Reads the context
Does the model read the allowed sections before it answers, and does the answer reflect them?
Roadmap
Benchmarks
Independent results are planned but not published yet. The benchmark will measure how models read Strap context, respect what it says, and propose updates worth keeping.
leaderboard · not yet publishedPreview
What the benchmark will measure
Three questions, scored per model, once the method is published.
Respects what it says
Do boundaries, preferences, and settled decisions hold across a multi-step task?
Proposes updates worth keeping
Are proposed changes narrow, durable, and accepted rather than rejected as noise?