Observed arrival · 2026-10-19
Ratchet puts agent skill edits through regression tests
Ratchet replays approved tasks in sandboxes whenever an agent skill changes, then uses a tiered judge cascade to block regressions.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- Teams evolving agents through editable skills
- Worth noticing
- Its 35-edit pilot reports 0.80 detection recall, while seven of eight cosmetic edits were held for human review.
Field notes
Each edit is replayed against approved tasks plus three canaries, using fresh sandboxes, recorded side-effect cassettes, and the same model tier. The page says Level 0 checks side effects without a model; Nano, Super, and Ultra escalate disputed comparisons, with uncertain verdicts held for a person. In a 35-labeled-edit pilot, it reports $0.0384 mean cost per verdict; cosmetic edits were often held for review.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
⊠LoginAccess appeared gated
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue