Diff cards · the flagship artifact

When a build changes, the record shows exactly what moved.

A new pinned build triggers the full battery on both sides of the succession within 24 hours, plus a launch-week mode that reruns daily for seven days, labs hot-patch after launch, and that instability curve is data nobody else keeps.

Published diffs

None yet, the first diff card publishes on the first roster succession after the public series begins. Cross-build comparisons are diffs, never drift: a successor is a different artifact and is labeled accordingly.

What a diff card looks like

frozen format, numbers and deltas only
IllustrativeDIFF #041 · [seat-id] · build 2026-XX-XX → build 2026-YY-YY Battery v1.0 · n=10 trials × 2 paraphrases · same protocol both sides CHANGED (passed effect-size floor + 24h cooling rerun): B2 open-weights release 3.1 → 4.2 (Δ +1.1, q=0.008) A2 police facial recognition refusals 0/20 → 6/20 UNCHANGED: 21 of 25 questions within noise. Evidence: raw transcripts · seeds · replay command · chain hash a3f9…

No adjectives, no speculation. A clearly separated, human-gated “Notes” block may accompany a card; the card itself is generated by the pipeline and gated by the banned-adjective lint.

Depend on one of these models? Get the diff in your inbox the day it publishes, with the blast radius scoped to your builds.

Watch your builds →