docs(taxonomy): #141 crossLink enrichment sample run (real output curated from #626)

[jsboige]

Summary

Primaire of dispatch lofjtd (ai-01). Concrete-output sample for the #141 crossLink enrichment — the real before/after the adaptation proposal #673 stages. Docs-only; 0 prod CSV write; gated post-tag + post-Stage-3.

What this sample is (transparency first)

The enrichment was already run in #626 — a fullscale gpt-5.5 closed-set pass over all 1232 non-card taxonomy nodes (3850 candidate links, 0 fabrication warnings). Rather than re-spend gpt-5.5 credits to reproduce output that already exists, this doc curates a 7-node sample from #626's real output (141-aif-candidates-fullscale.json, model gpt-5.5-2026-04-23) and shows the before/after, quality, and cost. Re-running would fabricate an illusion of new work; the generation is done, what remains is expert adjudication + the ratified write.

Before state (code=truth, prod taxonomy CSV)

Surface	Empty
crossLink_* (8 verb cols × 1408 rows)	11242 / 11264 (99.8%)
AIF_skos* (4 cols × 1408 rows)	5470 / 5632 (97.1%)

The 7-node sample

Representative spread across depth (d=0,4,4,4,5,6,7) and confidence (0.38–0.98):

decimal_path	label	#cand	conf range	ratifiable (≥0.70)
0	Argument fallacieux (root)	4	0.62–0.78	2
1,322	Appel au bénéficiaire	3	0.78–0.90	3 (all)
4,211	Transposition impropre	4	0.62–0.98	3
7,2111	Changement de sujet	4	0.58–0.88	2
7,32211	Tactique du complice	1	0.95	1
4,117	Avoir toujours raison	3	0.45–0.62	0 (all defer)
6,311227	Compensation du risque	2	0.38–0.45	0 (all defer)

All 7 had empty crossLink_* cells in prod (before-state verified).

Quality eval (21 links across 7 nodes)

• Confidence min/mean/max: 0.38 / 0.70 / 0.98
• Ratifiable (≥0.70): 11/21 (52%); defer 10/21 (48%)
• Verbs exercised: 4/8 (Leverages ×9, IsRelatedTo ×8, Mirrors ×3, Allows ×1)
• Fabrication warnings: 0 — closed-set anti-fab held (the docs(taxonomy): #141 AIF Stage-1 full-scale results — 1232 nodes, 0 fabrication #626-validated result; confirms proposal docs(taxonomy): #141 crossLink DatasetUpdater adaptation (additive proposal + sample, gated) #673's UseFunctionCalling=true design)

Findings surfaced for the Stage-3 expert gate

1. Multi-target serialization (the real adjudication question). 5/7 nodes have ≥1 verb with multiple candidate targets, but the prod cell convention holds ONE dotted decimal_path. Expert must decide: highest-conf? canonical? or schema-extend to ;;-joined lists? Judgment, not engineering — confirms the closure recommendation.
2. Confidence threshold. 0.70 cleanly separates clean structural mappings (root/family-level) from deep-leaf generic associations — but the threshold decision is the expert gate's.
3. SkipNonEmpty=true is safe — none of the 7 sample nodes had pre-existing values; the 22 expert cells sit on different decimal_paths.

Cost

• This sample: 0 new gpt-5.5 spend (curated from docs(taxonomy): #141 AIF Stage-1 full-scale results — 1232 nodes, 0 fabrication #626).
• docs(taxonomy): #141 AIF Stage-1 full-scale results — 1232 nodes, 0 fabrication #626 fullscale: 1232 nodes, 3850 links, single closed-set pass (already paid). Fresh re-run not warranted unless the docs(taxonomy): #141 crossLink DatasetUpdater adaptation (additive proposal + sample, gated) #673 prompt design changes (it uses the same closed-set + 8-verb schema as docs(taxonomy): #141 AIF Stage-1 full-scale results — 1232 nodes, 0 fabrication #626).

What this PR does NOT do

❌ No prod Cards/ write · ❌ No AssetConverter C# change · ❌ No fresh gpt-5.5 call · ❌ No auto-ratification · ❌ Does not resolve the multi-target question (surfaces it for Stage-3).

Reproducible

build_141_sample.py (scratchpad) — read-only curation from #626 fullscale JSON + prod taxonomy CSV. Deterministic (no RNG). Output 141_sample.json. 0 write under Cards/.

Relates #141, #673, #626, #609, #620, #623, #130, #136, #498, #595. Base d5913862.

🤖 Worker po-2024 (dispatch lofjtd, primaire)