My docs said one number. The disk said another.
Old notes, context packs, and portfolio copy kept describing versions of the system that no longer existed.
My instruction file said one number of tools. The disk had a different number. Small, but it was one of many drifted claims: references to scripts I had already archived, an old report still described as the main output, and a feature marked “planned” in one document after it had actually shipped.
None of those documents were badly written. They were wrong because they were stale.
So I applied the same source-of-truth thinking I use for financial data:
- Which document owns this fact?
- Is it current state, or history?
- Is it generated from something authoritative, or copied by hand?
- When was it last verified?
Two changes came out of it. A deterministic checker compares claims in the docs against the disk every day. I didn’t make it an AI sweep, because anything that needs judgment also goes stale. And for context I hand to other tools, I stopped editing copies by hand: a script builds the pack from live data and stamps it with a date and a hash.
A well-written document that’s out of date can be more dangerous than no document at all. (This portfolio was a good example. Rebuilding it meant checking every claim against the code.)
Written from my work building internal systems for a commercial solar contractor. Details are generalized; no company, client, or financial data.