diff --git a/docs/taxonomy/141-aif-stage3-adjudication.md b/docs/taxonomy/141-aif-stage3-adjudication.md
new file mode 100644
index 00000000..03c282be
--- /dev/null
+++ b/docs/taxonomy/141-aif-stage3-adjudication.md
@@ -0,0 +1,170 @@
+# #141 AIF Stage-3 — expert adjudication package (DRY-RUN, no auto-apply)
+
+**Author**: po-2024 (worker) · **Date**: 2026-07-01 · **Base**: master `33b1c0bc`
+**Source of truth**: [`141-aif-stage2-diff.csv`](141-aif-stage2-diff.csv) (16 rows, reproducible via
+[`141-aif-stage2-diff.py`](141-aif-stage2-diff.py)) · derived from the
+[full-scale Stage-1 sidecar](141-aif-candidates-fullscale.csv) (#626, merged).
+**Status**: **ADVISORY PACKAGE FOR THE EXPERT GATE** — recommendations only. **Nothing here is
+auto-applied.** Every row marked ⚖️ is a human (jsboige / argumentation expert) call.
+
+This is the one-pass adjudication table for the **12 non-card nodes** that already carry an AIF
+value on the taxonomy (0.97 % of 1232). For those 12, the full-scale generator (gpt-5.5,
+closed-set) was compared against the existing expert annotation. The other **1220 nodes are entirely
+net-new** — no existing value to conflict with, so they are out of this adjudication (they go to the
+general Stage-3 ratification, prioritized by confidence > 0.8).
+
+## Verdict summary
+
+| Verdict | Rows | Nodes | Action |
+|---|---:|---:|---|
+| CONFIRM | 2 | 2 | gpt-5.5 re-derived the **exact** expert token — **ratify, no change** |
+| CONFLICT | 9 | 6 | gpt-5.5 proposed a different token/field — **adjudicate below** |
+| SILENT | 5 | 5 | gpt-5.5 stayed silent on a filled field — **preserve existing** |
+
+## CONFIRM — ratify as-is (2)
+
+| dp | label | field | token (existing = proposed) |
+|---|---|---|---|
+| `1.111` | Appel à l'ignorance | DirectRef | `Ignorance_Inference` |
+| `5.321` | Expression vague | DirectRef | `VagueVerbalClassification_Inference` |
+
+Both are independent corroboration: the closed-set generator reproduced the expert's exact Walton
+token with no prompt priming. Highest-confidence signal in the whole run.
+
+## CONFLICT — adjudicate (9 rows, 6 nodes)
+
+Each node below: existing annotation vs gpt-5.5 proposal, an **advisory recommendation** (⚖️ expert
+decides), and a one-line Walton rationale. A cross-cutting convention underpins several recos
+(see [Convention](#convention)).
+
+### `4.116` Pente glissante — field-swap + token churn (4 rows)
+
+| field | existing | gpt-5.5 proposed |
+|---|---|---|
+| DirectRef | `RequiredSteps_Conflict` | `CausalSlipperySlope_Inference; NegativeConsequences_Inference` |
+| ExceptionRef | `CausalSlipperySlope_Inference` | `RequiredSteps_Conflict` |
+| ExceptionRef | `FullSlipperySlope_Inference` | `RequiredSteps_Conflict` |
+| ExceptionRef | `PrecedentSlipperySlope_Inference` | `RequiredSteps_Conflict` |
+
+⚖️ **Reco: MERGE — adopt the field-swap, preserve the dropped tokens.**
+- **Adopt**: `_Inference` schemes as **DirectRef** (primary reasoning scheme), `RequiredSteps_Conflict`
+  as the conflict/structure side — gpt-5.5's field allocation aligns with Walton's scheme-vs-conflict
+  distinction (see [Convention](#convention)).
+- **Preserve** (gpt-5.5 dropped these, but they are legitimate Walton slippery-slope subtypes):
+  `FullSlipperySlope_Inference`, `PrecedentSlipperySlope_Inference` → keep in ExceptionRef.
+- **Adopt the addition**: `NegativeConsequences_Inference` (valid subtype gpt-5.5 surfaced).
+- Net = **union of both sets under the swapped field allocation**, not swap-and-drop.
+
+### `6.31123` Apophénie (1 row)
+
+| field | existing | gpt-5.5 proposed |
+|---|---|---|
+| ExceptionRef | `Sign_Inference` | `OtherCausalFactorsInvolved_Conflict; SignFromOtherEvents_Conflict` |
+
+⚖️ **Reco: PROMOTE `Sign_Inference` to DirectRef; absorb the proposed `_Conflict` tokens as
+ExceptionRef critical questions.** Apophénie's core is a *sign* mistaken for a real relationship —
+`Sign_Inference` is the primary scheme, arguably mis-located in ExceptionRef. The proposed
+`OtherCausalFactorsInvolved` / `SignFromOtherEvents` are well-formed critical questions (alternative
+causes / alternative signs) and enrich the exception structure. Expert: confirm the field promotion.
+
+### `6.311234` Preuve anecdotique (1 row)
+
+| field | existing | gpt-5.5 proposed |
+|---|---|---|
+| ExceptionRef | `InductiveInference_Scheme` | `WeakestLink_Conflict` |
+
+⚖️ **Reco: KEEP `InductiveInference_Scheme` (canonical).** Anecdotal evidence = hasty induction
+from a single case — `InductiveInference_Scheme` is the textbook Walton fit. `WeakestLink_Conflict`
+is a critical question about the weakest link in a chain; it can be added as a *companion*
+critical question but does **not** replace the scheme. gpt-5.5 is over-conservative here (dropping
+the scheme). Expert: keep scheme, optionally add `WeakestLink_Conflict` as a critical question.
+
+### `6.312` Biais émotionnels (1 row)
+
+| field | existing | gpt-5.5 proposed |
+|---|---|---|
+| DirectRef | `OppositeConsequences_Conflict` | `Bias_Inference` |
+
+⚖️ **Reco: ADOPT `Bias_Inference` as DirectRef; demote `OppositeConsequences_Conflict` to
+ExceptionRef.** The fallacy *is* a bias — `Bias_Inference` (gpt-5.5) is the stronger primary scheme.
+`OppositeConsequences_Conflict` is a critical question (consequences going the other way) that fits
+the exception structure, not the primary scheme. gpt-5.5 is right; existing DirectRef is mis-tiered.
+
+### `6.321` Biais d'attribution (1 row)
+
+| field | existing | gpt-5.5 proposed |
+|---|---|---|
+| ExceptionRef | `Example_Inference` | `OtherCausalFactorsInvolved_Conflict` |
+
+⚖️ **Reco: ADOPT proposed; flag both as weak DirectRef fits.** Attribution bias = misattributing a
+cause. `Example_Inference` (existing) = generalizing from an example — weak fit (attribution isn't
+example-based). `OtherCausalFactorsInvolved_Conflict` (gpt-5.5) better captures the error
+(*ignoring alternative causes*) but is a `_Conflict` token (critical question, not a scheme).
+**Expert**: check whether a cleaner `Causal_Inference` / `CauseToEffect_Inference` scheme exists in
+the Walton vocabulary for the primary DirectRef; absent that, adopt gpt-5.5's token as the exception
+structure and keep the field flagged.
+
+### `6.322` Essentialisme (1 row)
+
+| field | existing | gpt-5.5 proposed |
+|---|---|---|
+| ExceptionRef | `Preference_Scheme` | `BiasedClassification_Conflict` |
+
+⚖️ **Reco: ADOPT proposed; existing looks mislabeled.** Essentialism = treating categories as
+innate/fixed. `Preference_Scheme` (existing) is an odd fit (preference ≠ essence).
+`BiasedClassification_Conflict` (gpt-5.5) captures the classification-bias core, but is a `_Conflict`
+token. **Expert**: check for a `Classification_Inference` / `VerbalClassification_Inference` scheme
+for a cleaner DirectRef; the existing `Preference_Scheme` appears mislabeled and should not be
+preserved as-is.
+
+## SILENT — preserve existing (5)
+
+gpt-5.5 did **not** re-propose a token for these already-filled fields. The correct action is to
+**preserve the existing expert value** — silence is not a signal to drop.
+
+| dp | label | field | existing (preserve) |
+|---|---|---|---|
+| `5.3212` | Pente glissante sémantique | ExceptionRef | `VerbalSlipperySlope_Inference` |
+| `6.31121` | Effet de simple exposition | ExceptionRef | `Example_Inference` |
+| `6.312` | Biais émotionnels | ExceptionRef | `Preference_Scheme` |
+| `6.31212` | Biais d'autocomplaisance | ExceptionRef | `PositiveConsequences_Inference` |
+| `6.312311` | Justification de l'effort | ExceptionRef | `Waste_Inference` |
+
+> Note: `6.312` Biais émotionnels appears in both CONFLICT (DirectRef) and SILENT (ExceptionRef
+> `Preference_Scheme`). The DirectRef conflict is adjudicated above; its ExceptionRef
+> `Preference_Scheme` is preserved here (gpt-5.5 was silent on that field).
+
+## Convention
+
+The recurring thread across the 9 CONFLICTs is a **scheme-vs-conflict tier distinction** in Walton's
+AIF:
+
+- **`*_Inference` / `*_Scheme` tokens** name a *reasoning scheme* (the pattern of inference) →
+  natural **DirectRef** (primary scheme the fallacy instantiates).
+- **`*_Conflict` tokens** name a *conflict/decision structure* or a *critical question* → natural
+  **ExceptionRef** (the exception/critical-question side).
+
+gpt-5.5's closed-set picks respect this distinction more consistently than some existing annotations
+(which occasionally put `_Conflict` tokens in DirectRef or `_Scheme` tokens in ExceptionRef). Four of
+the six conflict nodes (`4.116`, `6.31123`, `6.312`, `6.322`) improve under the convention; two
+(`6.311234`, `6.321`) are weak on both sides and need a cleaner scheme token. **Adopting the
+convention is itself a gate decision** — it is surfaced here, not imposed.
+
+## What this package does NOT do
+
+- **No auto-apply.** Every ⚖️ row awaits the expert gate. No `Cards/` write, no DatasetUpdater run.
+- **No verdict on the 1220 net-new nodes.** Those go to general Stage-3 ratification, prioritized by
+  confidence > 0.8 (the high-signal subset, cf. the [full-scale report](141-aif-fullscale-report.md)).
+- **No schema-extension decision** (the 87 `bad_map:*` WARNs — `relatedMatch`/`exactMatch`/`none`/
+  `noMatch`) — that is a separate gate call, already surfaced in the full-scale report.
+
+## Reproducibility
+
+The data in this package is regenerated deterministically from
+[`141-aif-stage2-diff.py`](141-aif-stage2-diff.py) (reads the taxonomy + the Stage-1 sidecar). The
+recommendations are curated prose over that data; re-running the diff after any taxonomy change will
+refresh the table, at which point the recos should be re-checked.
+
+Relates to #141, #609, #620, #626, #130, #136, #192. Honors pre-tag freeze (0 `Cards/` write, 0
+AssetConverter code change). Memory honored: Anti-Fab "Walton scheme = WARN" — this package treats
+every Walton-scheme mismatch as adjudication, never silent overwrite.
diff --git a/docs/taxonomy/141-aif-warn-triage.csv b/docs/taxonomy/141-aif-warn-triage.csv
new file mode 100644
index 00000000..0ab9de59
--- /dev/null
+++ b/docs/taxonomy/141-aif-warn-triage.csv
@@ -0,0 +1,88 @@
+source_dp,source_label,mappingType,bad_map_verdict,recommendation
+"2,131",Jeu de sonorités,none,none,"DROP — weak hedge, no defensible mapping"
+"2,1312",Assonance,none,none,"DROP — weak hedge, no defensible mapping"
+"2,132",Répétition,none,none,"DROP — weak hedge, no defensible mapping"
+"2,1322",Anaphore,none,none,"DROP — weak hedge, no defensible mapping"
+"2,13222",Symploque,none,none,"DROP — weak hedge, no defensible mapping"
+"1,2224",Appel au nombre,skos:exactMatch,skos:exactMatch,ADOPT — extend observed set (legit SKOS predicate)
+"3,211",Confusion entre corrélation et causalité,skos:exactMatch,skos:exactMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,1346",Zeugma,skos:noMatch,skos:noMatch,"DROP — weak hedge, no defensible mapping"
+"6,3111112",Rappel en série,skos:noMatch,skos:noMatch,"DROP — weak hedge, no defensible mapping"
+"1,1116",Argument d'omniscience,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"1,1211",Argument par ouï-dire,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"1,23321",Blasphème,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,1141",Appel au slogan,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,1142",Petite phrase,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,1144",Technique du carpaccio,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,1151",Obscurcissement,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,11513",Obscurantisme,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,115242",Langage des titres de presse,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,1211",Blague,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,12124",Bathos,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,122221",Wellerisme,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,122222",Antiparémie,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,1223",Ironie,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,122323",Second degré,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,123",Cynisme,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,1231",Humour noir,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,1311",Allitération,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,1314",Onomatopée,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,1321",Accumulation,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,13221",Épiphore,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,1332",Paraphrase,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,1341",Aposiopèse,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,1345",Syllepse,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,136",Substitution,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,1362",Astéisme,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,2123",Piège émotionnel,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,2133",Appel à la fierté,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,2234",Hyperbole forcée,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,3111",Objectif caché,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,312",Conditionnement opérant,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,3124",Renforcement Intermittent,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,3151",Induction hypnotique,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,3152",Message subliminal,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,322112",Amorçage,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,322113",Ce-n'est-pas-tout,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,3221211",Porte-au-nez,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,322122",Pied-dans-la-bouche,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,3224",Appel au dernier effort,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,323",Jeu psychologique,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,323231",Sealioning,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,324",Conditionnement social,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,3241",Abus de pouvoir,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,32422",Frapper le chat,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,32423",Kiss up kick down,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,32424",Sycophantie,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,332",Chronémique,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,334231",Poignée de main,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,334234",Salut romain,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,33424",Idéologique,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,334243",Imposition des mains,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"2,334252",High five,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"3,1231",Stéréotype,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"3,232",Mauvais argument de la régression infinie,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"3,323",Erreur numérique,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"4,124",Contre-induction erronée,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"4,3111",Sophisme génétique,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"4,3311",Milieu non distribué,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"6,11112",Exagération,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"6,1115",Faux effet,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"6,11216",Lance à incendie de mensonges,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"6,1131",Contre-vérité ironique,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"6,2262",Départ optionnel,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"6,311121211",Effet d'humour,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"6,311131",Effet suivant-en-ligne,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"6,311235",Sophisme animiste,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"6,31211212",Amplification et minimisation,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"6,312313",Biais de croyance,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"6,3123222",Avoir toujours raison,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"6,3213",Biais de proportionnalité,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"6,322244",Evasion,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"6,322246",Appropriation culturelle,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"6,3323",Radicalisme politique,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"6,3325",Extrémisme,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"7,1211",Argument autocollant,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"7,1332",Delenda Carthago,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"7,2123",Gish gallop,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
+"7,2124",Ignorer la question,skos:relatedMatch,skos:relatedMatch,ADOPT — extend observed set (legit SKOS predicate)
diff --git a/docs/taxonomy/141-aif-warn-triage.md b/docs/taxonomy/141-aif-warn-triage.md
new file mode 100644
index 00000000..29a198da
--- /dev/null
+++ b/docs/taxonomy/141-aif-warn-triage.md
@@ -0,0 +1,70 @@
+# #141 AIF WARN `bad_map:*` triage (DRY-RUN, no auto-apply)
+
+**Author**: po-2024 (worker) · **Date**: 2026-07-01 · **Base**: master `33b1c0bc`
+**Source of truth**: [`141-aif-candidates-fullscale.json`](141-aif-candidates-fullscale.json)
+(structured sidecar, #626) · triage reproduced by
+[`141-aif-warn-triage.py`](141-aif-warn-triage.py) → reviewable
+[`141-aif-warn-triage.csv`](141-aif-warn-triage.csv) (87 rows).
+**Status**: **ADVISORY — the schema-extension decision for the expert gate.** Nothing is auto-applied.
+
+This completes the expert-gate package for #141 (alongside the [Stage-3
+adjudication](141-aif-stage3-adjudication.md) of the 12 existing-AIF nodes and the
+[closure recommendation](141-closure-recommendation.md)). It addresses the **other** gate
+decision surfaced in the [full-scale report](141-aif-fullscale-report.md): the 87 WARN nodes.
+
+## What the 87 WARNs are (and are not)
+
+Each WARN is `bad_map:*` — the generator proposed an AIF `mappingType` **outside the *observed*
+on-disk set** `{broadMatch, closeMatch, narrowMatch}`. The validator is strict-by-construction: it
+flags anything not in the observed set. **These are legitimate SKOS predicates or explicit hedges —
+not fabrications.** There are **zero** fabricated scheme names and **zero** invented targets across
+all 1232 nodes; the closed-set design contains fabrication completely. The 87 are a **schema call**:
+extend the observed set, or down-grade.
+
+## Triage (87 nodes, 7 % of 1232)
+
+| mappingType | Nodes | Recommendation |
+|---|---:|---|
+| `skos:relatedMatch` | 78 | **ADOPT** — extend observed set |
+| `skos:exactMatch` | 2 | **ADOPT** — extend observed set (but ⚠️ strong claim — verify per-node) |
+| `none` | 5 | **DROP** — weak hedge, no defensible mapping |
+| `skos:noMatch` | 2 | **DROP** — explicit no-match hedge |
+| **Total** | **87** | **80 ADOPT / 7 DROP** |
+
+### ADOPT — extend the observed set (80 nodes)
+`skos:relatedMatch` (78) and `skos:exactMatch` (2) are standard SKOS mapping predicates.
+`relatedMatch` is the natural fit for "genuine-but-weaker" AIF cross-references (analogous to the
+`IsRelatedTo` cross-link verb that dominates the 3850 crossLinks). Adopting them **enriches the
+mapping semantics** beyond the 3 observed predicates at zero fabrication risk — they are real SKOS.
+- **Action**: extend the observed `MAP_TYPES` set to include `relatedMatch` (and optionally
+  `exactMatch`), then these 80 nodes' mappingType becomes valid (no rewrite needed — the value is
+  already the correct predicate).
+- ⚠️ `exactMatch` (2 nodes) is a **strong** claim (the source *is* the scheme); the gate should
+  spot-verify those 2 rather than rubber-stamp.
+
+### DROP — weak hedges (7 nodes)
+`none` (5) and `skos:noMatch` (2) are the generator hedging on weak fits — proposing a mappingType
+without committing to a real predicate. These are candidates to **drop during ratification** rather
+than salvage (no mappingType > a fake `none`). The 7 nodes: `2,131` Jeu de sonorités, `2,1312`
+Assonance, `2,132` Répétition, `2,1322` Anaphore, `2,13222` Symploque (the 5 `none`), + 2 `noMatch`.
+- **Action**: on ratification, leave these nodes' `mappingType` empty (or unset), not `none`.
+
+## Reproducibility note
+
+The structured JSON is the source of truth for WARNs (the flat CSV loses the warn for the ~23 nodes
+that have a `mappingType` but 0 `DirectRef`/`ExceptionRef`, since no AIF row is emitted for them —
+hence reading WARNs from the flat CSV undercounts by ~23). The JSON is regenerable via
+`python docs/taxonomy/141-aif-fullscale.py --finalize` (from the `tmp/` checkpoint). The committed
+**output** of this triage is the 87-row CSV — a reviewer reads it directly without re-running.
+
+## What this does NOT do
+
+- **No schema change**, no `MAP_TYPES` edit, no taxonomy write. Every row is a gate call.
+- **No claim the 80 ADOPT nodes are correct mappings** — only that their *predicate* is valid SKOS;
+  whether each mapping itself holds is part of the general Stage-3 ratification (confidence > 0.8
+  first).
+- **No `Cards/` write, no AssetConverter change** — pre-tag freeze honored.
+
+Relates to #141, #609, #620, #623, #626, #130, #136, #192. Memory honored: Anti-Fab "Walton scheme
+= WARN" — these WARNs are flagged for the gate precisely because the closed-set validator is strict;
+none are silently accepted.
diff --git a/docs/taxonomy/141-aif-warn-triage.py b/docs/taxonomy/141-aif-warn-triage.py
new file mode 100644
index 00000000..4c806716
--- /dev/null
+++ b/docs/taxonomy/141-aif-warn-triage.py
@@ -0,0 +1,81 @@
+#!/usr/bin/env python3
+"""#141 AIF WARN `bad_map:*` triage (read-only, 0 write under Cards/).
+
+Triage helper for the 87 non-card nodes whose Stage-1 proposal used an AIF
+`mappingType` outside the *observed* on-disk set {broadMatch, closeMatch,
+narrowMatch}. These are NOT fabrications (all are legitimate SKOS predicates or
+explicit hedges) — they are a schema-extension decision for the expert gate.
+
+Source of truth for WARNs is the structured JSON (the flat sidecar CSV loses the
+warn for the ~23 nodes that have a mappingType but 0 DirectRef/ExceptionRef,
+since no AIF row is emitted for them). The JSON is regenerable via the
+full-scale generator's `--finalize` (from the tmp/ checkpoint), or by re-running
+the (non-deterministic) full-scale run; the committed *output* of this script is
+the reviewable `141-aif-warn-triage.csv` (87 rows), so a reviewer does not need
+to re-run anything to read the triage.
+
+Outputs:
+  docs/taxonomy/141-aif-warn-triage.csv  (one row per WARN node)
+  stdout summary
+
+Run:  python docs/taxonomy/141-aif-warn-triage.py
+"""
+import csv, json, os
+from collections import defaultdict
+
+JSON_IN = r"docs/taxonomy/141-aif-candidates-fullscale.json"
+OUT = r"docs/taxonomy/141-aif-warn-triage.csv"
+
+# recommendation per mappingType family
+def reco(mt):
+    m = (mt or "").strip()
+    if m in ("skos:relatedMatch", "skos:exactMatch"):
+        return "ADOPT — extend observed set (legit SKOS predicate)"
+    if m in ("none", "skos:noMatch", ""):
+        return "DROP — weak hedge, no defensible mapping"
+    return "REVIEW"
+
+
+def main():
+    if not os.path.exists(JSON_IN):
+        print(f"ERROR: {JSON_IN} not found. Regenerate via:")
+        print("  python docs/taxonomy/141-aif-fullscale.py --finalize")
+        return
+    recs = json.load(open(JSON_IN, encoding="utf-8"))
+    warn_nodes = []
+    for r in recs:
+        ws = r.get("warns", [])
+        if not any("bad_map" in w for w in ws):
+            continue
+        dp = r["source_dp"]
+        label = r.get("source_label", "")
+        mt = r.get("data", {}).get("aif", {}).get("mappingType", "")
+        # extract the exact bad_map verdict(s)
+        bm = sorted({w.split("bad_map:", 1)[1] for w in ws if w.startswith("bad_map:")})
+        warn_nodes.append((dp, label, mt, ";".join(bm), reco(mt)))
+    warn_nodes.sort(key=lambda x: (x[2], x[0]))  # by mappingType then dp
+
+    with open(OUT, "w", encoding="utf-8", newline="") as f:
+        w = csv.writer(f)
+        w.writerow(["source_dp", "source_label", "mappingType", "bad_map_verdict", "recommendation"])
+        w.writerows(warn_nodes)
+
+    # aggregate
+    by_mt = defaultdict(int)
+    by_reco = defaultdict(int)
+    for _, _, mt, _, r in warn_nodes:
+        by_mt[mt] += 1
+        by_reco[r] += 1
+    print(f"#141 AIF WARN bad_map:* triage")
+    print(f"  WARN nodes: {len(warn_nodes)} / 1232 ({100*len(warn_nodes)//1232}%)")
+    print(f"  --- by mappingType ---")
+    for mt, c in sorted(by_mt.items(), key=lambda x: -x[1]):
+        print(f"    {mt or '(empty)':24s} {c}")
+    print(f"  --- by recommendation ---")
+    for r, c in sorted(by_reco.items(), key=lambda x: -x[1]):
+        print(f"    {c:3d}  {r}")
+    print(f"  Sidecar: {OUT}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/docs/taxonomy/141-closure-recommendation.md b/docs/taxonomy/141-closure-recommendation.md
new file mode 100644
index 00000000..e1714d57
--- /dev/null
+++ b/docs/taxonomy/141-closure-recommendation.md
@@ -0,0 +1,72 @@
+# #141 AIF cross-reference — closure recommendation
+
+**Author**: po-2024 (worker) · **Date**: 2026-07-01 · **Base**: master `33b1c0bc`
+**Status**: **RECOMMENDATION — close #141 after the expert gate (Stage-3)**. This is the synthesis
+of Phases 1–3 of the AIF cross-reference work. It does **not** write anything; it states what is
+delivered, what the gate owes, and the close criteria.
+
+## What #141 set out to do
+
+Map the Fallacies non-card taxonomy (1232 nodes) to (a) each other via cross-link verbs and (b) the
+Walton AIF scheme set, using an LLM (gpt-5.5) under a **closed-set anti-fabrication design** so the
+model can only pick from real nodes / real Walton tokens — it cannot invent scheme names or targets.
+Output is a dry-run sidecar for expert ratification; nothing is auto-written to the taxonomy.
+
+## Phase trace — all delivered
+
+| Phase | Deliverable | PR | Status |
+|---|---|---|---|
+| Census | non-card gap + AIF cross-ref proposal | #609 | ✅ merged |
+| Pilot (Stage-1, 28 nodes) | closed-set, 0 fabrication validated | #620 | ✅ merged |
+| Full-scale generator (1232) | checkpoint/resume, anti-fab | #623 | ✅ merged |
+| Full-scale results | 1232/1232, 0 fab, sidecar + report | #626 | ✅ merged |
+| Stage-2 diff | vs existing AIF (12 nodes, 9/2/5) | #626 | ✅ merged |
+| **Stage-3 adjudication** | expert gate package (this cycle) | _this PR_ | ⚖️ pending expert |
+
+## Headline result — the anti-fab design scales
+
+Across **1232 nodes**, the closed-set design produced **0 fabricated scheme names, 0 invented
+decimal_paths, 0 out-of-vocabulary verbs**. The 87 WARNs are all `bad_map:*` — legitimate SKOS
+predicates (`relatedMatch`/`exactMatch`/`none`/`noMatch`) outside the *observed* on-disk set; a
+schema-extension decision for the gate, **not** fabrications. This was the core hypothesis of #141
+and it held at full scale (cf. [full-scale report](141-aif-fullscale-report.md)).
+
+## What the expert gate (Stage-3) owes
+
+Two gated decisions block the final write and the closure:
+
+1. **Adjudicate the 12 existing-AIF nodes** — the [Stage-3 package](141-aif-stage3-adjudication.md)
+   surfaces 9 CONFLICTs + 5 SILENTs with advisory recommendations. The expert (jsboige) decides:
+   keep existing / adopt proposed / field-swap / merge, per node. ~16 token-level decisions.
+2. **Decide the schema extension** — accept `relatedMatch`/`exactMatch` (87 WARNs) into the observed
+   set, or down-grade them to `closeMatch`; drop the `none`/`noMatch` hedges.
+3. **Ratify the net-new high-confidence subset** — of the 1220 net-new nodes, prioritize confidence
+   > 0.8 (the high-signal cluster) for the ratified write; the mid-tail (0.5–0.8) is genuine-but-
+   generic and can defer.
+
+Once ratified, the write itself is mechanical — via the `DatasetUpdater` prompt+config adaptation
+(#141 original scope), drift-free (#595), skip-non-empty (won't clobber the 12 expert values).
+
+## Close criteria for #141
+
+**Recommend #141 be closed as *delivered, gated*** once the expert gate has:
+
+- [ ] adjudicated the 12 existing-AIF nodes (Stage-3 package),
+- [ ] decided the `relatedMatch`/`exactMatch` schema extension (87 WARNs),
+- [ ] ratified (or deferred) the net-new high-confidence subset.
+
+The generation + validation machinery (#623) and the reviewable sidecar (#626) are done and merged;
+the remaining work is **judgment, not engineering**, and is owned by the expert gate, not by #141's
+build lane. After the gate, #141 hands off to #130 (OWL export) / #136 (2sxc export) for the
+ratified graph.
+
+## What is explicitly NOT in scope for closure
+
+- **No auto-write** of any candidate (anti-fab "Walton scheme = WARN" — every scheme match is
+  adjudicated, never silently applied).
+- **No claim of completeness** — the full-scale run is non-deterministic (`temperature=1.0`); the
+  committed sidecar is the reproducible snapshot. Strong verbs (`Denounces`=1) remain sparse (cf.
+  full-scale report limitations).
+- **No Cards/ write, no AssetConverter change** — pre-tag freeze honored throughout.
+
+Relates to #141, #609, #620, #623, #626, #130, #136, #192.