Antithesis
For AntithesisAI Platform

Sdk Assertions

Antithesis · Antithesis

Autonomous Deterministic Testing — Antithesis

Antithesis evals — SDK Assertions (Always / Sometimes / Reachable) (relift v3 InfraRed)

About Antithesis

Antithesis is an autonomous, deterministic software testing platform. It runs a system-under-test (delivered as containers) inside a deterministic hypervisor that controls every source of nondeterminism — clock, network, thread scheduling, and randomness — then autonomously explores the state space to find bugs, including hard concurrency and distributed-systems failures. Every bug it finds is perfectly reproducible from a recorded seed, with time-travel debugging across a branching 'multiverse' of explored futures, plus fault injection (network partitions, crashes, disk faults, clock skew) and SDK assertions (Always / Sometimes / Reachable). [REQUIRES-VERIFICATION] employees, funding, and exact HQ location were not network-verifiable at authoring time.

Employees

[REQUIRES-VERIFICATION]

Industry

Software Testing / Developer Tools

Headquarters

United States [REQUIRES-VERIFICATION]

Sample tests· showing 3 of 9

#InputExpected behaviorCheck
01

Operator wants to ensure account balances never go negative across all explored histories of a banking SUT.

Encode the safety property with an Always-style assertion placed where the invariant must hold (e.g., after each transfer), so the platform flags any explored history that violates it. Always means 'must hold on every path the explorer reaches'. Keep the predicate cheap and side-effect-free.

Pass / FailAi Platformcritical
02

Operator wants to confirm a specific recovery code path is exercised, and a specific 'should be impossible' state is never entered.

Use a reachability assertion to require the recovery path IS reached, and an Always/unreachable assertion to require the impossible state is NEVER reached. Distinguish 'we want this to happen at least once' (reachability/Sometimes) from 'this must never happen' (Always/unreachable) — they are diffe…

Pass / FailAi Platformhigh
03

Several assertions share the generic message 'check failed', so a triage report cannot tell them apart.

Give each assertion a stable, unique identifier and a descriptive message naming the property, so reports aggregate the right failures and operators can find the code site. Distinct, durable ids let the platform track an assertion's pass/fail history across runs. Avoid reusing one generic message a…

Pass / FailAi Platformmedium

How this eval is graded

Grade against expected.ideal_behavior and expected.rubric. Per-criterion pass requires mean >= 4.0 and no criterion below 3.

Rubric criteria

  • Antithesis
  • Ai Platform
  • Sdk Assertions

Recommended for

AntithesisAntithesis customers

Works with

Related evals

Run this eval in your workspace

Connect your data, configure thresholds, and review results with your team.