1 Training data← Back to the data viewer2 TokenizerCompareAuditExplainer slidesCoordinate tokens3 Image budget4 MP vs 8 MP4 NotesData processingAssamese ra checkSynthetic checkSmall held-out blocksData EDAJaigyan latticeJaigyan v4 re-readEmpty pagesManipuri: 4B vs pipeline5 TrainingEmbedding warm-upSFT ablationsSFT merges2B distillation0.8B distillationDistillation explainedRL planRL ablationsRL training

Jaigyan lattice

v4 · 4B-RL re-read — progress, yields by rule and Bodhan vs 4B-RL comparisonsOpen →

Small working set: 500 pages per language from the 6M trusted list. A block is accepted (Bodhan's text) by exact agreement, or after the lattice when every disagreeing slot has exact support from another model. Comparison keys fold: quote styles, dashes, Gujarati two-sign o, visarga = colon, Urdu optional diacritics, line-end hyphens, "$". A page is rejected early when a model has no transcript for a block (English: Gemma may be missing), when Bodhan writes a foreign script, or when the language label is wrong. B = Bodhan, G = Gemma 4 31B, S = Surya OCR 2. Green slot: Bodhan supported; red slot: no support. Inside a red slot, green words of Bodhan still match a model word for word (hover: which model). Supported cells show as accepted text (hover: the other variant). Lexicon = decided by the in-domain dictionary (built from the accepted text of the full run): Gemma = Surya and only their word is valid, or all differ and only Bodhan's word is valid. Solo = all three models differ and Bodhan's word is kept, when a block has at most 2 such cells (ground truth: Bodhan right ~2/3, CER 1.7%). Amber rule = Gemma = Surya and Bodhan differs by ≤ 2 characters: their word is taken (not in pa, te). Punctuation from G+S = Gemma and Surya agree and differ from Bodhan only in punctuation: their punctuation is taken (hover: Bodhan's version; user, 2026-11-04). Bodhan form = accepted: Bodhan's cell equals a model after a reduction (old doubled consonant; Gujarati ો read as ા; LaTeX super/subscript markup removed); hover shows which. Near = red cell accepted because Bodhan differs from one model by ≤ 1 character or ≤ 2% CER (not amber cells; user, 2026-11-04). Click a bar segment to see samples of that tier. In the pages view, rejected-after-the-lattice pages (1-10 red cells) are sorted fewest first; boxes: green = accepted block, red = red cells, dark red = rejected early. In near cells and everywhere else, the final text is always Bodhan's own. amber cell = Gemma and Surya agree with each other against Bodhan (not accepted; the judge pilot decides per language).