Without verification there is no way to tell whether the work moved toward done or away from it. So in LifeOS nothing counts as finished until a tool has shown it: a test that passed, a probe of the live site, a screenshot from a real browser, an eval that scored the output. The evidence ships with the work.
Every criterion carries its check
Each criterion in an ISA names what would prove it false, and carries that check as something the system can actually run. Deterministic checks run first. When a criterion can only be judged, such as whether copy sounds right, an eval with a calibrated judge scores it. Bunker runs all of them from one command. The check has to match the criterion: a page’s appearance is checked by looking at rendered pixels, and a 200 response from the server doesn’t prove the page looks right.
”Should work” is banned
The phrase is banned from every LifeOS report. Either something was verified, or the report says it wasn’t. The builder also never grades its own work on anything that matters. High-stakes changes get a second look from a separate model that reads the result fresh, without the builder’s reasoning to lean on.
Saying “I don’t know”
The same rule applies to answers. A confident guess is worse than an admitted gap, because you might act on it. When the system hasn’t checked something, it says “I haven’t verified that” in the same sentence. When it doesn’t know, it says so and explains what it would take to find out. A second model agreeing with an answer doesn’t count as evidence.
In practice
A deploy report lists what was probed: the live URL returned the new page, a real browser rendered it at desktop and phone widths, and the link check found nothing broken.