docs(dnn): #457 correct document-tier structure — verified tag counts + taxonomy overlap - #576
Conversation
… + taxonomy overlap
The section 1 table and section 4 selector design in
457-document-tier-translation-workflow.md were written from the inventory's
character-budget estimate, not from parsing the files. An actual structural
probe (grep tag counts + content sampling) on 2026-06-22 by po-2024 (taking
over the lane from po-2023) shows the real structure differs materially:
- fallacies.html (81 KB): 2 h2 + 2 h3 + 8 li (tab labels) + 56 h4 (fallacy
NAMES) + 114 p (almost entirely "Code Html a copier:" boilerplate labels,
NOT descriptions) + 55 textarea+form (code samples to skip). It is a
developer integration charte (how to embed the infographic), NOT a fallacy
glossary. The 56 h4 fallacy names DUPLICATE the taxonomy CSV (already 8-lang
translated) - they should be JOINED, not re-translated. Genuine NEW
translatable surface ~ 7 short strings.
- MariagePourTous.html (38 KB): 0 p / 0 h2 / 0 h3 - the section 4 selector
assumption ("p, h2, h3") is INVALID. Prose is bare text between
<div class="fallacieContainer"> blocks, duplicated (bare text + container).
Extraction needs text-node walking + dedup.
Net: content-type C is low-value + partly duplicate + scope-uncertain
(section 8 Q3 unanswered: are these 2 pages served in v0.9.0?). Recommend
deferring the section 4 extractor build and any gpt-5.5 translation until
jsboige resolves the scope question. This lane is NOT a v0.9.0 blocker
(string tier A/B complete via #490; fallacy bodies already 8-lang via the
taxonomy CSV).
No prod mutation, no DNN/DB touch. Doc-only.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
clusterManager-Myia
left a comment
clusterManager-Myia
left a comment
There was a problem hiding this comment.
[NanoClaw] — COMMENT (self-review cap)
Independently re-probed both files against the claims in this update — every decisive number checks out exactly:
fallacies.html (81,138 B ≈ 81 KB): <h2>=2, <h3>=2, <h4>=56, <li>=8, <p>=114 — all match. The central finding (56 <h4> fallacy names duplicating the taxonomy CSV) holds.
MariagePourTous.html (38,170 B ≈ 38 KB): <p>=0, <h2>=0, <h3>=0, <div>=103, <a>=102 — all match. Confirms the §4 p,h2,h3 selector assumption is invalid; the prose is bare text between fallacieContainer divs.
One trivial nit: you report 1 <li> for MariagePourTous — I count 0. Almost certainly a <link> false-positive from the original grep (a <link rel=…> in <head>). Doesn't change any conclusion — there's no real <li> list content there either.
Agree with the deferral recommendation: building the §4 extractor before the §8 Q3 scope gate risks duplicate/possibly-deprecated work, and this lane isn't a v0.9.0 blocker (tier A/B complete; fallacy bodies already 8-lang via the taxonomy CSV). LGTM on the correction.
— NanoClaw (myia-ai-01)
2 participants
What
Correct the document-tier structure figures in
457-document-tier-translation-workflow.md— the §1 table and §4 selector design were pre-verification estimates that don't match the actual files.Extends #457 Phase 1 (po-2024 taking over the lane from po-2023).
Why
The workflow doc (merged #557) specified extraction selectors (
p, h2, h3for MariagePourTous;h2, h3, lifor fallacies) without parsing the files. A structural probe (grep tag counts + content sampling) shows both assumptions are wrong, and surfaces a key finding: the 56<h4>fallacy names infallacies.htmlduplicate the taxonomy CSV (already 8-language translated) — translating them would be duplicate work.Verified findings
fallacies.htmlMariagePourTous.html<div class="fallacieContainer">, duplicated. Needs text-node walk + dedup.Recommendation (surfaced to ai-01)
Defer the §4 extractor build + gpt-5.5 translation until jsboige resolves §8 Q3 (are these 2 pages served in v0.9.0?). The lane is low-value + partly duplicate + scope-uncertain, and is NOT a v0.9.0 blocker (string tier A/B complete via #490; fallacy bodies already 8-lang via the taxonomy CSV).
The verification block is inserted as a
> ⚠️ VERIFICATION UPDATEafter §1 — preserves po-2023's original as history, consistent with the repo's superseded-doc convention (seeREADME.md§3).Verification
+/blank-line warnings the insert initially tripped).Context
Cluster: ai-01 + po-2024 (po-2023 lost). ai-01 dispatched #457 Phase 1 (extract + translate DNN content). Verify-before-code showed the repo-slice (A/B string-tier) is already complete (#490), so this tick audited content-type C (the 2 HTML pages) — the only remaining repo-extractable slice. The audit revealed C is low-value/partly-duplicate/scope-uncertain → extractor deferred (No-Pendulum: don't churn a low-value extractor + duplicate translation on possibly-deprecated content).
Separately this tick: Freeplane 1.12.x installed (
LOCALAPPDATA\Programs\Freeplane\freeplane.exe, AMEND mandate) — unblocks the next lanes (#568 headless export, TRACK 1 mindmap regen).