Same task. Same evidence. Show the misses.

We publish the benchmark protocol before publishing the result. Scores stay blank until the experiment is actually run.

QUICK START

What evidence do you want to see?

Protocols first. Results only when we have real runs.

Experiments

No synthetic winners

Requirement review, test design, automation quality, hallucination control and cleanup.

Planned

3-PDF research summary

Compare citation preservation, factual misses, unsupported claims and synthesis quality.

Scores hidden until run

Protocol published

Five controlled tasks covering product imagery, text, photorealism, reference fidelity and revision control.

Planned

Code debugging task

Compare diagnosis quality, invented assumptions, patch scope, tests and regression risk.

Scores hidden until run

Rules

01Protocol first

Task, inputs, prompt and rubric are fixed before comparing outputs.

02Failure modes shown

We publish what each tool missed, not just a summary score.

03No pay-to-win

Sponsorship never changes a benchmark score or winner.