12. How Mova reads this English today: pipeline, fixes, failures

This chapter follows a JPS 1917 sentence through Mova's annotation cycle and says, stage by stage, what happens to its EModE forms. Every number has a date; all are reproducible from the files named.

12.1 The pipeline (Torah cycle, 2026-10-17)

The run is cycle-torah-2026-10-17 (scripts run*.sh, root root/, five Torah flows: rounds 1 torah-gen, 3 torah-exo, 5 torah-rest, 7 torah-left, 9 torah-final; round 121 is the 0.3 tales flow). For each round (tools/cycle/src/pipeline.rs, main.rs):

12.2 Timeline of the EModE fixes (2026-10-17)

12.3 The refusals, round by round

From root/roundK/prepare.log, lines draft … refused by the gates (2026-10-17):

Torah total: 1 789 refusals of 1 087 distinct sentences. By lemma: shalt 969 (+ Shalt 3), hath 401 (+1), hast 167 (+6), art 141 (+6), didst 45, doth 20, dost 9, canst 7, hadst 5, wast 4, mighty 3, sawest 1, bear 1. The refused sentences were picked up by the later flows («torah-left»), so the cost was rounds and money, not lost text: the db holds 6 065 of the 6 105 sentences (40 missing, 2026-10-17 21:29).

The tales round 121 ran before both fixes; its refusals include dialect forms of other varieties (wuz, kin, twa, sha', mus', maun, wer, bin, ai, sorter, from Uncle Remus (PG 2306), Scots and northern English tales) and EModE forms not on any list yet: mayst, may'st, couldst, wouldst, Wouldst, shouldst, 'twould, 'tis, forsooth, mayhap, and the tokenizer fragments St, st, e, er, k, lt (of may'st, could'st, ne'er…).

12.4 What the small model does (SLM now)

en annotate $MOVA_DATA/runs/mmm-parse-2026-10-15/en-agreed.bin over the 6 105 lines of torah.tsv, built from main at f3e616747 (both fixes in), 4.95 s, two runs byte-identical (2026-10-17). Aligned with the db by sentence id and token index where the forms match: 6 013 sentences, 180 871 tokens.

Archaic tokens are 6.2 % of the text and 41 % of the UPOS disagreements (3 573 + 546 + 151 = 4 270 of 10 393). If they were all right, agreement would be about 96.6 %.

Sentences (of 6 013): 1 821 have no UPOS disagreement at all; 2 245 have a disagreement on an archaic token; 704 have disagreements only on archaic tokens — they would be clean with a lexicon. By class:

The full list of the largest single disagreements (SLM → db), from the same alignment: shalt NOUN→AUX 840, LORD NOUN→PROPN 774, thee NOUN→PRON 568, thereof ADJ→ADV 359, thy ADJ→PRON 231, hath NOUN→AUX 196, hast NOUN→AUX 109, thine NOUN→PRON 92, thee VERB→PRON 87, lest ADJ→SCONJ 62, hath ADV→AUX 50, me PROPN→PRON 46, ye NOUN→PRON 43, art NOUN→AUX 43, mayest ADJ→AUX 42, didst NOUN→AUX 40, even ADV→NOUN 38, whither ADJ→ADV 35, giveth ADJ→VERB 32, wherewith NOUN→ADV 31, thyself NOUN→PRON 27, goest ADJ→VERB 27, thy PROPN→PRON 26, thereon ADJ→ADV 26 (these from the round drafts, round{1,3,5,7,9}/mmm.conllu, which agree with SLM now to within a few tokens).

12.5 How wrong lemmas and features survive

Three mechanisms, each traced in the data:

(a) The tagger's tag decides the lemmatizer's path. archaic() acts only with a verb tag (except the 19 verb-only forms). The SLM tags giveth NOUN or ADJ, so Morph::lemmatize returns giveth. Round 5: mmm.conllu has -eth forms NOUN|NN 319, ADJ|JJ 83, ADP|IN 37, all with lemma = form; only the 50 tagged VERB are lemmatized.

(b) The graph parser fixes the tag but keeps the lemma. Lex::fill(form, lemma, upos, xpos) looks up (form, UPOS, XPOS) in GP_LEX (the school corpus, which has no giveth), and falls back to «the most frequent FEATS of the new tag, with the old lemma». Round 5 mmm-gparse.conllu: -eth VERB|VBZ with lemma = form 405, lemmatized 39. The same fallback gives a re-tagged thou the FEATS of the most frequent PRP, which is it: Case=Nom|Gender=Neut|Number=Sing|Person=3 — the source of the 115 third-person thou and 98 sentences in the db (chapter 3.8).

(c) The teacher patches what it is shown. The teacher's patch format (annot/src/batch.rs, flow patch) corrects the draft; a lemma that looks plausible (giveth) is not corrected. The gate checks the lemma only for AUX. Result in the db: 400 VERB/AUX tokens whose lemma equals an archaic form (giveth 35, goeth 20, toucheth 19, cometh 18, lieth 13, hath 10, saith 10 …; chapter 4.5), in 307 + 21 sentences.

12.6 Errors that pass every gate (db audit, 2026-10-17)

These do not stop a sentence from entering the db; they change what the world layer reads from it.

12.7 Effects downstream

• Pragmatics (prag/src/feats.rs): the subject's person comes from FEATS → D4 makes 98 sentences third-person; the modal comes from the form (norm_word) → shalt, wilt, mayest, canst, shouldest are features apart from shall, will, may, can, should; polarity looks for a negator on the root → D10 turns «I know not the day of my death» into a positive root; the negator list lacks nought, naught, nay, lest. The second-person detector (you list) includes thee thou thy thine ye but not thyself.
• World facts (world/src/worldbook.rs, stands_alone): a sentence with Thee, Thou, Thy, Thine, thee, thou, thy is never chosen to word a principle («no prayer or verse») — 1 580 of the 6 105 Torah sentences. The filter misses ye, thine, hath, -eth; a register score (density of Style=Arch) would be a cleaner test, if the filter is wanted at all for the Torah course.
• Reranker (world/src/rerank.rs): the ArchaicAdverb repair knows 11 adverbs; the most frequent Torah ones are not among them (chapter 10.1).
• Lessons: the Torah course (world/data/school/…, Torah ★ 252 on 2026-10-17 per f3e616747) reads the db. Watching the SLM think: the diagnosis of its answers lists where the small model fails on the course; the EModE layer is one of the causes that does not need a model to fix.

12.8 What works

• Lemmas of the auxiliaries in the db: shalt → shall 890 of 890; hath → have 351 AUX; art → be 43; didst → do 42; mayest → may 50; wilt → will 35 (db, 2026-10-17).
• unto ADP case 2 339 of 2 359; behold INTJ discourse 222; lo INTJ 24 (db).
• ye → PRON nsubj 839; thou PRON 1 452 nsubj (db).
• Hebrew letter names in -eth (Beth, Heth, Daleth, Teth) stay PROPN: the verb-tag guard works.
• art the noun stays a noun (db art|NOUN 2, tales round 121 four nouns): the negative control of archaic_verbs holds in the data.
• The pipeline is deterministic: en annotate gave byte-identical output in two runs.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/en-x-emode/12-mova-pipeline/