Skip to content
Voidex
Sign inStart free
Voidex Bench

Measure reliability with reproducible runs — not anecdotes.

Suites of versioned cases run against freshly provisioned simulator tenants with optional fault injection, under your workspace policy. Every number comes from a recorded run, with its configuration hash.

Try it in the sandboxAPI & SDK
  1. 1Action

    Typed request from an agent, workflow or API

  2. 2Policy

    Deterministic allow / deny / require approval

  3. 3Approval

    Bound to the exact payload hash

  4. 4Execute

    Isolated worker, allowlisted egress

  5. 5Verify

    Independent read-back of real state

  6. 6Receipt

    Hash-chained evidence of what happened

A failed or inconclusive check never becomes “success”. A model’s own claim is never counted as verification.

Capabilities and current status

Suites and versioned cases

Import/export in the documented voidex.bench.v1 JSON format. Replacing cases bumps the suite version; old runs keep theirs.

Available

Deterministic fixtures

Fresh simulator tenants per run; faults such as dropped writes, lost responses and transient errors.

Simulated

Metrics

Task success vs expectation, verification outcome, policy violations, exceptions, retries, latency and estimated provider cost (0 when no model is used).

Available

Side-by-side comparison

Compare two runs case by case, with a flag when configurations differ.

Available

Model/agent comparisons

Running cases across real model providers requires provider adapters, which are planned.

Planned

Voidex does not publish claims that one model or provider is better than another. Bench reports what your configurations did on your cases.