What's Changed
- feat(evals): B3a W3 — judge evidence, hosted score rows, and the second derivation pass by @devin-ai-integration[bot] in #4327
- feat(evals): B4 W1 — the versioned run verdict policy contract by @devin-ai-integration[bot] in #4329
- fix(evals): B4 W1 follow-up — nested rate arithmetic, exact decision reasons, execution variants by @devin-ai-integration[bot] in #4330
- Fix/host runtime config retry by @ignaciojimenezr in #4331
- evals: run under the v2 verdict policy end to end (B4 Phase 2) by @devin-ai-integration[bot] in #4332
- fix(conformance): protocol suite entry mapping + JSON-safe report persistence by @chelojimenez in #4333
- fix(github-checks): stop the MCP OAuth handler hijacking the bind callback by @chelojimenez in #4336
- ci(release): parallelize the release pipeline (~53m → ~27m) by @chelojimenez in #4338
- Evals v2 — G4b: the inspector half of pre-run disclosure by @chelojimenez in #4334
- Evaluate: landing page, create-suite page, suite overview by @chelojimenez in #4344
Full Changelog: v2.46.0...v2.47.0