Sep 19 12:16 ReNikud-yi v3 (labels from the v3 lattice ear, fine-tuned from v2); gold lexicon stress fixes (ืืืืจืื, ืืืืกื, ืืืืืืืืกืืืจื, ืคืจืขืืืืขื ื) Sep 18 11:32 ReNikud-yi v2 (lattice-ear labels, stress head used by the engine); v9 pointing unchanged Sep 13 08:27 phonikud-yi v9 pointing model (retrain9: attested tier with ReNikud-yi as second witness) replaces v8 00:12 ReNikud-yi: leave hand-pointed words alone 00:09 ReNikud-yi context model (int8) on the rule path Sep 9 07:13 phonikud-yi v8: retrain8 (audio-attested tier) replaces v6 โ 48:4 over v6 on the audio yardstick; onnx_yiddish_v6 removed Sep 2 08:05 Rename the eighth table to printed_respelling_lk (reason 'printed-respelling') 06:17 Eighth table: Niborski phonetic-index LK readings (1,991 entries, rescue #3 between Sefaria and the model guess) Aug 28 08:52 Weinreich-44 ืื default (ษj), closed oส รป-class list, ืืจืืืึพ and ืคืืื/ืคืืืื, ืื+ืืืื context read Aug 21 18:11 Drop the superseded v5 export; the loader resolves onnx_yiddish_v6 first and doubling the snapshot doubles every cold boot 18:10 phonikud-yi v6: retrained on chain-attested supervision (65.8% coverage); paired phonetic eval vs v5: +25/-10 on the held-out test set; model_pointed_lk +59 types Aug 18 15:42 Drop stray __pycache__ artifacts 15:41 phonikud-yi engine bundle: G2P + 7 tables + v5 ONNX pointing model 15:40 initial commit 
