docs/OSS_TRANSLATOR_METHODS.md

Open-source translator methods → FSOT-safe orientation

We study classical / historical OSS pipelines for structure, not to import free neural MT cores. PFLT stays FSOT 2.1–aligned (scalar teacher, zero free params on the scalar path).

What successful classical tools do

System / tradition	Approach	Transferable idea
Whitaker’s WORDS	Exhaustive Latin prefix + stem + ending tables → lemma → English gloss	Finite analyze-then-gloss; never invent from embedding alone
Morpheus / Perseus	Greek/Latin morphological analysis → dictionary headword	Surface form ≠ meaning unit; lemma is the unit
CLTK	Pipeline: tokenize → normalize → POS/morph → lexicon	Stage separation; dual metrics (lex vs open)
LEMLAT / Collatinus	Rule morphology + large lemma lexicon	Precision over recall on short stems
Historical linguistics pipelines	Cognate tables, sound laws, Swadesh/ASJP	Lineage neighborhood under gates (we already have ASJP hooks)
Gazetteers (Pleiades)	Entity table + multilingual aliases	Names off the morph path (we split dual-track)

What we deliberately do not do

• LLM / NMT as the translation core
• Soft generative “guessing” that breaks FSOT teacher gates
• Mixing unique entities into morph open-set metrics

Applied in PFLT (v0.1 → now)

1. Dual track — core lexicon vs name gazetteer (Pleiades contacts)
2. Analyze-then-gloss — reverse morph + paradigm expand + prefix strip
3. Lemma-first donors — inject prefers content glosses for transfer
4. Precision gates — short-stem collisions blocked unless high form similarity

Next increments

• Richer Latin/Greek prefix inventories (Whitaker-style)
• Sense preference by FSOT domain context
• Cognate bridges only when multi-source support exists