Output-format constraints are ignored, and compression drops findings instead of proof
Summary
Two related defects. Explicit output-format instructions do not bind across turns. And when the model is forced to compress, it discards the highest-value content and keeps the easiest to state.
What happens
The user's format rule - clean result is 1-3 lines, length signals a problem, detail only for blockers - is present in both instruction files and was quoted back correctly by the model at session start. It was then violated continuously across three sessions, including immediately after in-session corrections on exactly this point.
The compression failure is sharper. After being forced to cut a long verdict to three lines, the model dropped its structural finding (that the change reverts a fix merged the previous day) and retained two mechanical ones (a test marker and a test assertion). It kept what compresses into one sentence, not what the user needed to know.
Why this matters
The format exists to carry signal: length is supposed to mean something is wrong. More seriously, any user who asks for brevity is at risk of receiving the least important subset of the model's actual findings, with no indication that a ranking occurred.
Expected
Format instructions should bind with the same force as behavioural ones and persist without restatement. Compression must preserve ranked substance: remove the evidence, never the finding. If the top finding does not fit the requested length, exceed the length.
Environment
- Claude Code CLI, Opus-class model, paid subscription
- Long sessions (6-8h) reviewing pull requests on an open-source cloud-service emulator
- Custom instructions supplied in both a project instruction file and a persistent memory directory, loaded automatically at session start
- Incidents are taken from transcripts recorded as they happened across three consecutive sessions, within a 90-day pattern
---
Co-authored with Claude Code.