📘 English Wikipedia Fish

The Fish annotate English Wikipedia a little at a time with the Codex teacher (Terra gpt-6.1-terra, ChatGPT login), the English counterpart of the French Fish (peta/fish/fr/README.md). Mova's English parse and pragmatic labels are the draft the teacher corrects; the validated replies are review-only silver (not gold, no training grant).

• Contract and executor: tools/fish/src/wiki.rs, binary fish-wiki (annotation-execute like the English and French ones, run by the unchanged pager fish-pages). Gates: bound fields, exact targets with their # sent_id and # text, CoNLL-U tree, the UD English registry of en::udreg (UPOS, features per UPOS, relations, auxiliary lemmas, the copula be), word rows only (no multiword ranges, as in the English contract), source-character alignment, the closed labels of prag::schema. Witnesses (findings, never a rejection): Mova's parse against the reply (parser_divergence when under half the labelled attachments agree), Mova's pragmatic labels against the teacher's (pragmatics_differs_from_draft), SpaceAfter marks. fish-wiki wiki-check CONLLU runs the gate over a file (control: UD English EWT dev, 1696 of 2001 blocks pass; the others carry multiword ranges or empty nodes only). wiki-prepare and wiki-render show the native evaluation and the request without a model call.
• Pages and corpus: fish-wiki-pages pages reads a finished treestore pass (finished rank-*.mts parts only; the pass keeps articles in page-view order) and writes pager pages: one page = the first 16 sentences with a tree of one article, two tasks of 8 targets, 4 context sentences. Sentences whose text is in a held-out treebank are rejected and never sent (--heldout, required); titles starting List of, Lists of, Deaths in, Index of, Outline of are skipped. --pagers w08,w09,… deals the pages round-robin to one plan and pager configuration per name over a shared ready/. fish-wiki-pages collect gathers validated jobs into a corpus and a snapshot (one article in ten, by FNV-1a of the page id, is dev and never trains). limit-guard.sh stops all pagers of a build when one turns uncertain on anything but a transport error or a bare provider-capacity refusal.

Build: cargo build --release here and in tools/fish.

Dry run (2026-11-06)

Pass wiki-all-v5 (page-view order), the first 2 articles (.xxx, 2026 FIFA World Cup), 32 targets in 4 tasks; held out: UD English EWT dev+test, GUM dev+test, PUD test (7784 texts). 4 Codex calls, all validated (32 of 32 targets); 124-180 s per call; tokens per call 18.1-20.9 k input (8.2 k cached), 6.5-9.7 k output; rendered requests 21-25 KB. Mova's parse agrees with the teacher on 685 of 990 words (69 %; 60-76 % per task). The teacher labels all 32 sentences expository: 30 decl/assert/0/0/none/none, 1 hedge, 1 character voice, where the draft said narrate/narrator for much of it. A held-out control (a text of the first article put into --heldout) is rejected and not sent.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/tools/fish-wiki/