent corroboration of the Trinity Accord.
- DOI registration, OTS and Arweave preservation do not certify truth, originality, peer review or Google Scholar indexing.
Evidence hierarchy
The strongest claims are formal theorems and analytical constructions. Synthetic oracle-proxy studies are sufficiency checks. Intervention-derived and recurrent studies provide controlled constructive demonstrations. The end-to-end process-aware long-horizon advantage is explicitly not established after eight-seed controller replication; this failed replication is preserved in the paper.
Replication added before publication
- Six controller seeds support learnability of contextual-bandit persistence-operation selection in the controlled task.
- Three independently trained GRU-plus-merger stacks reproduce the core redundancy/complementarity/stale-restore/merge lineage structure.
- Eight controller seeds do not reproduce a reliable process-aware long-horizon return advantage.
Originality boundary
The manuscript does not claim that causal states, random sets, total correlation, Möbius transforms, causal representation learning, reinforcement learning, checkpoint/fork/restore/merge operations, or personal-identity fission problems are individually new. Its narrower contribution is their combination into a set-valued, representation-disciplined and current-process-grounded functional-continuity framework with exact decision consequences and controlled persistence-agent benchmarks.
See the manuscript references for the complete bibliography.