WOMO LABS / MIRAGE
MIRAGE: A Benchmark for Enterprise World Models.
Today's AI is graded on how convincing its output looks. Looking right and being right about the world are not the same thing. MIRAGE is a coherence benchmark for world models: it tests whether a system holds one consistent picture of a situation, a state that moves under rules and reacts to intervention. Grading is exact and judge-free: every instance is minted from a private seed, its answer computed by a simulator and re-derived by an independent solver. This release publishes the protocol, the anti-contamination design, and reference baselines validating the grader; frontier model results are the next deliverable, produced by the released harness.
MIRAGE is our measurement layer, not our product.
GROUNDED REASONING
Can the model reason from what is currently true of the world, rather than from stale training priors? Scenarios are anchored to dated, verifiable states.
CAUSAL COHERENCE
Do predicted effects follow defensible causal chains? Credit requires the mechanism, not just the outcome — chains are scored edge by edge.
HALLUCINATION RESISTANCE
Unanswerable queries must fail loudly. Confident answers to unidentifiable questions are penalized more than abstentions.
MIRAGE · Publishing: full protocol, anti-contamination design, and reference baselines validating the grader. Frontier results are the next deliverable, produced by the released harness.
Methodology: counterfactual market scenarios with ground truth resolved by subsequent real-world outcomes; anti-contamination protocol; identification checked before scoring. Full protocol in the MIRAGE paper.
Evaluate your model, request access →