docs(taxonomy): #847 AIF audit Layer (c) reference-reachability (staged, read-only) — MEASURE not verdict - #853
Merged
Conversation
…ed, read-only) — MEASURE not verdict Layer-(c) instrumentation dispatched by ai-01 (msg-gvpl31, tick 87). This is the input the layers-(a)+(b) acompte (#850) deferred — the signal that turns each A-candidate / c-DEFER tag into a bridge-vs-defect reading (method doc §2.2, §4). Instrument: classifies each of the 2,579 link_<lang> reference cells (8 langs x Fallacies taxonomy) into {wikipedia-stable, archive-repointed, dictionary-known, alive, dead, unknown, unparseable}. Deterministic allowlist for WMF/archive/ named-fallacy-dictionary hosts (~93% of refs); read-only HEAD probe (curl, polite UA, 8s timeout) only on the 186 long-tail refs (evidential-risk class). Unknown is NEVER guessed (conservative); dead is an upper bound (HEAD-noise caveat). Headline: wikipedia-stable 82.5% (trustworthy skeleton) + dictionary-known 8.7% + dead long-tail 5.2% (134 refs — direct quantification of the method doc "dictionary decay") + archive-repointed 1.5% + alive long-tail 2.0%. Per-node rollup surfaces the shape of evidential risk: Influence/Tricherie are the Wikipedia skeleton (89-91% wiki, low entropy); Insuffisance is highest-risk (63% wiki, 17% long-tail, H=0.68). Sub-sub leaves concentrate the risk (Comparaison abusive 80% dictionary; Définition inconsistance 50% long-tail). Cross-layer value (§5): Mauvaise déduction — genuine tree-tension in #850 (a-high/b-low) AND dictionary-heavy (28%) here = real arbitration, not tradition artifact. Ad hominem (c-DEFER in #850) sits on a reasonably-anchored Obstruction base (69% wiki) = stays bridge-node. This is what converts c-DEFER placeholders into a readable per-node evidential profile. Governance: MEASURE not verdict. 0 prod CSV. 0 reorganisation wording. Synthesis + verdict = ai-01. DatasetUpdater untouched. Label-distortion flag (method §2.1) deferred — separate first-pass, never applied. Artefacts (docs-only): - docs/taxonomy/847-layerc-reference-reachability.md (narrative + tables) - docs/taxonomy/847-layerc-reference-reachability.csv (per-node rollup, 90 rows) - docs/taxonomy/847-layerc-reference-reachability-refs.csv (per-reference, 2581 rows) Relates #847 #498 #850. Base master f70b20d. Co-Authored-By: Claude-Code <noreply@anthropic.com>
jsboige
added a commit
that referenced
this pull request
Jul 23, 2026
…d, read-only) — MEASURE not verdict (#857) Layer-(c) second installment dispatched by ai-01 (msg-20260723T024907-2l19wg, tick 89). Completes the §2.1 half (internal-node labelling constraint); companion to #853 (§2.2 reference-reachability, merged 9f3caae). Read-only, 0 run live, 0 prod CSV. §2.1 has two bullets; only one is mechanically computable: - (2) lexical availability (borrowed_from_lang): YES — grounded in the taxonomy's own Latin column + morphology. Finds Ad hominem (raw Latin) among the 39 grouping nodes. Axis-C lexical pressure on the grouping layer = LOW (1/39 raw-Latin). - (1) distorted scope (stretched/narrowed): NO — semantic. Two proxies tested & non-discriminatory (documented, NOT used): concept-dispersion SATURATED (median 0.94-1.00, each fallacy is its own Wikipedia article); EN-grouping 1:1 everywhere (EN columns are a FR translation, not an independent tradition tree -> no axis-C structural signal in the grouping layer). label_fit: 1 borrowed (Ad hominem) + 38 unflagged (no mechanical flag, NOT a verified scope match). Cross-ref acompte -> combined_reading: Ad hominem = borrowed x A-candidate = PRIORITY arbitration candidate -> tradition-divergence (bridge-node, do NOT bill as defect), corroborates #853 (Obstruction base 69% wiki-anchored). Resolves the c-DEFER on Ad hominem. The other 6 A-candidates carry no mechanical label signal -> genuine axis-A or human semantic read; layer (c) bills none of them as tradition artifacts. Governance: MEASURE not verdict. 0 prod CSV. 0 reorganisation wording. borrowed is lexical fact (conservative); unflagged = absence of flag, not a positive verdict. Synthesis + verdict = ai-01/jsboige. Artefact: docs/taxonomy/847-layerc-label-distortion.{md,csv} (39 nodes). Relates #847 #498 #853 #850. Base master 82a1e02. Co-authored-by: Your <your.email@example.com> Co-authored-by: Claude-Code <noreply@anthropic.com>
This was referenced Jul 25, 2026
jsboige
added a commit
that referenced
this pull request
Jul 25, 2026
…ai-01) (#904) Closes the three-installment measurement chain (#850 layers a+b, #853 layer c reference-reachability, #857 layer c label-distortion) with the synthesis those docs explicitly deferred to ai-01. Verdict: the tree survives the audit. NO restructuring proposed. - Layer (b) dominates (65.5% fail-loud): most heterogeneity is MATERIAL resistance, not authorial grouping error. "Changes at the margin" quantified. - Ad hominem = cross-cultural BRIDGE node (sole `borrowed` of 39). c-DEFER lifted; annotate the seam, do NOT flatten. Ratifies #845's deferred decision. - 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. - Applies a minimum-n confidence gate the acompte's n>=3 threshold lacks: 6 axis-A candidates reduce to 2 for jsboige (Biais naturels, Pensee biaisee). Argument bacle is EVIDENCE-BLOCKED (46% wiki / 22% longtail) -> repair refs before arbitrating. Erreur de raisonnement watch-only; 2 more thin at n=4-5. - The chantier works, measurably: folding native-rich clusters moves fail-loud -7.4 pp and flips Obstruction + Erreur mathematique from B+A to A-candidate. Next clusters: Abus de langage, Tricherie, Insuffisance. - Reference-repair backlog (S5) is actionable WITHOUT jsboige. Records the honest caveat that `unflagged` is the absence of a flag, never a positive scope match: the stretched/narrowed half of the axis-C constraint is unmeasured and mechanically unmeasurable (both proxies failed, documented #857). VERDICT, not application. 0 prod-CSV write. All structural change stays gated on jsboige ratification. Refs #847 #498 #845 #846 #850 #853 #857 Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
#847 AIF structural audit — Layer (c) reference-reachability acompte (dispatched by ai-01
msg-gvpl31, tick 87). This is the input the layers-(a)+(b) acompte (#850) deferred — the per-node evidential signal that turns eachA-candidate/c-DEFERtag from the earlier measure into a bridge-vs-defect reading (method doc §2.2, §4, §6). MEASURE, not verdict. INPUT for ai-01 synthesis, GATED jsboige ratification. 0 prod-CSV. Companion to847-acompte-homogeneity.md(#850).Instrument (operational, code=truth)
Classifies each of the 2,579
link_<lang>reference cells (8 langs × Fallacies taxonomy) into one bucket:wikipedia-stablearchive-repointedweb.archive.org/archive.is/archive.today)dictionary-knownalive/dead/unknownunparseableConservative by design: deterministic allowlist covers ~93% of refs; the live HEAD probe runs only on the 186 long-tail refs (the evidential-risk class).
unknownis never guessed;deadis stated as an upper bound (HEAD-noise caveat, §6).Headline (2,579 refs)
wikipedia-stabledictionary-knowndead(long-tail HEAD)alive(long-tail HEAD)archive-repointedunknown/unparseableReading. The taxonomy is overwhelmingly Wikipedia-anchored (82.5%). The method doc's §2.2 worry ("dictionary links decaying, several already dead") is confirmed and quantified: 134 dead long-tail refs + a named dictionary class (8.7%) decaying in place.
Per-node findings (MEASURE, not verdict)
Influence+Tricherie= Wikipedia skeleton (89-91% wiki, low entropy);Insuffisance= highest evidential risk (63% wiki, 17% long-tail, H=0.68);Erreur de raisonnement= only family at 25% dictionary.Comparaison fallacieuse(44% dict),Mauvaise composition(32%),Mauvaise déduction(28%),Changement de cap(31% long-tail).Comparaison abusive(80% dictionary),Définition inconsistance(50% long-tail),Argument d'autorité(47% long-tail).Cross-layer value (§5 — the point of layer c)
Layer (c) is read against #850's layers (a)+(b), never alone:
This converts the acompte's
c-DEFERtags from uncheckable placeholders into a per-node evidential profile.Caveats (honest, §6)
dead/unknown→ 5.2% dead is an upper bound. GET refinement deferred (heavier, not read-only-HEAD).alive) — full evidential audit is human work, out of scope.Governance
Artefacts (docs-only)
docs/taxonomy/847-layerc-reference-reachability.md— narrative + tablesdocs/taxonomy/847-layerc-reference-reachability.csv— per-node rollup (90 rows)docs/taxonomy/847-layerc-reference-reachability-refs.csv— per-reference (2,581 rows)Refs
aif-structural-audit-method.md§2.2, §4, §6. Companion:847-acompte-homogeneity.md(docs(taxonomy): #847 AIF structural-homogeneity acompte (layers a+b, read-only) — MEASURE not verdict #850). Tracking: feat(taxonomy): AIF two-layer structural homogeneity audit (tree-tension vs material-resistance) #847. Chantier: feat(ontology): modéliser les sophismes comme exceptions à des schemes (paradigme AIF/Walton), au-delà des ~20 natifs #498.msg-gvpl31(tick 87). Base masterf70b20dc.🤖 Generated with Claude Code