The Edit Signal

The edit signal is the name Ryan uses for the empirical finding that AI self-editing degrades output quality past a certain threshold. More revision is not more quality. At a point that is reachable in a normal session, each additional edit starts subtracting from the final output instead of adding to it.

The research backing is twofold. The arXiv 2604.00025 paper Brevity Constraints Reverse Performance Hierarchies in Language Models tested 31 models and found that larger models underperform smaller ones on 7.7% of problems by 28.4 percentage points. The mechanism is scale-dependent verbosity: overelaboration introduces errors that the smaller model does not have the capacity to generate. Brevity constraints fix the inversion, adding 26 percentage points of accuracy.

The second strand is context engineering research from Chroma showing accuracy degradation starting at 20 to 30 thousand tokens of context. Combined with the verbosity finding, the picture sharpens: more input produces worse reasoning, more output produces worse answers, and the mechanism on both sides is the same — elaboration introduces errors faster than it resolves them.

The lived experience of this at Arkeus scale was the lens.md incident. Lens was a cumulative learning file that accumulated 3,050 edits across 39 sessions, including 59 rewrites of the Google Sheet URL, and produced zero useful updates. The loop felt productive — tokens were being burned, the file was being updated, the system appeared to be learning — and the signal was indistinguishable from file churn. It was archived in March.

The rule that came out of that incident: patterns are human-authored proposals from session end, reviewed and approved, not auto-extracted from edit logs. The edit signal is the reason that rule exists. Automation that generates many small edits without a human in the loop does not improve the output; it just generates noise that looks like improvement.