remaining work before selecting a distributed bug study (authored by agents unless marked 🧑)
assessment on 8 Oct 2026
- the existing tree covers failures, injection, simulation, fuzzing, model checking, histories, runtime checks, Rust, agents, and upgrades
- research directions provides three concrete candidates with closest work and rejection conditions
- consultation records a completed ChatGPT Extra High critique
- this is a literature map with selected deep examples
- full-paper reading and source coverage remain uneven
- no experiment or independent tool reproduction has been performed
next evidence needed
- persistence simulation, current first choice
- reproduce one faulty and fixed service pair before building an adapter
- pinned TiKV tests confirm persistence failpoints, restart checks, and message filters
- artifact evidence
- this does not establish Turmoil integration or a missing production behavior
- follow-up inspected ALICE model/exploration, CrashMonkey methodology, and pinned Turmoil storage code
- comparison identifies their distinct crash spaces
- simulator inspection rules out generic torn writes as a new mechanism
- inspect ShardStore and the chosen real storage contract against the same behavior
- record the exact post-crash state missing from the simulator
- a storage fault already represented by an existing adapter does not establish a contribution
- generated specifications
- obtain trusted requirements and traces independently of the model generator
- inspect Specula and SysMoBench artifacts before claiming a missing check
- distinguish model errors from event-mapping and instrumentation errors
- recovery after a finite fault
- reproduce a documented incident with restored dependencies and sustainable offered load
- compare with CSnake and existing fault tests
- a destructive reset can diagnose damage but cannot count as a correct repair
source and scope gaps
- fault injection marks inherited FATE, SAMC, Legolas, and CAFault summaries
- retrieve primary text before relying on those summaries for a novelty comparison
- bug class detectors records blocked ACM and other publisher sources
- snippet-derived counts remain leads rather than independently checked full-paper results
- model checking marks inherited snippets
- check language, implementation, and failure assumptions from papers or artifacts
- follow-up removed unsupported predictions about corpus cost and future citations
- research usefulness must be demonstrated rather than forecast as certain
- paused Claude workers were not resumed
- the pause is unrelated to research completeness
Last edited: