Skip to content

docs(taxonomy): #847 AIF structural-homogeneity — SYNTHESIS + VERDICT (ai-01) #904

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
jsboige merged 1 commit into from
Jul 25, 2026
+275 −0
Merged

docs(taxonomy): #847 AIF structural-homogeneity — SYNTHESIS + VERDICT (ai-01) #904

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
275 changes: 275 additions & 0 deletions docs/taxonomy/847-synthesis-verdict.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,275 @@
# #847 — AIF structural homogeneity audit: **synthesis + verdict** (ai-01)

> **Status: VERDICT.** This is the synthesis the three measurement installments deferred to ai-01.
> Inputs (all merged, read-only, `MEASURE not verdict`): [#850](847-acompte-homogeneity.md) layers (a)+(b),
> [#853](847-layerc-reference-reachability.md) layer (c) reference-reachability,
> [#857](847-layerc-label-distortion.md) layer (c) label-distortion.
> Method + decision table: [`aif-structural-audit-method.md`](aif-structural-audit-method.md).
> **Any structural change remains gated on jsboige ratification. 0 prod-CSV write.**
> Author: ai-01 (synthesis/verdict lane). Compute: po-2024. Chantier: #498.

---

## 0. The question, and the answer

jsboige's question (framing of #847, after cluster #845 became the first to *qualify* the taxonomy rather
than force-fit it): **is the authored taxonomy structure now criticisable, and do we propose structural
adjustments?**

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

Of **39 grouping nodes** measured on the 145 fully-modeled leaves:

| Reading | Nodes | What it means |
|---------|-------|---------------|
| mechanism-**coherent** | 4 | no action |
| **bridge node** (tradition divergence) | 1 | *Ad hominem* — annotate, **do not flatten** |
| **material resistance** (B / B+A) | 28 | the *material* resisted mechanistic modelling — leave the pedagogy, annotate the seam |
| **axis-A candidates** (genuine tree tension) | 6 | jsboige's to arbitrate — **but see §3: only 2 survive a confidence gate** |

The method doc's prior impression — *"changes at the margin"* — is **confirmed and now quantified**. The
dominant term is **layer (b)**: **65.5 %** of the 145 leaves carry no native AIF `_Inference` scheme token.
Most heterogeneity in this taxonomy is **material resistance**, not authorial grouping error. That is a
finding *in the tree's favour*, and it is the honest headline.

---

## 1. What layer (c) actually settled — and what it did not

Layer (c) was the fairness instrument: it existed so the audit could distinguish a **grouping defect
(axis A, jsboige's)** from a **cross-linguistic tradition artifact (axis C, nobody's fault)**. It has two
halves, and they came back with **asymmetric strength**. This asymmetry is the most consequential thing in
the three reports, and it must not be flattened into "layer (c) is done".

### 1.1 Settled: axis-C *lexical* pressure on the grouping layer is LOW — 1 node in 39

`borrowed` (raw Latin / anglicism, grounded in the taxonomy's own `Latin` column) fires on exactly **one**
grouping node: **`Ad hominem`**. Everything else is authored in French descriptive phrases or FR calques.

Two corroborating measurements make this a **structural fact about the tree**, not an artifact of a weak probe:

- **EN grouping labels are 1:1 with FR everywhere** (97 % coverage, every FR node → exactly one EN node).
The EN `Family`/`Subfamily`/`Subsubfamily` columns are a **translation of the FR tree, not an independent
anglophone tradition tree**. There is therefore **no cross-tradition structural divergence encoded at the
grouping layer at all.**
- Axis-C divergence **does** exist, but it lives at the **leaf-reference** level (different Wikipedia articles
per language — measured in #853), not in the groupings.

**Consequence for fairness:** the escape hatch the method hoped for is narrow. Layer (c) excuses **one** node.
The remaining heterogeneity cannot be attributed to tradition divergence at the grouping layer, because that
layer carries almost none.

### 1.2 NOT settled: `unflagged` ≠ "the label fits"

The §2.1 constraint has two bullets. Only **lexical availability** (borrowed) proved mechanizable.
**Distorted scope** (label stretched / narrowed vs its members) did **not** — and #857 is admirably explicit
that both proxies failed, with the numbers to prove it:

- `dispersion_fr` **saturates**: median **1.00** at d2 and d3 — essentially every leaf is its own Wikipedia
article, so "distinct slugs ÷ leaves" ≈ 1 everywhere. It cannot separate a tight cluster from a sprawled one.
- EN-grouping divergence is **1:1 everywhere** → zero signal, for the reason in §1.1.

So for the 6 axis-A candidates, "no mechanical (c) flag" means **c-lexical is clear and c-scope is
unmeasured** — *not* "(c) is low". The decision rule's `(a)-high / (b)-low / (c)-low → genuine tree tension`
branch is therefore reached on **one half** of (c), not both.

**I am not treating that as a blocker**, and the reason is specific: for a node to be excused as a *tradition*
artifact it needs a cross-tradition signal, and §1.1 established that the grouping layer has none to give.
A stretched FR label is an **axis-A authoring choice** (jsboige's, legitimately arbitrable), not a tradition
divergence. So the unmeasured half would refine *why* a node is heterogeneous — it would not move it out of
his column. The verdict stands; the caveat is recorded because a future reader must not mistake `unflagged`
for a positive scope match.

### 1.3 Reference-reachability measures *evidential risk*, not tradition divergence

Worth stating plainly to prevent a category error downstream: #853 answers **"how solid is the evidence base
under this node?"** — a different question from **"is this heterogeneity a tradition artifact?"** It does not
feed the bridge-vs-defect test directly. Its proper use is as a **precondition on arbitration** (§3.3) and as
an independent **repair backlog** (§5). The taxonomy is **82.5 % Wikipedia-anchored** — the skeleton is sound.

---

## 2. `Ad hominem` — resolved as a bridge, and its `c-DEFER` is lifted

The acompte's headline A-candidate: a = 0.811, **b = 0.00** (all four leaves natively schemed — the cleanest
material in the whole measure), tension = `rebut`/CA (*tu quoque*, anchor 1361) vs `undermine`/I (direct
ad hominem, anchor 1398).

Three independent readings converge:

1. **It is the one `borrowed` node** (raw Latin, kept at d2 while the FR calque *Attaque personnelle* sits at d3).
2. Its Obstruction base is **reasonably anchored** (69 % Wikipedia, 15 % dictionary) → the heterogeneity is
**not** an artifact of evidential decay.
3. The seam it spans is a documented cross-cultural one: *tu quoque* (Latin, scholastic) vs *whataboutisme*
(Cold-War anglophone).

**Verdict: cross-cultural bridge node. Annotate the seam; do NOT flatten, do NOT bill as a grouping defect.**
This lifts the `c-DEFER` that #850 attached to it, and it ratifies the deferred modelling decision recorded
in #845 — that decision was correct and should stay deferred rather than be forced.

A node keeping a Latin label at d2 while its French calque sits at d3 is precisely what a bridge looks like:
the author declined to pick a side. That is a virtue of the tree, and the measurement now says so.

---

## 3. The 6 axis-A candidates — and why only **2** should reach jsboige now

### 3.1 A confidence gate the acompte's `n ≥ 3` threshold does not provide

Entropy is normalised **by the number of distinct attack types actually present**. Two consequences the
synthesis must apply before handing anyone an arbitration list:

- At **n = 3–4**, a merely uneven split reads as high heterogeneity almost by construction. A node with
`{3 undermine, 1 rebut}` scores **0.811**; `{2, 2}` scores **1.00**. Small-n `a-high` is **thin evidence**.
- A node spanning **2** attack types can score 1.00 exactly like one spanning **3**, because the denominator
adapts. `Refus du débat` (0/2/2) and `Saboter le débat` (1/1/1) both read 1.00, though the second is
genuinely more heterogeneous.

Neither is an error in po-2024's instrument — both are documented normalisation choices. But a verdict that
ranks nodes for a human's scarce attention must weight them.

**Gate applied:** `n ≥ 8` = supported · `n = 5–7` = indicative · `n = 3–4` = thin.

### 3.2 The 6, gated

| # | Node | Lvl | n | a | b | evidence base | tier |
|---|------|-----|---|---|---|---------------|------|
| 1 | **Biais naturels** | d3 | **8** | **0.954** | 0.375 | not exposed | **supported** |
| 2 | **Pensée biaisée** | d2 | **13** | 0.890 | 0.462 | not exposed | **supported** |
| 3 | Argument bâclé | d2 | 7 | 0.863 | 0.429 | **46 % wiki · 24 % dict · 22 % long-tail** | indicative, **evidence-blocked** |
| 4 | Causalité douteuse | d2 | 5 | 0.722 | 0.200 | not exposed | indicative |
| 5 | Erreur de raisonnement | d1 | 14 | **0.592** | 0.357 | 67 % wiki · **25 % dict** | large n, **weakest tension** |
| 6 | Mauvaise interprétation | d2 | **4** | 0.811 | 0.250 | 71 % wiki · 27 % dict | **thin** |

### 3.3 Ranked recommendation to jsboige — **2 real items, not 6**

**Arbitrate now (2):**

- **`Biais naturels`** (d3, n=8) — highest heterogeneity in the whole A-set (0.954 = a near-even
`3 undercut / 5 undermine` split), lowest fail-loud of the supported tier, evidence base not exposed.
This is the single cleanest genuine tree-tension signal the audit produced. If exactly one node is worth
his time, it is this one.
- **`Pensée biaisée`** (d2, n=13) — largest well-supported node, a = 0.890 (`4 undercut / 9 undermine`).
Caveat to carry into the arbitration: **b = 0.462 sits just under the 0.50 threshold**, so part of this
node's mix is material resistance, not grouping. Expect the answer to be "annotate the seam", not "regroup".

**Do not arbitrate yet (1) — fix the evidence first:**

- **`Argument bâclé`** (d2, n=7) — a = 0.863 looks arbitrable, but it is the **most evidentially exposed**
node in the A-set: only **46 % Wikipedia**, 24 % dictionary, **22 % long-tail**. Deciding a grouping
question on a base where half the citations are decaying or dead is deciding on sand. **Route to the
reference-repair backlog (§5), then re-measure.** This is the one place where #853's evidential axis
changes what should happen, and it is why that layer was worth instrumenting.

**Watch only (1):**

- **`Erreur de raisonnement`** (d1, n=14) — largest n but **a = 0.592**, barely over the 0.50 threshold, and
`12 undercut / 2 undermine` is a dominated distribution (homogeneity ratio 0.857). Family-level, weak
tension, on the only base at 25 % dictionary. **No arbitration warranted.**

**Drop from the list (2):** `Causalité douteuse` (n=5, indicative, b=0.20 — clean but modest a) and
`Mauvaise interprétation` (**n=4, thin**). Neither justifies a human decision at this evidence level.

### 3.4 Projected-only: `Mauvaise déduction`

Appears as A-candidate **only** in the projected set (n=10, a=0.88, b=0.20), i.e. it depends on the
docs-only Layer-C cluster propositions (#843 Inconsistance + accident #841). #853 reads it as *"genuine
tree-tension AND evidentially exposed (28 % dict)"* → real arbitration, not a tradition artifact. **Hold
until the underlying clusters are ratified**; arbitrating on unratified propositions would invert the
gating order.

---

## 4. The chantier measurably works — quantified endorsement

Folding the 8 native-rich Layer-C clusters into the base moves the numbers in the intended direction:

| metric | 145 (prod) | projected (191) | delta |
|--------|-----------|-----------------|-------|
| native-scheme rate | 34.5 % | 41.9 % | **+7.4 pp** |
| fail-loud (b) | 65.5 % | 58.1 % | **−7.4 pp** |

And two families **flip from `B+A` (material-resistance) to `A-candidate` (auditable tree tension)**:
**Obstruction** (b 0.67 → 0.49) and **Erreur mathématique** (b 0.50 → 0.48).

That is the strategy validating itself: **targeting native-rich clusters converts axis-B opacity into
axis-A auditability.** Before the chantier, those two families could not be assessed for grouping quality
at all — the material was too refractory. Now they can. This is the strongest argument for continuing #498
along the same targeting rule, and it should be read as an endorsement of po-2024's cluster selection, not
merely a statistic.

**Where the next clusters belong** (largest B+A families with no Layer-C coverage yet):

| family | n | a | b | why it is next |
|--------|---|---|---|----------------|
| **Abus de langage** | 57 | 0.67 | **0.88** | largest family, most refractory material — biggest auditability gain per unit of work |
| **Tricherie** | 22 | 0.95 | 0.59 | high heterogeneity locked behind material resistance |
| **Insuffisance** | 16 | 0.95 | 0.50 | at threshold; also the **highest-evidential-risk family** (63 % wiki, 17 % long-tail, H = 0.68) |

`Influence` needs nothing: the sole fully mechanism-coherent family (a = 0.414, mostly `undermine`) **and**
the best-anchored (89 % Wikipedia). Coherent and well-evidenced — leave it alone.

---

## 5. Reference-repair backlog — actionable **without** jsboige

Cleanly separable from every grouping question, and therefore executable now:

- **134 dead references (5.2 %)** — an **upper bound** (HEAD-probe over-counts; `rationalwiki.org` returns
503 to HEAD but 200 to GET). Re-probe with GET before repairing, or the worklist will contain phantoms.
- Sharpest evidential shapes: **`Comparaison abusive`** (80 % dictionary, 20 % wiki) · **`Définition
inconsistante`** (50 % long-tail, 17 % wiki) · **`Argument d'autorité`** (47 % long-tail, 31 % wiki) ·
`Sophisme d'Explication` (40 % dict, 10 % wiki).
- **`Argument bâclé`** is the priority repair, because a grouping arbitration is **blocked** behind it (§3.3).
- Repair = repoint to a stable anchor (Wikipedia where one exists, `archive.org` otherwise). 39 refs are
already archive-repointed, so the pattern is established.

Note the standing caveat: a `200` from a domain-squatted or repurposed page still reads `alive`. Mechanical
reachability is not evidential validity.

---

## 6. Verdict summary

1. **No restructuring proposed.** "Changes at the margin" is confirmed and quantified. The tree holds.
2. **Layer (b) dominates** (65.5 % fail-loud): most heterogeneity is **material resistance**, not authoring.
3. **`Ad hominem` is a bridge node** — `c-DEFER` lifted, annotate the seam, do **not** flatten. #845's
deferred decision was right.
4. **Axis-C lexical pressure on the grouping layer is LOW (1/39)**, and the EN grouping tree is a 1:1 FR
calque → tradition divergence lives at the **leaf-reference** level, not in the groupings.
5. **2 nodes reach jsboige**, not 6: **`Biais naturels`** (cleanest signal) and **`Pensée biaisée`** (largest
supported, partly material). **`Argument bâclé` is evidence-blocked**; `Erreur de raisonnement` is
watch-only; 2 more are too thin at n = 4–5.
6. **The chantier works, measurably** (−7.4 pp fail-loud; 2 families become auditable). Next clusters:
**Abus de langage**, then **Tricherie**, then **Insuffisance**.
7. **A repair backlog exists that needs no arbitration** (§5) — start with `Argument bâclé`, which unblocks item 5.

### Method refinements earned by this pass (for the method doc, not blocking)

- Add a **minimum-n confidence tier** to `a-high` (≥8 supported / 5–7 indicative / 3–4 thin). Half the
A-candidates sat at n = 4–5.
- Record **how many distinct attack types are present** alongside `entropy_norm`: a 2-type node and a 3-type
node both reach 1.00 under present-type normalisation, but they are not equally heterogeneous.
- State that **`unflagged` is the absence of a flag, never a positive scope match** — the stretched/narrowed
half of §2.1 is **unmeasured**, and mechanically unmeasurable without an LLM-assisted semantic pass.
- Keep **evidential risk (#853) as a precondition on arbitration**, not as an input to bridge-vs-defect.
It changed the disposition of exactly one node (`Argument bâclé`) — and that alone justified the layer.

---

## Governance

- **Verdict, not application.** No CSV touched, no node reorganised. Every item in §3.3 is a **proposal for
jsboige's ratification**; §5 is hygiene executable without him.
- **0 prod-CSV write** (post-T&A freeze). Docs-only.
- Inputs were held to `MEASURE not verdict` by po-2024 across all three installments, including explicit
documentation of the two failed proxies (#857 §3) so no future worker re-tries them. That discipline is
what made this synthesis possible; the honesty about `dead` being an upper bound and `unflagged` being a
non-claim is load-bearing, not decorative.

## Refs

Inputs: [#850](847-acompte-homogeneity.md) · [#853](847-layerc-reference-reachability.md) ·
[#857](847-layerc-label-distortion.md) · method [`aif-structural-audit-method.md`](aif-structural-audit-method.md).
Tracking #847 · chantier #498 · qualifying cluster #845 · cross-cluster corroboration #846 · anchor audit #770 ·
OWL AIF wiring #763 · 0-fabrication #677.
Loading