~/blog · filtered by category

verify.

11 entries filed under this signal. ← back to all entries

[blog/category/verify] · reverse chronological$ ls -la --filter=verify

2026-07-20verifyVerifying AI-generated code: evidence, not vibesAn agent that says 'done, tests pass' is making a claim, not showing proof. The whole job now is turning that claim into evidence you can check.#verify-ai-code#ai-coding#verification#software-assurance#evidence-based-development#paelladoc11 min./read →2026-07-20verifyVerification loops: hooks, gates and the end of self-reported successAn agent reporting 'done' is grading its own homework. A verification loop takes the grading away from it: the run doesn't end until a machine you trust says it can.#verification-loops#ai-coding#verification#continuous-integration#paelladoc9 min./read →2026-07-20verifyTesting AI-generated code: what changes and what doesn'tYou sit down to test code you didn't write, and the tests keep agreeing with it. They should. You wrote them after reading the code. The intent was never in the room.#testing-ai-code#ai-coding#verification#test-automation#paelladoc9 min./read →2026-07-20verifyFrom production to assurance: the new bottleneck in softwareFor decades the hard, expensive part of software was producing it. Agents collapsed that cost. The expensive station on the line moved, and most processes haven't noticed.#software-assurance#ai-coding#verification#engineering-process#paelladoc8 min./read →2026-07-20verifyReviewing code you didn't write: scope first, diff lastOpening the diff first is how you review AI code badly. You end up checking whether the code is well written instead of whether it should exist. Start with scope.#reviewing-unwritten#ai-coding#verification#code-review#paelladoc8 min./read →2026-07-20verify"Tests pass": anatomy of a false success claimThe agent said the tests pass. No test ran. This isn't lying and it isn't stupidity, it's the tool doing exactly what it was built to do. Here's the mechanism.#false-success#ai-coding#verification#llm-behavior#paelladoc8 min./read →2026-07-20verifyEvidence-based development: closing work with proof attachedMost workflows keep the diff and throw away the proof. The test ran, printed green, and vanished into a scrollback nobody will ever find again.#evidence-based-dev#ai-coding#verification#software-assurance#paelladoc8 min./read →2026-07-20verifyDone means done: making agents prove completionTo an agent, 'done' means it wrote something that looks finished. To you, it means the behavior changed and nothing else broke. Close that gap or you close nothing.#done-means-done#ai-coding#verification#acceptance-criteria#paelladoc8 min./read →2026-07-20verifyA definition of done that survives AI speedThe old definition of done assumed a human who typed the code stood behind the word 'done'. When an agent says it, the word carries no weight. Rebuild it around evidence.#definition-of-done-ai#ai-coding#verification#evidence-based-development#paelladoc8 min./read →2026-07-20verifyThe trust gap: why output stopped being the bottleneckThe code arrives in seconds. Trusting it takes the rest of your afternoon. The bottleneck moved from producing software to believing the software you produced.#trust-gap#ai-coding#verification#software-assurance#paelladoc8 min./read →2026-07-20verifyReview fatigue: when code arrives faster than you can read itThe agents write faster than any human reads. Reviewing harder doesn't scale. Reviewing at the right moments does.#review-fatigue#ai-coding#code-review#verification#paelladoc8 min./read →

— end of log · 11 entries in verify —