Every open-weights agent model writes file contents between one framing newline on each side. We rendered 1,762 real files as tool calls and parsed them back with five parsers: the format preserves 1,759, a .strip() parser preserves 15. Serving stacks get this right; agent clients re-deriving the parse get it wrong, and a benchmark that grades file contents scores a perfect model at 0.9%.
§ Notes
Notes
Papers, and the method around them. Written when there is something to record, not on a schedule.
15 November 2026
paper
paper
27 October 2026
lab note
lab note
The study above ran on two vCPUs and no GPU. That constraint is part of the method, not a footnote — it sets the model scale, the number of cells, and what the results are allowed to claim.
27 October 2026
paper
paper
Ask for the middle of a list from the beginning and from the end. Same item, same ordinal, one word different. On Qwen2.5-1.5B-Instruct that one word costs 33 points of accuracy — and the loss depends on how the question is phrased.