Build log · 9 August 2026
Done means verified: replaying the fix in a real browser
Written up in November 2026 from our engineering log · 2 min read
In early August we introduced Tasks — one unit of work, with its evidence attached — and Rounds, a fixer that works through them. Two days later we added the part that matters most: proving the fix worked.
Evidence says what is wrong; a flow says what is right
A task carries evidence — a screenshot, a scan issue — that shows the problem. It can also carry a test flow, which shows what working looks like. Until then only the first half ran. Now a fix asks for its flows to be replayed, either headless or in the user's own browser after they accept.
Honest outcomes
A step that cannot be replayed is recorded as skipped and turns the run PARTIAL. Only FAILED blocks a push; a declined prompt or an unreachable server means we did not check, never that it is broken. The first false block teaches everyone to switch verification off.
Later, clicking was allowed in these checks, because most of a product sits behind a tab — but clicks that delete, pay, touch credentials, sign out or broadcast are refused — and so is any click that does not say what it targets.