Launch Week 02 wrapped — explore all five launches

September 4, 2026

  • Evals
  • Observability
  • Security
  • Integrations
  • Quality of Life

By the Book

TGIF! Thank god it's features, here's what we shipped this week:

The headliner: governance policies now generate themselves from the documents you already wrote. Upload your existing compliance material—PDF, Word, plain text, or Markdown—and Confident reads it and drafts the policy, controls included. Elsewhere: the Eval Gate stopped being a one-repo relationship, threads total their own tokens and cost, exports went async with notifications attached, and email verification landed alongside cleaner SSO account linking. Everything strictly by the book.

Changelog September 4, 2026

Added

  • Governance Policies from Your Documents - Upload the compliance material you already have—PDF, Word, plain text, or Markdown—and Confident drafts governance policies straight from it. The framework you spent a quarter writing stops being a file in a shared drive and starts being a policy with controls you can actually track, no retyping your own standards into a form. Read the docs, write the policy.
  • Multi-Repository Eval Gate - The Eval Gate now guards every repository in your project instead of making you pick a favorite. Each one carries its own pinned dataset, metric collection, and regression tolerance, runs its own setup, and keeps its own run history—so the service with strict thresholds and the one you're still prototyping can live under the same roof without sharing a config. Monorepo, polyrepo, and that one service nobody wants to touch: all gated. Repo after repo.
  • Thread Token & Cost Totals - Threads now total input and output tokens across the whole conversation, alongside input and output cost. A multi-turn session's real spend is one number in the thread overview instead of a spreadsheet you assemble span by span. Every turn, accounted for.
  • Async Exports with Notifications - Exports stopped holding you hostage. Every export now runs as a background job that notifies you in-app when it's ready—or when it failed—with the download waiting right there in the notification. You can also export a hand-picked selection of traces or threads instead of everything in a window, with the time range now optional. Kick it off, walk away. Export and forget.
  • Email Verification & SSO Account Linking - Email verification is here, with expiring links that sign you straight in and clear success and error states when they don't. Account security settings picked up verification controls on eligible plans, SSO-provisioned users arrive already verified, and account linking now requires a verified local email everywhere except on-prem deployments, which keep working as before. Identity, confirmed.
  • Native Streaming for Experiments, Arena & Test-Run Summaries - Experiments, arena runs, and AI test-run summaries now stream natively, so outputs, metrics, and traces show up as they happen instead of all at once at the end. Test-run summaries also gained topic grouping and progressive overview insights, and the whole path picked up better budget controls, batching, and concurrency—plus partial results that survive an error instead of disappearing with it. Stream as you mean to go on.

Changed

  • Provider Validation on Credential Save - Provider integrations are now validated when you save credentials and when you select a model, so a bad key or a misconfigured provider tells you immediately instead of halfway through an evaluation run. Fail at setup, not at scale. Valid from the start.

That's the drop for this week—see you next Friday.

Built byConfident AI