tiltmetertiltmeter
The paneldeath-condition guard

Models

This page is deliberately not a leaderboard. No score, no rank, and no cross-suite aggregate exists anywhere in this project — not hidden, not computed and withheld. There is nothing to sort by.

Every published number is scoped to (suite, harness commit, model). A model is not good or bad here; a harness either still fires the way it did last week, or it does not. See each suite's own page for readings.

The panel this project runs against: model, its role, its cited release date, and the source
ModelPanel roleReleasedSource
Claude Haiku 4.5standing2025-10-15cited source
Claude Sonnet 5standing2026-06-30cited source
Claude Haiku 4.5null2025-10-15cited source

The null role

One panel entry runs the cheapest model a second time with identical axes. Nothing about the model or the harness differs between the pair, so whatever moves between them is noise — the negative control that bounds what "moved" is allowed to mean.

Alias substitution

No model has ever been resolved via alias substitution in this observatory. When it happens it publishes here as a labelled entry, never a silent swap (SPEC §4).