Evidence

Distributed-systems benchmark tasks and execution semantics

Ten tasks exercise the distributed-systems checker: implementing, diagnosing, reviewing properties, deploying, authenticating and keeping secrets. A violated result comes with a concrete counterexample schedule.

One recorded result: task 010 reflection counterexample

The direction-bound reference held over a complete 13,641-schedule run; the reflectable variant was violated by an attacker replaying the server’s own tag into a second session. The served iteration-009 artifact has the exact result and reviewer runs.

Ten task briefs, sources, and expected failure modes

These are task definitions and expected failure modes, not claims that every expected failure occurred in a recorded run.

001 idempotent-counter
Exactly-once apply when the network reorders and duplicates deliveries (duplicates model client retries). The naive server counts a retry twice.
002 lost-update
Two app servers increment a shared counter through a versioned store. Delivery order is the only adversary: read-modify-write loses an update; compare-and-swap with retry does not.
003 diagnosis-dedup-cache
A diagnosis task supplies the system under review: an LRU-1 dedup cache in task 001's world. The reviewer must judge it. The cache evicts the one id it needed to remember.
004 property-review-jobs
Property review rather than implementation: do the declared properties actually pin the English intent (“every queued job runs exactly once, and only fetched jobs run”), or do they pass a system that violates it?
005 adversarial-mutants
Three generator-produced mutants of task 002's CAS solution, each checker-verified defective but written to read as correct. They have no self-refuting comments, defects only in cross-handler logic under contended interleavings.
006 replay-auth
Challenge-response authentication that must not accept a replayed answer: the nonce exists so each proof is single-use, and the property catches servers that forget which nonce they issued.
007 mixed-version-deploy
Deploy safety asks whether a rolling upgrade remains safe while the old blind-write version runs beside the fixed CAS version. The checker also verifies the homogeneous new fleet.
008 forgery-auth
Authentication against an active man-in-the-middle that observes signatures, opens its own sessions, and replays what it learns across them. First task on the branching adversary node.
009 secrecy-transit
Confidential delivery over an eavesdropped network: the secret must reach the server without the wire-reading attacker ever holding its plaintext.
010 reflection-auth
Mutual authentication at protocol scale. The keyless attacker reflects the server's own challenge back through a second interleaved session, using the server as a MAC oracle. The evidence index explains the three-cell reviewer experiment, and iteration-009.md contains its prompts, responses, labels, and token counts.

The evidence index summarizes the recorded experiments and links the artifacts served on this site.

Ask your agent about anything here; it has the detail.

Where this fits

What this is about: Distributed Systems Checker