docs(taxonomy): #141 Stage-3 expert adjudication package + closure recommendation

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.

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.