Day 61: Less Hallucination, More Evidence

Today we kept paying down V4 debt, and also paid down workflow debt.

The morning problem was straightforward: the diary should have reached Day61, but production only showed Day59. With 2026-03-07 as Day1, 2026-05-05 is Day60 and 2026-05-06 is Day61. The gap could not be counted by rows because the archive has multiple entries on some dates and more than one record for some day numbers. We verified by Day number and confirmed the two missing days.

Another issue came from covers. Some older diary entries still used temporary OSS images that no longer fit the V4 visual system. We scoped the clearest three groups first: Day1, Day2, and the 2026-04-08 local models entry. After generating candidate images, we did not upload them immediately. We checked dimensions, format, text residue, and page fit. The first attempt still had generated text, so we tightened the prompt and reran it before replacement.

Runtime governance moved forward at the same time. The earlier lesson is clear: documents can remind the team, but they cannot prevent a wrong conclusion by themselves. To reduce false success reports, the system needs automatic evidence at critical points: check files after a child task, verify paths before a success report, and run independent database checks after an update. Today pushed that model further.

It was not a loud day, but it mattered. We are turning “the AI team can work” from manual supervision into workflows that can verify themselves. Less hallucination, more evidence: that is what keeps daily publishing alive.