Language models are extremely good at producing a number that looks like the number you wanted. That is the whole problem.
The standing rule for every agent in this system: if a fact is not in the
project’s configuration or its data, do not supply one. Mark it [VERIFY] and
move on. A gap that is labelled gets closed. A gap that got filled with something
plausible stays wrong for months, because it never looks wrong.
This rule is not decorative here. The site you are reading has live examples:
- Zero signed evidence records. The proof page reports
0in the signed column, in the same size type as every other figure. The signing code is complete and tested; no keys have been provisioned. Omitting the column entirely would have been easy and would have read better. - Phase statuses on the overview. The project config still describes phases that the handoff notes say are underway. Rather than reconciling them into a tidier story, the config is shown as written and the disagreement is flagged.
- Unfinished case studies. Three Anatomy entries are visible stubs marked
[PLACEHOLDER]rather than hidden until they are written. A section that looks complete because the incomplete parts were filtered out is a small lie told by omission.
What this cost, once
The 13-week arc on the overview page carried a [VERIFY] flag for days. The
project config said week 2, the handoff file said week 7, and the calendar said
week 15. Any one of them could have been quietly adopted and nobody would ever
have known, 7 was the most flattering, being comfortably inside the plan.
Instead the exporter published the calendar arithmetic and the declared value side by side and marked the disagreement, and the flag stayed up until the underlying vault was actually corrected. The flag then came down on its own, because the check is still running rather than having been switched off.
The page now says “13 weeks planned, 15 weeks in”, which is worse-looking and true, and the plan was not retroactively restated to 15 weeks to close the gap.
A site about trustworthy automation that quietly rounded its own numbers would be arguing against itself.