docs(taxonomy): #847 AIF structural-homogeneity — SYNTHESIS + VERDICT (ai-01)

[jsboige]

Closes the three-installment measurement chain with the synthesis those docs explicitly deferred to ai-01.

Inputs (all merged, all held to MEASURE not verdict by po-2024):

• docs(taxonomy): #847 AIF structural-homogeneity acompte (layers a+b, read-only) — MEASURE not verdict #850 — layers (a) attack-type entropy + (b) fail-loud scheme rate
• docs(taxonomy): #847 AIF audit Layer (c) reference-reachability (staged, read-only) — MEASURE not verdict #853 — layer (c) reference-reachability
• docs(taxonomy): #847 AIF audit Layer (c) label-distortion flag (staged, read-only) — MEASURE not verdict #857 — layer (c) label-distortion

Verdict: the tree survives the audit. No restructuring proposed.

Of 39 grouping nodes: 4 coherent · 1 bridge · 28 material-resistance · 6 axis-A candidates.

1. Layer (b) dominates. 65.5 % of the 145 fully-modeled leaves carry no native AIF scheme token. Most heterogeneity is material resistance, not authorial grouping error. The method doc's "changes at the margin" impression is confirmed and quantified — and it is a finding in the tree's favour.

2. Ad hominem = cross-cultural bridge node. It is the sole borrowed node of 39, its four leaves are the cleanest material in the measure (b = 0.00), and its base is decently anchored (69 % wiki) — so the heterogeneity is not evidential decay. The tu quoque / whataboutisme seam is real. c-DEFER lifted: annotate, do NOT flatten, do not bill as a grouping defect. This ratifies the deferred modelling decision in #845.

3. Axis-C lexical pressure on the grouping layer is LOW — 1 node in 39. Corroborated structurally: the EN grouping columns are 1:1 with FR everywhere (97 % coverage), i.e. a FR calque, not an independent anglophone tradition tree. Tradition divergence therefore lives at the leaf-reference level, not in the groupings. The escape hatch the method hoped for is narrow — it excuses exactly one node.

4. A confidence gate the n >= 3 threshold does not provide. Entropy normalises by present attack types, so at n = 3–4 a merely uneven split reads as high heterogeneity almost by construction ({3,1} -> 0.811). Gating at n>=8 supported / 5–7 indicative / 3–4 thin, the 6 axis-A candidates reduce to 2 for jsboige:

Disposition	Node	n	a	b	note
arbitrate	Biais naturels (d3)	8	0.954	0.375	cleanest genuine tension in the audit
arbitrate	Pensée biaisée (d2)	13	0.890	0.462	b just under threshold -> partly material
evidence-blocked	Argument bâclé (d2)	7	0.863	0.429	46 % wiki / 24 % dict / 22 % long-tail — repair refs first
watch only	Erreur de raisonnement (d1)	14	0.592	0.357	barely over threshold, dominated 12/2
dropped	Causalité douteuse · Mauvaise interprétation	5 · 4	—	—	indicative / thin

Argument bâclé is the one place the evidential layer changed a disposition — deciding a grouping question on a base where half the citations are decaying is deciding on sand. That alone justified instrumenting #853.

5. The chantier works, measurably. Folding the 8 native-rich Layer-C clusters: fail-loud 65.5 % -> 58.1 % (-7.4 pp), and Obstruction (0.67 -> 0.49) + Erreur mathématique (0.50 -> 0.48) flip from B+A to A-candidate — families that could not be assessed for grouping quality at all now can. Targeting native-rich clusters converts axis-B opacity into axis-A auditability. Next: Abus de langage (n=57, b=0.88), then Tricherie, then Insuffisance. Influence needs nothing — sole coherent family and best-anchored (89 % wiki).

6. A repair backlog that needs no arbitration (§5): 134 dead refs (upper bound — re-probe with GET, rationalwiki 503s on HEAD), Comparaison abusive 80 % dict, Définition inconsistante 50 % long-tail. Start with Argument bâclé — it unblocks item 4.

Honesty caveat, recorded not buried

unflagged is the absence of a flag, never a positive scope match. Only the lexical availability half of the axis-C constraint proved mechanizable; the distorted scope half is unmeasured and mechanically unmeasurable — #857 tested two proxies and documented both as dead ends (dispersion_fr saturates at median 1.00; EN divergence is 1:1 by §3 above). I do not treat it as a blocker, for a specific reason: a node needs a cross-tradition signal to be excused as a tradition artifact, and the grouping layer has none to give. A stretched FR label is an axis-A authoring choice — legitimately jsboige's. The caveat exists so a future reader never mistakes unflagged for "the label fits".

Governance

• Verdict, not application. No node reorganised, 0 prod-CSV write (post-T&A freeze), docs-only.
• Every item in §3.3 is a proposal for jsboige's ratification; §5 is hygiene executable without him.
• po-2024 held the MEASURE not verdict line across all three installments, including publishing the two failed proxies so no future worker re-tries them. That discipline — and the honesty about dead being an upper bound — is what made a defensible verdict possible.

Refs #847 #498 #845 #846 #850 #853 #857 #770 #763

🤖 Coordinator ai-01