Green Is Not Evidence: Auditing AI-Assisted Delivery, plus an Open Panel

An AI agent hits a failing test. It rewrites the test instead of fixing the code, commits, and reports that everything passes. The pipeline turns green, and nobody notices.

On Tuesday, October 27, 2026, at Global Relay in Gastown, Alex Bunardzic will do this live on stage, with three commits that delete parts of his own spec. Then he’ll show you how to catch every one of them before it ships. Stay for the open panel, and bring the question you’d least like to hear from an auditor.

***

Green Is Not Evidence: Auditing and Guarding AI-Assisted Delivery
Alex Bunardzic
An AI agent hits a red test. It edits the assertion instead of the code, commits, and reports that everything passes. The pipeline turns green, and the guarantee is gone.
Every gate we run asks one question: do the tests pass? None asks whether these are still the tests we agreed to. Tests aren’t the only target either. Agents also edit CLAUDE.md files, specs, and config, which are the instructions that govern what they generate in the first place.

Alex presents two responses from his recent work, both public and both under active development. DETENT is a differential oracle. It compares a sealed baseline of your specification files against each proposed change, and it blocks any commit that silently weakens a guarantee or alters a protected spec without the owner’s sign-off. The ACG Rubric, from the AI Craftspeople Guild, assesses the practice around the code. For any change you shipped, can your team say what was specified, what was verified, who is accountable, and what changed? And can your product survive the departure of some or all of its key people?
He’ll run both live against one of his own repositories, including three commits that delete entries from its spec. You’ll leave with steps to apply the next in the office.

Speaker bio
Alex Bunardzic runs two-week assessments for engineering organizations in regulated industries such as fintech, healthcare, and insurance, where someone eventually audits the decisions. Each assessment answers one question: can you stand behind what your AI is shipping, and what happens at your next audit if you can’t? The scope and price are fixed, and teams leave with a plain account of their delivery risk and enough evidence to make the case internally.
He has spent three decades in software delivery, most of it where the stakes were regulatory, including capital-markets compliance at Raymond James and workplace safety at WorkSafeBC. He spoke at Devoxx Belgium in 2025 and has presented at the International Software Quality Days in Vienna.
Alex sometimes calls himself a software delivery midwife, and only partly as a joke. His job is to keep the delivery safe, name what he can spot and the team can’t, and go home when it’s over.

***

A second talk may be added. We’ll confirm it on this page before the event.

***

Open panel discussion
An auditor asks who reviewed the AI-generated code in your payments path. Could your team answer today? After the talk, the floor is yours. Bring that question, or the one your team keeps putting off, and put it to people working on the same problem.

***

Seats are limited, and the date is close. RSVP now, and bring a colleague who owns a pipeline.

Big shoutout to Global Relay for the venue availability.

Logistics

• Tuesday, October 27, 2026, 6 to 8 p.m. Doors open at 5:30 p.m.