Blog

Notes on autonomous test generation, the 5-signal quality verdict, and what we learn building TFactory.

• Jul 17, 2026 Four releases that tightened the verifier itself 0.9.8 through 0.9.11 in one day: the contract schema stops drifting, the web-server suite stops rotting, a whole flake class dies, a trust boundary gets a name, VAL-3 runs on a real cluster, and the verify lane finally follows the deliverable.
• Jul 13, 2026 A benchmark number is only as honest as the verifier behind it This cycle the fleet published its first external score. The reason that number is worth trusting is the same discipline TFactory was built on: verification with visible evidence, and no credit for work that was not actually run.
• Jul 7, 2026 Verification joins the shell — and takes the screenshots for this post TFactory picks up the fleet-wide shell and the latest models. The same MFA browser harness that verifies the portals captured every authenticated screenshot you see here. A one-page showcase is ready to download.
• Jun 23, 2026 The verify Job went live, and now regression never sleeps Two weeks ago we said the Job-native verify executor was close but the production default flip was not live, and we said exactly why. This week it flipped. Here is what that unlocked: every verdict now comes from a per-task Kubernetes Job in a contract-pinned Nix toolchain, and a regression suite re-runs the persisted test corpus over that same substrate without anyone re-triggering it.
• Jun 21, 2026 Verdicts you can trust: VAL ladders, real pytest runs, and an honest verify-Job rollout This week TFactory shipped the Verification Assurance Level ladder end to end, made the host-venv pytest fallback the thing that actually runs tests on real code, and got the Job-native verify executor close — but the production default flip is not live yet, and this post says exactly why.
• Jun 18, 2026 Evidence you can see: screenshots, recordings, and a login that beats 2FA A green checkmark is not proof. The newest TFactory work makes the verdict legible — visible screenshots and recordings in the cockpit, an honest per-criterion ledger, and a real two-factor login captured on its way through.
• Jun 13, 2026 SSRF guards, auth that fails closed, and one version that can't lie An audit found three ways the verifier could be turned against the things it verifies — a metadata-reachable test lane, an unauthenticated network bind, and a schema version that could drift. This cycle closes all three, additively, behind the required backend gate.
• Jun 10, 2026 v0.9.0: the enterprise foundation A merge gate, a no-AIFactory front door, multi-tenant groundwork, and the start of real polyglot — without diluting the one thing that matters: proving a generated test is worth keeping.
• Jun 8, 2026 When "done" has to survive a crash An event you only send once isn't a contract — it's a hope. v0.7.0 makes TFactory's completion handoff durable, idempotent, traceable, and bounded, then fixes the login that was quietly bouncing everyone back to the door.
• Jun 5, 2026 TFactory stops working alone The test work now arrives governed, leaves on a schema the whole line can read, and shows up in the catalog. Three changes that turn a tool into a node.
• Jun 3, 2026 Your tests found a bug. Now what? Finding the problem was never the hard part. TFactory now hands the fix back to AIFactory, re-tests, and knows when to stop — plus two more things that shipped this cycle.
• May 31, 2026 Your test suite is green. It's also lying to you. A test that runs but never fails isn't a test — it's a decoration. Here's how mutation testing catches the fakes.
• May 31, 2026 One spec, five lanes, three languages Most test generators speak one language. Real features don't. Here's how TFactory fans a single handoff across pytest, Jest, and Playwright.
• May 31, 2026 How do you let an AI touch your cloud without handing it the keys? Real integration tests need real credentials. Real credentials must never touch the repo. The Credential Broker squares that circle.
• May 31, 2026 Run the whole thing on your own GPU (yes, with the Wi-Fi off) If your code can't leave the building, your test generator shouldn't phone home either. TFactory runs fully on local models.
• May 31, 2026 Introducing TFactory: test quality, not test count Why we grade generated tests on five signals instead of dumping fifty and calling it coverage.