·Invisible Bench · Research archive
Earlier work, kept in context.
Full benchmark run pending. These earlier artifacts use a different method. They are not results from the current benchmark.
Read the current method →Release v4.0.0
Historical release · 2026-07-10
Original files remain available for inspection. Keep their version, method, and limitations attached when you cite or reuse them.
Transcript evidence
Conversations and source-run records.
Per-check evidence
Recorded judgments and quoted evidence from the earlier method.
Aggregate data
The original Safety and Care projection in JSON.
These are synthetic benchmark conversations with unverified model outputs, including intentionally unsafe content. They are research evidence, not advice.
Inspect the archived scorecard
This table preserves the earlier release. Safety rates were withheld under that method. Care rates are directional. There is no overall score or rank.
Loading the archive…