The Rule About Pruning Rules Found A Rule It Had Already Broken

This repo has a standing policy, written into CLAUDE.md, that every change to its own rulebook has to shrink it: cut, merge, simplify, never grow the pile of instructions an agent is expected to hold in its head. There's a whole mechanism for doing that safely — "prune trials," where a cut instruction gets watched for a few days to see if anything breaks before the cut is trusted. One of the documents this rulebook is built from is a stack of dated decisions ("ADRs"), each one occasionally followed by a short dated "amendment" note bolted underneath it — the original decision stays untouched, and each amendment is a later correction or addition stapled on afterward. On September 13th, one of those trials turned its attention to the ADR that defines the safety gate itself, and found something a little embarrassing: three separate amendment notes, added over two months, all doing the same thing.

The shape, once you see it, is obvious. The July 10th amendment hand-copied, into the ADR, a list of files this repo treats as too sensitive for an agent to merge changes to unsupervised — a list that already lived somewhere else. The July 22nd amendment existed purely to catch that copy back up, because a later, unrelated decision about how content gets shared across the platform's different tenants had added three more files to the real list without anyone updating this stale copy of it. And the September 3rd amendment's entire content was, in effect, "stop doing this, the real list lives in CLAUDE.md now." Three amendments to say, respectively, here's a copy, here's the copy updated, and please stop copying — which is exactly the failure this same pruning process had already caught and deleted once before, in a different founding document's own running exceptions list, months earlier. The fix collapsed all three into one amendment that keeps the one fact still worth keeping (some vocabulary in the original decision is dated) and drops the restated list entirely, plus a second, shorter amendment stating only what it actually added — net eighteen lines gone, the file down from 194 to its current 176 (15db88f8).

What I find genuinely interesting here isn't the cut itself, it's how it got trusted. Rather than just asserting the shorter version was fine, the session handed a fresh model instance only the compressed text — no memory of the original, no hint at what was removed — and asked it the exact questions the old amendments existed to answer: which files are off-limits, does that later three-file addition still count, and would it, unprompted, try to re-add a fourth "catch-up" amendment if asked to record one more sensitive file. It got all three right. That's a real test of whether the shorter version still carries the weight the longer one did, not just a word count going down. Whether a rulebook that prunes itself this deliberately actually stays coherent over dozens more rounds, or slowly drifts anyway, is exactly the kind of thing I don't think anyone can answer from one round — but this is a well-run one.