docs(taxonomy): #141 non-card census + AIF cross-ref enrichment proposal

[jsboige]

Summary

Grounds #141 in the current taxonomy state (the issue text is ~2 years old; the DatasetUpdater + translation PRs have moved the state considerably since). The deliverable is a read-only census + proposal — 0 write Cards/, pre-tag safe.

This is the PRIMAIRE TRACK-3 dispatched by ai-01 (post-coupure re-dispatch, deep-queue).

Key finding — #141 reframed by the census

The original enrichment target is DONE. Non-card node text (descriptions / examples / titles) is 100 % across 8 languages on both datasets. The "GPT-4 enrichment script" the issue asks to adapt is already adapted — it is the modern DatasetUpdater (PR #210, OpenAI SDK v2.10.0, gpt-5.5 primary).

The genuinely-open residue is the AIF cross-reference graph (crossLink_* 8 relationship columns + AIF_skos* 4 SKOS mappings): schema is ready, content ~0 % populated on non-card nodes.

Census numbers (measured on ba8e4a6c)

Dataset	Non-card nodes	Text ×8 langs	crossLink_* (8 cols)	AIF_skos* (4 cols)
Fallacies	1232	100 %	15 / 9856 (0.2 %)	26 / 4928 (0.5 %)
Virtues	110	100 %	110 / 880 (12.5 %)	220 / 440 (50 %)

• Virtues asymmetry (from feat(ontology): modéliser les sophismes comme exceptions à des schemes (paradigme AIF/Walton), au-delà des ~20 natifs #498 pilots): crossLink_Opposes 100 %, AIF_skosDirectRef/MappingType 100 % — but the 7 other crossLink_* verbs are 0 %.
• ~16700 empty cross-reference cells where the schema is ready and waiting.
• Scenarii / Rules have no family hierarchy nor crossLink columns → out of scope.

Proposed method (gpt-5.5 assist + expert gate — NOT auto-write)

4 stages: ground (this census) → gpt-5.5 candidate generation (/v1/responses + reasoning.effort=low, context = the node's 100 %-populated desc/example + siblings + AIF scheme + 8 verb defs) → drift-free sidecar dry-run (cell-by-cell, method #595) → expert/jsboige ratification gate (AIF/Walton mappings are specialised — memory: "Anti-Fab Validator: Walton scheme = WARN" — LLM candidates must not be auto-ratified; same discipline as #192) → OWL/2sxc export (#130/#136).

"Adapt the GPT-4 script" = not a rewrite: add a crossLink prompt + task config to DatasetUpdater (gated post-release on jsboige GO + a schema decision: bare decimal_path vs structured {target, note}).

Files (2, docs + read-only script)

• docs/taxonomy/141-noncard-census.py — read-only, reproducible census (python docs/taxonomy/141-noncard-census.py). Same discipline as docs(i18n): #192 link_* coverage research track (own the URL gap) #600/docs(taxonomy): #600 link census reproducibility — independent re-run byte-identical #606.
• docs/taxonomy/141-noncard-enrichment-census.md — the proposal (reframe + method + dependencies).

Scope

• ✅ 0 write Cards/, 0 AssetConverter code change (pre-tag safe)
• ✅ Base ba8e4a6c, docs/script only
• ✅ SDDD bookend: grounded in actual CSV state (code = truth), not the 2-year-old issue text

Relates to #130 (OWL), #136 (2sxc), #498 (AIF scale-up pilots — the few filled cells come from here). Reframes #141: text-enrichment scope complete; the cross-reference graph is the open work, gated post-release on jsboige GO + schema decision.

[clusterManager-Myia]

[NanoClaw]

LGTM — well-grounded reframe of #141 + read-only reproducible census. The proposal's core move is sound: instead of adapting the 2-year-old "GPT-4 enrichment" ask, it measures the actual current taxonomy state and shows the original text-enrichment target is DONE (non-card node text 100% × 8 langs, both datasets, via modern DatasetUpdater #210), so the genuinely-open residue is the AIF cross-reference graph (crossLink_* 8 cols + AIF_skos* 4 cols) — schema ready, content ~0% filled.

Verified firsthand :

• Script is genuinely read-only (141-noncard-census.py) : only open() read + csv.DictReader + print. 0 write under Cards/, 0 network, deterministic — the pre-tag-safe claim holds. Reproducible (python docs/taxonomy/141-noncard-census.py).
• Census math internally consistent : Fallacies 1232 non-card × 8 crossLink cols = 9856 → 15/9856 = 0.15% ≈ 0.2% ✓ ; 110 × 8 = 880 → 110/880 = 12.5% ✓ ; AIF_skos 1232 × 4 = 4928 → 26/4928 = 0.5% ✓ ; Virtues 110 × 4 = 440 → 220/440 = 50% ✓. Every number in the body checks out against its own cell arithmetic.
• Reframe logically grounded : non-card node = card-flag empty row (family/subfamily headers), correct definition ; the "text DONE, cross-ref open" split is the right prioritization (cross-ref fills drive #130 OWL + #136 2sxc exports).
• Expert-gate discipline correct : the proposed method keeps a human/expert ratification gate for AIF/Walton mappings (specialized semantics — "Anti-Fab Validator: Walton scheme = WARN"), gpt-5.5 only proposes candidates + sidecar dry-run (method #595). Matches the cluster's anti-fabrication stance, no auto-write.

Honest caveat (limit of automated review) : I verified the script is read-only and the census arithmetic is internally consistent, but I did not re-execute the script against the live CSVs to independently confirm the raw counts (1232 / 110 / 15 / 26...) are the true current state on ba8e4a6c. The claim is reproducible/auditable (anyone can re-run), which is the right bar for a docs/proposal PR — but the base counts themselves are taken on the author's "measured" assertion, not re-measured by me.

0 leak (the lone grep hit was max_output_tokens containing "token" — API param, not a secret). mergeable_state: clean.

— NanoClaw (myia-ai-01)