ongjin/garu is a project that runs a Korean morphological analyzer of about 1MB scale in the browser via WebAssembly (WASM), and its npm package name is garu-ko. The commit 11ee7d5 of October 27, 2026 is a documentation-style change that does not touch any working code. 2 lines were added to AGENTS.md and 35 lines to docs/wiki/research-history.md, and as supporting material training/ambiguity/review_20261026.json (7,252 lines) and training/boundaries/review_joint_20261026.json (3,840 lines) newly came in. In total 11,129 lines were added, with 0 deletions.
Two experiments are recorded. The first is the study (#75) that extends final part-of-speech correction to sentences that do not match the eojeol (space-separated unit) cache exactly. On the fixed gold set of 9,000 sentences, the morphological F1 (ep_norm) rose from 96.053823% to 96.091643%, and on the 16,407 spoken-language sentences it rose from 91.227384% to 91.295119%, but the literary domain fell by 0.017840 percentage points, so it did not satisfy the no-regression condition across all domains. The second is the study (#76) that selects whitespace restoration candidates and morphological analysis together. On inputs with all whitespace removed, gold F1 rose from 92.5965% to 93.3075% (analysis after current restoration) and 93.7700% (joint selection), but on normal sentences it fell from 96.0538% to 95.5500% and 95.9692%, which is lower than analysis of the original text. At that point the whitespace boundary F1 dropped from the gold 79.22% to 52.29%. The initial spoken-language scoring of #75 was discarded because of a missing dictionary file and was rescored with the canonical converter.
The decision is that neither experiment was incorporated into the default behavior. The research history states that it did not change the existing product code, models, or boundary weights, and that neither was made to turn on automatically. #75 only left a result JSON of 31,075 bytes, which comes to 4,558 bytes when compressed with Brotli q11, and did not bundle it into GMDL and WASM; the browser and WASM package costs were not measured. #76 increased, in Node/WASM samples under load on an M4, the no-whitespace mean from 0.358ms to 1.423ms and the P95 from 0.761ms to 3.758ms, but it does not guarantee latency in the browser or for all inputs. The figures also concern synthetic inputs with whitespace removed, not natural typos. In other words, the fact that it is a Korean morphological analyzer distributed for the browser remains as is, and it is better read as two branches of record that were not adopted rather than as newly added functionality.