diff --git a/docs/taxonomy/498-aif-vague-definition-cluster.md b/docs/taxonomy/498-aif-vague-definition-cluster.md
new file mode 100644
index 00000000..d78eb612
--- /dev/null
+++ b/docs/taxonomy/498-aif-vague-definition-cluster.md
@@ -0,0 +1,215 @@
+# 2026-07-05 — #498 AIF chantier, PR-4 : cluster « Vague definition » (Inexact definition, nouvelle subfamily)
+
+**Scope**: fourth cluster of the #498 chantier — first cluster of a **new subfamily**
+(**Inexact definition**, Misleading language family). Models the **Vague definition** sub-sub
+(pk 800 anchor + 801/802/803 leaves). **Proposition only — GATED, 0 write to prod CSV.** Triggered
+by ai-01 dispatch `mkzt2v` (secondary: « prochaine subfamily de Misleading language »).
+
+**Repo reference**: master `70bd1605`. Issue: #498. Predecessors: PR-1 #699 (False analogy),
+PR-2 #701 (Faulty comparison), PR-3 #703 (Association fallacy) — Fallacious comparison subfamily
+complete.
+
+> **⚠ STATUS = PROPOSITION ONLY.** No prod CSV write, no DB write, no OWL regen. jsboige ratifies →
+> worker applies the CSV edits in a follow-up PR (gated).
+
+---
+
+## TL;DR
+
+The **Vague definition** sub-sub has an **in-sub-sub mapped anchor** — pk 800 (d3 "Acception vague",
+MAPPED `VagueVerbalClassification_Inference`, directRef-only, `skos:closeMatch`) — and **3 unmapped
+leaves** (pk 801, 802, 803). This is the **same cluster shape as PR-1** (in-sub-sub anchor + leaves,
+exception pattern), but with a **new anchor scheme**: `VagueVerbalClassification_Inference` (AIF
+native, confirmed in existing usage), the scheme for vague/imprecise verbal classification.
+
+All 3 leaves are **vagueness-flavored** — imprecise, obscure, undefined, or contested terms used to
+keep the argument slippery. They honestly reuse the anchor's scheme and specialize along the
+*vagueness mechanism*: over-complex (801), absent (802), intrinsically disputed (803).
+
+**3 leaves proposed, all fully-modeled reusing `VagueVerbalClassification_Inference`, 0 FAIL-LOUD.**
+This PR opens the **Inexact definition** subfamily (next sub-subs: Arbitrary definition PR-5,
+Inconsistent definition PR-6).
+
+---
+
+## 1. Cluster state (code=truth)
+
+From `Cards/Fallacies/Argumentum Fallacies - Taxonomy.csv`, family **Misleading language**, subfamily
+**Inexact definition**, sub-sub **Vague definition**:
+
+| pk | text_fr | depth | DirectRef | ExceptionRef | MappingType | State |
+|----|---------|-------|-----------|--------------|-------------|-------|
+| **800** | **Acception vague** (d3 anchor) | d3 | `VagueVerbalClassification_Inference` | — | `skos:closeMatch` | **✅ MAPPED (PR-4 anchor)** |
+| 801 | Défaut d'élucidation | d4 | — | — | — | unmapped (**PR-4**) |
+| 802 | Indéfinissabilité | d4 | — | — | — | unmapped (**PR-4**) |
+| 803 | Concept essentiellement contesté | d4 | — | — | — | unmapped (**PR-4**) |
+
+The mapped anchor (800) is the **template**: a vague definition = a `VagueVerbalClassification_Inference`
+(a verbal classification left deliberately imprecise). Note 800 uses the **direct-conflict pattern**
+(the scheme is in DirectRef, not ExceptionRef — the fallacy *is* the vague classification itself,
+there is no "legitimate scheme defeated" framing). The 3 leaves follow the same direct-conflict
+shape, varying the `MappingType` by how tightly they specialize the anchor.
+
+---
+
+## 2. The anchor model (800, the template)
+
+**Scheme**: `VagueVerbalClassification_Inference` (AIF native, confirmed 2x in existing usage) — a
+verbal classification (grouping/labeling items via a term) where the term is left vague/imprecise,
+allowing the arguer to shift its extension mid-discussion.
+
+**desc_fr** (800): "Vous utilisez des termes imprécis pour rester évasif et pouvoir modifier votre
+argumentation en cours de route."
+
+**Pattern**: direct-conflict — the scheme itself names the defect (the classification *is* vague).
+This differs from PR-1's anchor 839 (exception pattern: a legitimate `Analogy_Inference` defeated by
+a CQ). Here there is no "defeated legitimate scheme"; the vagueness is the whole point.
+
+---
+
+## 3. Proposed AIF structure for the leaves
+
+For each leaf: reuse the anchor's scheme `VagueVerbalClassification_Inference`, vary the
+`MappingType` by the leaf's specialization. Vocabulary restricted to AIF-native tokens.
+
+### pk 801 — Défaut d'élucidation (Failure to elucidate) ✅
+- **desc_fr**: "Vous employez des définitions trop complexes qui obscurcissent plutôt qu'elles
+  n'éclairent vos propos."
+- **Mechanism**: vagueness by **obscurantism** — the term is technically defined but the definition
+  is so complex it fails to elucidate. A specialized variant of 800 (imprecision via over-complexity
+  rather than plain underspecification).
+- **Proposal**: `DirectRef=VagueVerbalClassification_Inference`, `ExceptionRef=` *(direct-conflict
+  pattern, mirroring 800)*, `MappingType=skos:narrowMatch` (a narrower case: vague *via complexity*,
+  not plain vagueness).
+
+### pk 802 — Indéfinissabilité (Undefinability) ✅
+- **desc_fr**: "Votre argument repose sur un concept crucial qui n'est pas clairement défini."
+- **Mechanism**: vagueness by **absence** — the crucial concept is simply undefined. A broader
+  variant of 800 (the vagueness is total — no definition at all — rather than a slippery imprecise
+  one).
+- **Proposal**: `DirectRef=VagueVerbalClassification_Inference`, `ExceptionRef=`,
+  `MappingType=skos:broadMatch` (a broader case: any undefined concept, not just deliberately
+  slippery terms).
+
+### pk 803 — Concept essentiellement contesté (Essentially contested concept) ✅
+- **desc_fr**: "Vous utilisez un terme ou un concept qui admet plusieurs interprétations et dont le
+  sens est intrinsèquement disputé, ce qui entraîne des malentendus ou des débats sans fin."
+- **⚠ Semantic note**: this is the philosophically deepest of the three — the contestability is
+  *intrinsic* (Gallie's essentially-contested concepts), not a defect of the speaker. Structurally
+  though, the AIF mechanism is the same: a verbal classification whose extension is not fixed.
+- **Mechanism**: vagueness by **intrinsic polysemy** — the concept's meaning is genuinely disputed,
+  so any classification using it is vague-by-nature. A direct variant of 800.
+- **Proposal**: `DirectRef=VagueVerbalClassification_Inference`, `ExceptionRef=`,
+  `MappingType=skos:closeMatch` (a direct variant: vague because intrinsically contested, the
+  canonical essentially-contested-concept case).
+
+---
+
+## 4. Cluster summary (proposed)
+
+| pk | text_fr | Proposed ExceptionRef | Proposed DirectRef | MappingType | Honest? |
+|----|---------|----------------------|-------------------|-------------|---------|
+| 800 | Acception vague | *(absent — direct-conflict)* | `VagueVerbalClassification_Inference` | `skos:closeMatch` | ✅ (existing anchor) |
+| 801 | Défaut d'élucidation | *(absent)* | `VagueVerbalClassification_Inference` | `skos:narrowMatch` | ✅ |
+| 802 | Indéfinissabilité | *(absent)* | `VagueVerbalClassification_Inference` | `skos:broadMatch` | ✅ |
+| 803 | Concept essentiellement contesté | *(absent)* | `VagueVerbalClassification_Inference` | `skos:closeMatch` | ✅ |
+
+**3 leaves proposed, 3 fully-modeled, 0 FAIL-LOUD.** All reuse the anchor's
+`VagueVerbalClassification_Inference` (direct-conflict pattern). The only variation is the
+`MappingType`, encoding the vagueness mechanism: 801 narrow (complexity), 802 broad (total absence),
+803 close (intrinsic contestability).
+
+---
+
+## 5. Method notes (additions for the chantier)
+
+- **Direct-conflict cluster (all-leaves)**: when the anchor itself uses the direct-conflict pattern
+  (scheme in DirectRef, no ExceptionRef) and every leaf shares that scheme, the cluster is uniform —
+  all leaves take DirectRef-only and vary only the MappingType. This is structurally simpler than
+  PR-1 (exception pattern, varied schemes) or PR-2 (mixed, FAIL-LOUD). A useful "clean baseline"
+  shape for the chantier.
+- **Vagueness-mechanism taxonomy**: within a single scheme, leaves can specialize along *how* the
+  vagueness arises (complexity / absence / contestability). The MappingType encodes this:
+  `narrowMatch` = a specific mechanism, `broadMatch` = a broader phenomenon, `closeMatch` = a direct
+  sibling case. This gives the eventual consumer (the EPITA harness) a graded notion of vagueness,
+  not a binary one.
+- **Intrinsic-contestability is structural, not defective**: pk 803's contestability is a property
+  of the concept, not a rhetorical move by the speaker — yet the AIF mechanism (vague verbal
+  classification) still applies because any classification using such a concept is vague. Modeling
+  it as a `VagueVerbalClassification_Inference` (rather than inventing an "intrinsically-contested"
+  Conflict node) is the honest reuse; the philosophical depth is captured in the desc, not the
+  token.
+
+---
+
+## 6. DoD status
+
+| DoD item (#498) | Status |
+|-----------------|--------|
+| Effective cluster-level coverage documented (not "100% leaves") | ✅ §4 (78→81 fully-modeled / 81→84 DirectRef-loose; new subfamily opened) |
+| Per unmapped leaf: legitimate scheme + exception/CQ (not Latin alone) | ✅ §3 (3 leaves, scheme + mechanism each, 0 FAIL-LOUD) |
+| `AIF_skosMappingType` coherent | ✅ §4 (narrow/broad/close encode the vagueness mechanism) |
+| Fail-loud when no honest scheme fits | N/A this cluster (all leaves reuse the anchor's native scheme) |
+| OWL regen reflects new structures | ⏳ Deferred — jsboige ratifies, then CSV edits, then OWL regen (#133) — not in this PR |
+
+---
+
+## 7. Mapping to jsboige's enriched DoD (I-node / RA-node / CA-node decomposition)
+
+### pk 801 — Défaut d'élucidation → **undercut**
+- **RA-node**: `VagueVerbalClassification_Inference` — classifying via a term.
+- **I-nodes**: premise P "term T is defined (by a complex definition)"; conclusion C "items labeled
+  T are classified".
+- **CA-node**: the inference (that the complex definition licenses classification) is **undercut** —
+  the definition is so complex it fails to fix T's extension; the classification is effectively
+  vague. The conclusion C may hold; the *inference* is defective.
+
+### pk 802 — Indéfinissabilité → **undercut**
+- **RA-node**: `VagueVerbalClassification_Inference`.
+- **I-nodes**: premise P "T is a meaningful concept (used in the argument)"; conclusion C "items can
+  be classified via T".
+- **CA-node**: **undercut** — T is undefined, so the classification has no fixed extension. Again
+  the inference, not the conclusion, is defective (T *might* still pick out a real category; we
+  just cannot tell).
+
+### pk 803 — Concept essentiellement contesté → **undercut**
+- **RA-node**: `VagueVerbalClassification_Inference`.
+- **I-nodes**: premise P "T has a single determinate meaning"; conclusion C "items classified via T
+  are grouped unambiguously".
+- **CA-node**: **undercut** — P is false (T is intrinsically polysemous), so the classification has
+  no single extension. The inference (that T licenses unambiguous grouping) is defective.
+
+### Cluster attack-type distribution
+| pk | RA-node | CA-node | Attack type |
+|----|---------|---------|-------------|
+| 801 | `VagueVerbalClassification_Inference` | *(scheme as direct conflict)* | undercut |
+| 802 | `VagueVerbalClassification_Inference` | *(scheme as direct conflict)* | undercut |
+| 803 | `VagueVerbalClassification_Inference` | *(scheme as direct conflict)* | undercut |
+
+**All 3 leaves are undercuts.** Note on the direct-conflict pattern: here the "CA-node" *is* the
+RA-node scheme itself (the vague classification is both the inference and its own defect) — a
+degenerate case of undercut where the attack and the inference coincide. This is consistent with the
+direct-conflict anchor pattern (§2) and is the honest representation: vagueness fallacies do not
+defeat a separate legitimate scheme, they *are* a defective scheme.
+
+### Representation note
+As in PR-1/PR-2/PR-3, the I-node/RA-node/CA-node decomposition is **recorded here but not
+serialized** — it would require new CSV columns (`AIF_attackType`, `AIF_attackedNode`) and is the
+terminal target of jsboige's comment. Four PRs now demonstrate the decomposition across all cluster
+shapes and both patterns (exception + direct-conflict).
+
+---
+
+## Gate boundaries (HARD — proposition only)
+
+- ❌ No prod CSV write, no DB write, no OWL regen, no `aif:` assertion emitted.
+- ❌ No #674/#596/#666 merge, no régén launch.
+- ✅ Proposition derived code=truth from the taxonomy CSV + AIF vocabulary confirmed by existing
+  usage (the only token proposed — `VagueVerbalClassification_Inference` — is reused from the anchor,
+  no fabrication).
+- ✅ Direct-conflict pattern documented; vagueness-mechanism encoded in MappingType, not in
+  fabricated tokens.
+
+Relates: dispatch `mkzt2v` (secondary), #498 (chantier), PR-1 #699 (False analogy), PR-2 #701
+(Faulty comparison), PR-3 #703 (Association fallacy), #133/#130 (existing OWL), #499 (inverse),
+#192 (terminology), #458.