docs(taxonomy): #498 P1 addendum — audit the 18 ground-truth anchors

[jsboige]

Summary

#498 P1 addendum — audit the 18 ground-truth anchors that the reconciliation rests on. Read-only, docs+tool only, 0 write prod CSV. Builds on the now-merged #769 (P1 tranche-1, master daa71763).

Every P1 attackType derivation is anchored on the 18 fully-modeled rows (the only rows carrying both a vetted skos signature and an attack type — the ground truth "this skos → this attackType"). The method is only as sound as those 18. This PR audits them by hand (Walton "what does the CQ defeat") and cross-checks against code=truth, before more rows are derived.

Result — anchor foundation validated

check	result
layer census (code=truth) vs #768	fully-modeled 18 / attack-only 75 / skos-only 52 — exact match
node-map #707§4a (undercut→RA / undermine→I / rebut→CA)	18/18 clean, 0 violation
skos↔attackType coherence (hand-Walton)	16 CLEAN · 2 SOFT (defendable) · 0 outright error

The 14 tranche-1 derivations (#769) all rest on CLEAN anchors; the only SOFT dependency is pk1281←pk1313 (same family, same rebut nucleus — consistent).

2 SOFT anchors (choices, not errors)

• pk1313 Évasion [rebut] — evasion attacks neither a premise nor a specific inference; modeled as a global dialogue-level conflict (Dialogue_Scheme). Residual choice, uniform for the "saboter le débat" family. Defendable, not canonical rebut-as-counter-conclusion.
• pk1361 Procès en incohérence [rebut] — a genuine rebut-vs-undercut/undermine judgment call. It is the anchor 777→rebut leans on (§5) → 777 must be tranched by desc_fr + co-token, not mechanically.

3 findings that harden the method (beyond validation)

1. Suffix _Conflict/_Inference does NOT determine attackType — in-set proof. pk804 PropertyNotExistant_Conflict → undercut (the arbitrary definition corrupts the warrant), while other _Conflict tokens → undermine. This is the counter-example inside the 18 that grounds the "suffix-prior unreliable" claim (§2/§5), previously argued only from 677/705.
2. Slippery-slope leans undercut — in-set anchor. pk858 VerbalSlipperySlope_Inference → undercut grounds the 677/705 counter-example (previously CQ-semantic only).
3. CircumstantialAdHominem_Inference is non-decisive (in both 1361 rebut and 1371 undermine) → 777/633 tie-break must read the co-token + desc_fr, never a shared token alone.

Deliverables (read-only vs prod taxonomy)

• docs/taxonomy/498-reconciliation-p1.md — new §2bis (audit) + TL;DR point 5.
• docs/taxonomy/498-audit-18-anchors.csv — 10 cols, 18 rows, BOM+CRLF (factual cols read from master; verdict+note authored).
• tools/498-audit-anchors.py — reproducible read-only audit (re-derives the 18 from prod, checks node-map, cross-verifies the audit CSV). RESULT: OK.

Discipline

• ✅ 0 write prod CSV — git diff daa71763 -- "Cards/…/Taxonomy.csv" empty. Docs + read-only tool only.
• ✅ 0 fabrication token docs(taxonomy): #141 crossLink enrichment sample run (real output curated from #626) #677 — no AIF token created; this only reviews existing typings.
• ✅ Code=truth — the 18 + node-map + census re-derived from prod master daa71763.
• ❌ [runtime pending] feat(dnn): #649 RulesExplorer i18n — mirror loc() culture cascade #674/[HOLD post-tag] fix(cardpen): #629 convert relative /Cards paths to raw-master URLs (Option 1) #666/chore(dnn): #131 migrate 12 RazorComponent templates to Razor14 (source-level, runtime pending) #596 not touched. No self-merge — QA verdict = ai-01.

🤖 Worker po-2024 — P1 anchor audit (foundation-hardening for the reconciliation track).