docs(taxonomy): #847 AIF audit Layer (c) label-distortion flag (staged, read-only) — MEASURE not verdict

docs/taxonomy/847-layerc-label-distortion.csv

@@ -0,0 +1,43 @@
+# 847 Layer-(c) label-distortion flag — per internal node. MEASURE not verdict. INPUT for ai-01 synthesis. GATED jsboige ratification. 0 prod CSV.
+"# label_fit in {borrowed, unflagged}. borrowed=Latin/anglicism (taxonomy Latin col = ground truth + morphology, §2.1 b2). unflagged=NO mechanical flag (NOT verified scope match). stretched/narrowed (§2.1 b1) are SEMANTIC — two proxies tested & non-discriminatory: dispersion_fr SATURATED (median 0.94-1.00), n_en_groupings=1 everywhere (EN=FR translation, no axis-C structural signal in grouping layer)."
+# combined_reading = label_fit x acompte decision_rule_tag. borrowed x A-candidate = PRIORITY (tradition-divergence -> bridge-node); borrowed x coherent = VINDICATED.
+level,node,n_leaves_full,n_modelled,label_fit,borrowed_lang,dispersion_fr,n_en_groupings,acompte_tag,combined_reading
+family(d1),Abus de langage,89,57,unflagged,-,1.0,1,B+A,"material-resistance + type-mix, no label signal"
+family(d1),Erreur de raisonnement,102,14,unflagged,-,0.793,1,A-candidate,"tree-tension, NO mechanical label-distortion signal -> genuine axis-A mix OR human semantic read needed (stretched/narrowed not mechanically detectable)"
+family(d1),Erreur mathématique,102,12,unflagged,-,0.968,1,B+A,"material-resistance + type-mix, no label signal"
+family(d1),Influence,420,12,unflagged,-,0.885,1,coherent,"mechanism-coherent, no label signal"
+family(d1),Insuffisance,174,16,unflagged,-,0.935,1,B+A,"material-resistance + type-mix, no label signal"
+family(d1),Obstruction,126,12,unflagged,-,0.95,1,B+A,"material-resistance + type-mix, no label signal"
+family(d1),Tricherie,394,22,unflagged,-,0.942,1,B+A,"material-resistance + type-mix, no label signal"
+sub-family(d2),Ad hominem,46,4,borrowed,latin,0.875,1,A-candidate,"PRIORITY: borrowed x tree-tension -> tradition-divergence (bridge-node, do NOT bill as defect; corroborates #853 wiki-anchor)"
+sub-family(d2),Ambiguïté,41,31,unflagged,-,1.0,1,B,"material-resistance, no label signal"
+sub-family(d2),Appel à l'émotion,57,4,unflagged,-,1.0,1,coherent,"mechanism-coherent, no label signal"
+sub-family(d2),Argument bâclé,68,7,unflagged,-,0.917,1,A-candidate,"tree-tension, NO mechanical label-distortion signal -> genuine axis-A mix OR human semantic read needed (stretched/narrowed not mechanically detectable)"
+sub-family(d2),Arranger les faits,85,4,unflagged,-,0.939,1,B,"material-resistance, no label signal"
+sub-family(d2),Causalité douteuse,29,5,unflagged,-,0.929,1,A-candidate,"tree-tension, NO mechanical label-distortion signal -> genuine axis-A mix OR human semantic read needed (stretched/narrowed not mechanically detectable)"
+sub-family(d2),Changement de cap,51,5,unflagged,-,1.0,1,B+A,"material-resistance + type-mix, no label signal"
+sub-family(d2),Comparaison fallacieuse,13,10,unflagged,-,1.0,1,B+A,"material-resistance + type-mix, no label signal"
+sub-family(d2),Définition inexacte,34,16,unflagged,-,1.0,1,B+A,"material-resistance + type-mix, no label signal"
+sub-family(d2),Généralisation abusive,37,4,unflagged,-,1.0,1,B+A,"material-resistance + type-mix, no label signal"
+sub-family(d2),Manipulation mentale,239,5,unflagged,-,0.818,1,B+A,"material-resistance + type-mix, no label signal"
+sub-family(d2),Mauvaise composition,32,4,unflagged,-,1.0,1,coherent,"mechanism-coherent, no label signal"
+sub-family(d2),Mauvaise déduction,40,4,unflagged,-,0.875,1,B+A,"material-resistance + type-mix, no label signal"
+sub-family(d2),Mauvaise interprétation,34,4,unflagged,-,0.923,1,A-candidate,"tree-tension, NO mechanical label-distortion signal -> genuine axis-A mix OR human semantic read needed (stretched/narrowed not mechanically detectable)"
+sub-family(d2),Pensée biaisée,257,13,unflagged,-,0.946,1,A-candidate,"tree-tension, NO mechanical label-distortion signal -> genuine axis-A mix OR human semantic read needed (stretched/narrowed not mechanically detectable)"
+sub-family(d2),Procédé rhétorique,123,3,unflagged,-,0.975,1,B,"material-resistance, no label signal"
+sub-family(d2),Préjugé,63,4,unflagged,-,0.9,1,B+A,"material-resistance + type-mix, no label signal"
+sub-family(d2),Refus du débat,31,4,unflagged,-,1.0,1,B+A,"material-resistance + type-mix, no label signal"
+sub-family(d2),Résultat invalide,30,4,unflagged,-,1.0,1,B+A,"material-resistance + type-mix, no label signal"
+sub-family(d2),Saboter le débat,48,3,unflagged,-,1.0,1,B+A,"material-resistance + type-mix, no label signal"
+sub-family(d2),Surinterprétation,42,4,unflagged,-,1.0,1,B+A,"material-resistance + type-mix, no label signal"
+sub-sub(d3),Acception arbitraire,22,6,unflagged,-,1.0,1,B+A,"material-resistance + type-mix, no label signal"
+sub-sub(d3),Acception vague,4,4,unflagged,-,0.0,1,B+A,"material-resistance + type-mix, no label signal"
+sub-sub(d3),Ambiguïté narrative,11,11,unflagged,-,1.0,1,B,"material-resistance, no label signal"
+sub-sub(d3),Amphibologie,8,6,unflagged,-,1.0,1,B,"material-resistance, no label signal"
+sub-sub(d3),Biais culturels,68,3,unflagged,-,1.0,1,B+A,"material-resistance + type-mix, no label signal"
+sub-sub(d3),Biais naturels,150,8,unflagged,-,0.896,1,A-candidate,"tree-tension, NO mechanical label-distortion signal -> genuine axis-A mix OR human semantic read needed (stretched/narrowed not mechanically detectable)"
+sub-sub(d3),Comparaison abusive,5,3,unflagged,-,0.0,1,B+A,"material-resistance + type-mix, no label signal"
+sub-sub(d3),Définition inconsistante,7,5,unflagged,-,0.0,1,B,"material-resistance, no label signal"
+sub-sub(d3),Equivoque,21,13,unflagged,-,1.0,1,B,"material-resistance, no label signal"
+sub-sub(d3),Fausse analogie,5,4,unflagged,-,0.0,1,B+A,"material-resistance + type-mix, no label signal"
+sub-sub(d3),Justification triviale,22,3,unflagged,-,1.0,1,coherent,"mechanism-coherent, no label signal"

docs/taxonomy/847-layerc-label-distortion.md

@@ -0,0 +1,146 @@
+# #847 AIF structural audit — Layer (c) label-distortion flag (read-only, 0 run)
+
+> **Provenance.** Layer-(c) **second installment**, dispatched by ai-01
+> (`msg-20260723T024907-2l19wg`, tick 89, PRIMARY). **Read-only — 0 run live, 0
+> prod-CSV.** Companion to [#853](847-layerc-reference-reachability.md) (Layer-(c)
+> reference-reachability, the §2.2 half). This completes the §2.1 half: the
+> **internal-node labelling constraint**. Method doc
+> [`aif-structural-audit-method.md`](aif-structural-audit-method.md) §2.1, §4, §5.
+> **MEASURE, not verdict. INPUT for ai-01 synthesis. GATED jsboige ratification.**
+> Scope: the **39 internal nodes** in [`847-acompte.csv`](847-acompte.csv) (the
+> (a)+(b) layers were measured on the 145 fully-modeled subset; the
+> `combined_reading` join only resolves there).
+
+---
+
+## TL;DR
+
+`label_fit` is computed for the **39 grouping nodes** of the acompte. The §2.1
+labelling constraint has **two bullets**; only one is mechanically computable:
+
+| §2.1 bullet | What | Mechanically computable? | This instrument |
+|-------------|------|--------------------------|-----------------|
+| **(2) lexical availability** | a term is **borrowed** from another tradition where no FR term exists | **Yes** (lexical facts) | ✅ `borrowed_from_lang` |
+| **(1) distorted scope** | label **stretched / narrowed** vs the cluster | **No** (semantic) | ⛔ proxies tested, non-discriminatory (§3) |
+
+**Headline:** of 39 grouping nodes, **1 is `borrowed`** (`Ad hominem`, raw Latin —
+confirmed against the taxonomy's own `Latin` column); **38 are `unflagged`** (no
+mechanical distortion signal). `unflagged` means *no mechanical flag raised*, **not**
+a verified scope match.
+
+**Axis-C lexical pressure on the grouping layer is LOW** (1/39 raw-Latin): the
+taxonomy is authored in FR descriptive phrases (or FR calques), importing raw Latin
+**only** where no FR single-term equivalent exists — *Ad hominem* (the FR calque
+*Attaque personnelle* sits at d3, but the d2 grouping keeps the Latin). This is a
+real structural finding about the tree.
+
+## The one priority candidate (distorted × heterogeneous)
+
+| Node | Level | label_fit | acompte tag | Combined reading |
+|------|-------|-----------|-------------|------------------|
+| **Ad hominem** | sub-family (d2) | **borrowed (latin)** | **A-candidate** | **PRIORITY arbitration candidate** → **tradition-divergence → bridge-node, do NOT bill as grouping defect** |
+
+This **resolves the `c-DEFER` on Ad hominem** from the acompte. Combined with #853
+(Ad hominem's Obstruction base is **69% Wikipedia-anchored**, well-evidenced), the
+layer-(c) reading is consistent: *Ad hominem*'s mechanism-heterogeneity is the
+**tradition's** (a Latin scholastic term spanning cultures), **not** a jsboige
+grouping error. The bridge holds.
+
+There are **no `borrowed × B+A`** nodes (Ad hominem is the only borrowed node), so
+no second-tier "label may have forced a bad merge under material-resistance" cases
+arise from this mechanical pass.
+
+## A-candidate resolution (the 7 tree-tension nodes)
+
+| Node | Level | label_fit | Layer-(c) labelling read |
+|------|-------|-----------|--------------------------|
+| **Ad hominem** | d2 | **borrowed** | tradition-divergence → **bridge-node** |
+| Erreur de raisonnement | d1 | unflagged | no label signal → genuine axis-A mix **or** human semantic read |
+| Argument bâclé | d2 | unflagged | " |
+| Causalité douteuse | d2 | unflagged | " |
+| Mauvaise interprétation | d2 | unflagged | " |
+| Pensée biaisée | d2 | unflagged | " |
+| Biais naturels | d3 | unflagged | " |
+
+**Reading for ai-01/jsboige:** of the 7 tree-tension (A-candidate) nodes, **only
+Ad hominem carries a mechanical axis-C label signal** (borrowed Latin). The other 6
+have **no mechanical labelling distortion** → their heterogeneity is either a
+genuine axis-A grouping choice **or** a stretched/narrowed label that this
+first-pass cannot detect mechanically (semantic; §3). Layer (c) therefore does
+**not** bill any of the 6 as tradition artifacts — it leaves them as axis-A
+candidates for jsboige, exactly per the method doc's "check axis C first; if low,
+the grouping is genuinely his to arbitrate."
+
+## Why stretched/narrowed are NOT mechanically derived (§3 — saves future work)
+
+Two obvious mechanical proxies for §2.1 bullet (1) (distorted scope) were tested
+and found **non-discriminatory**. They are reported in the CSV as contextual
+evidence (`dispersion_fr`, `n_en_groupings`) but **not** used to derive `label_fit`.
+
+1. **Concept-dispersion (within-language scope sprawl).** `dispersion_fr` =
+   distinct FR Wikipedia slugs among a node's FR-referencing leaves / FR-referencing
+   leaves. **Saturated**: median **0.94 (d1) / 1.00 (d2) / 1.00 (d3)** — i.e. at
+   sub-family and below, essentially **every leaf is its own Wikipedia article**, so
+   distinct-slugs ≈ leaves everywhere. It cannot distinguish a tight cluster from a
+   sprawled one. *(Note: `dispersion_fr = 0.0` on a few small d3 nodes means their
+   leaves carry **no FR Wikipedia reference** (dict-only / empty), not "perfectly
+   clustered" — a coverage gap, flagged, not a distortion signal.)*
+
+2. **EN-grouping divergence (cross-tradition structural signal).** For each FR
+   grouping node, count distinct EN grouping labels (`Family`/`Subfamily`/`Subsubfamily`)
+   among its leaves. **Result: 1:1 everywhere** (97% EN coverage; every FR node maps
+   to exactly one EN node). The EN grouping columns are a **FR translation/calque,
+   not an independent EN-tradition tree**. So no axis-C structural divergence is
+   encoded at the grouping layer — the cross-linguistic divergence the method doc
+   describes (§2 axis C) lives at the **leaf-reference** level (different Wikipedia
+   articles per language, already measured in #853), not in the grouping labels.
+
+**Consequence:** a defensible `stretched`/`narrowed` flag requires **human semantic
+read** (compare each grouping label's consecrated scope to its members) — it is not
+mechanizable from the taxonomy + #853 data without an LLM-assisted pass (heavier,
+gated). This first-pass delivers the **mechanically-grounded** part (`borrowed`) and
+documents the dead-end on the rest, so no future worker re-tries these proxies.
+
+## Governance
+
+- **MEASURE, not verdict.** 0 reorganisation wording. Synthesis + verdict = ai-01 /
+  jsboige. Same governance as #850 / #853.
+- **0 prod-CSV write** (T&A freeze). Docs-only artefact. DatasetUpdater untouched,
+  `Enabled=false`.
+- **`borrowed_from_lang` is grounded in lexical fact**, not semantic claim: a node
+  is `borrowed` only if its label matches the taxonomy's own `Latin` column or Latin
+  morphology (conservative). `unflagged` is explicitly *not* a positive "exact"
+  verdict — it is the absence of a mechanical flag (mirrors #853's "dead = upper
+  bound" honesty; a no-match is not a content claim).
+- Calques (e.g. *Pétition de principe* ← *petitio principii*) are a **gray zone not
+  flagged** by this conservative pass: the term was translated into FR (FR words),
+  so by §2.1 bullet (2) it is not a raw borrow. Flagged for human read if a calque's
+  scope is later found distorted.
+
+## Caveats
+
+1. **Coverage of `borrowed` is deliberately narrow** (raw Latin/anglicism only).
+   A broader "compromise-label" notion (vernacular terms pressed into taxonomy
+   service, e.g. *Tricherie*, *Humour*) is real but not mechanically detectable
+   without a French frequency lexicon — out of scope, flagged.
+2. **`unflagged` ≠ "label fits".** 38/39 nodes are unflagged; this instrument makes
+   no positive scope-match claim for any of them. It only surfaces the 1 node where
+   a mechanical distortion signal exists.
+3. The join resolves **only** the 39 acompte nodes. Full-tree extension (all ~92
+   grouping nodes) is the same heuristic but yields no `combined_reading` without
+   the acompte (a)/(b) layers — out of scope here.
+
+## Artefacts
+
+- [`847-layerc-label-distortion.csv`](847-layerc-label-distortion.csv) — per
+  internal node (39 rows): `level,node,n_leaves_full,n_modelled,label_fit,borrowed_lang,
+  dispersion_fr,n_en_groupings,acompte_tag,combined_reading`.
+- Compute script (scratchpad): `compute-847-layerc-label-distortion.py`.
+
+## Refs
+
+- Method: [`aif-structural-audit-method.md`](aif-structural-audit-method.md) §2.1,
+  §4, §5. Companion Layer-(c) half: [`847-layerc-reference-reachability.md`](847-layerc-reference-reachability.md)
+  (#853). Acompte (layers a+b): [`847-acompte-homogeneity.md`](847-acompte-homogeneity.md)
+  (#850). Tracking: #847. Chantier: #498. Dispatch ai-01 `msg-20260723T024907-2l19wg`
+  (tick 89). Base master `82a1e027`.