Steven Hé (Sīchàng)

consensus and replication research study (authored by agents unless marked 🧑)

start here

• agent recommendation: study recovery when snapshots and membership changes overlap
  • test whether a crash during snapshot installation or voter changes restores the right data and voters
  • a snapshot saves application state; a membership change replaces the machines allowed to vote
  • smallest proposed experiment: one pinned implementation and crashes between storage operations
  • check restored data, saved voting decisions, and permitted quorums
  • check request deduplication separately if the application promises it
  • no proposed topic has established novelty
• strongest comparison to read first
  • CCF and Ellsberg for connecting protocol models to running implementations
  • Grove for proofs combining recovery, leases, and membership changes
  • XLL for shared logging and crash recovery
  • IronSpec for checking whether specifications express the intended guarantees
  • sources and qualifications in the linked studies
• merged research shortlist
  • experiment choices, closest comparisons, and remaining evidence gaps

reading tree

• crash consensus and recovery
  • protocols, voting quorums, membership changes, snapshots, and shared logs
  • source ledger
• agreement despite malicious machines
  • leaders, dissemination graphs, asynchronous agreement, early transaction results, and proactive recovery
  • source ledger
  • data availability, backlog controls, and retention follow-up
  • Byzantine testing, restart, and voter-change follow-up
  • protocol performance assumptions and checked models
• blockchains and validators
  • fork choice, finality, execution, signing history, and client diversity
  • source ledger
• verification boundaries
  • which guarantees proofs and runtime checks cover
  • three experiments about persistence, membership, and adversary assumptions
• communication between replicated groups
  • receipt, durability, application effects, and recovery
• convergent replication
  • when copies can merge updates without ordering each one first
  • deletion records and retired replicas
• extended reviews
  • crash-tolerant consensus and real recovery failures
    • library interfaces, leases, client retries, and selected 2025–2026 work
  • verification and testing of consensus code
    • proofs, checking recorded executions against models, simulation, and LLM assistance
  • validator software and incidents
    • Agave, Ethereum clients, execution determinism, and historical failure replay
• coverage and consultation record
  • reading limits, search failures, and independent review

proposed sequence

• inspect one implementation’s recovery and membership tests
  • choose a target only after confirming the needed storage events are observable
• reproduce one known failure or documented corner case
  • pin source, dependencies, storage contract, and schedule
• compare network-only traces with storage-aware traces
  • measure new coverage, false alarms, and instrumentation cost
• decide the contribution from evidence
  • a concrete defect and fix
  • a missing contract with a proof or checked model
  • a resource-cost result under an explicit adversary budget
  • a useful negative result
• move to malicious replicas or cross-group delivery after the first contract works
  • do not infer either guarantee from crash-only tests

relation to existing work in these notes

• Agave verification sizing
  • existing pilot boundaries remain distinct from these distributed-system proposals
  • no consensus effort estimate is inferred from the sanitizer estimates
• sibling studies cover distributed bug finding, storage and databases, and other distributed-system areas
  • all files created by this worker stay in this folder

status

• selected studies written on 2026-10-07 UTC
• extended reviews integrated on 2026-10-08 UTC
• the interrupted additional protocol draft was absent
  • a new bounded follow-up now covers Angelfish and the Pipes author tutorial
  • selected full proofs and artifacts remain unread
• notes pushed to the repository on 2026-10-08 UTC
  • website navigation is coordinated separately
  • live index returned HTTP 404 before navigation integration was confirmed
• requested ChatGPT opinion captured and assessed on 2026-10-08 UTC
  • primary-source corrections and consultation limits
• recent primary sources include work published or updated in 2026
• literature breadth does not establish completeness
• paper performance results are author reports
  • none independently reproduced in this study
• proposals remain hypotheses
  • no experiment or implementation defect is claimed

Last edited: 2026-10-08