Benchmarks
Two reproducible benchmarks of what an AI agent spends editing documents. The addressing benchmark: on the same documents and the same 47 edits, saying where an edit goes costs 13,109 bytes in Markdown and 611 in GEML, 21.45× less. The mixed-toolchain benchmark: giving each of a real day's 14 edits to the GEML verb built for it cut the reading by 3.65× and the bytes spent saying where by 7×.
| script | what it measures |
|---|---|
addressing-cost.mjs | One edit's cost, on a controlled corpus: four documents this repository keeps as both Markdown and GEML, ~12 blocks sampled from each, both arms executed for real. |
real-session-replay.mjs | A full day of real editing, with each edit done the way that suits it — mechanical bulk replacement left on the original commands, edits that need an address moved to GEML. |
GEML_SRC=../geml node benchmarks/addressing-cost.mjs
GEML_SRC=../geml node benchmarks/real-session-replay.mjsBoth take --json for the per-edit rows.
real-session-edits.json is the frozen baseline for the replay: 33 edits one agent made to README_CN.md over a single day, recovered from its session log, each carrying what it actually cost. The log itself is not published — that file is everything derived from it, and it can be audited row by row. The GEML side is executed on every run, so the numbers move when the CLI does.
Design decisions for both — corpus, sampling, what is counted, and the two choices that deliberately favour the Markdown side — are stated in each script's header comment, and were fixed before the first run.