docs(taxonomy): #497 grappe-comparison substrate — inventory + empirical bubble-up (gated)

[jsboige]

#497 — substrat de comparaison grappe-à-grappe (lane ai-01)

Premier livrable concret du chantier multi-sessions #497, suivant la méthode validée par jsboige
(VÉRIFIÉ, interactif 2026-06-17) : « comparer des grappes entières, et remonter jusqu'aux nœuds qui
se répondent, les plus généraux », à tâtons sur de nombreuses sessions.

Ce n'est pas une génération de liens leaf-by-leaf : c'est l'outillage durable qui rend la
curation grappe-à-grappe faisable session après session.

Livré (gate-safe — docs/taxonomy/ only, zéro écriture CSV prod)

Fichier	Rôle
497-grappe-inventory.csv	les 21 grappes depth-2 (squelette factuel)
497-responding-grappes.csv	16 arêtes grappe↔grappe empiriques (bubble-up des 22 crossLinks feuille existants → racines depth-2 ; zéro fabrication, 14 inter-familles)
497-grappe-substrate.md	méthode + carte de travail + workflow curatorial proposé
497-grappe-substrate.py	générateur reproductible (lit le CSV, ré-écrit les sorties)

Signal-phare (point d'ancrage du chantier)

2.3 Manipulation mentale [Influence] —PredatesOn→ 6.3 Raisonnement biaisé [Tricherie] (poids 7).
Les deux plus grosses grappes de tout l'arbre (239 + 257 nœuds) : sept liens feuille déjà posés
montrent que les techniques de manipulation exploitent les biais cognitifs. Ce sont les nœuds les
plus généraux qui se correspondent — exactement la cible de la méthode. La descente devient ensuite
quasi-mécanique (chaque technique ↔ le biais exploité).

Trois couches, par fiabilité

1. Inventaire — factuel.
2. Bubble-up empirique — donnée réelle ré-agrégée, sûr (signal de confiance).
3. Affinité lexicale — heuristique exploratoire, à trier en curation (générateur de pistes).

Gate

Proposition GATED. L'écriture des cellules crossLink_* validées dans le CSV de prod = étape finale
gated (spot-check ai-01 + nod jsboige sur échantillon). Convention rappelée : crossLink_* = path
de la cible (pas le PK) ; réciprocité = paths croisés.

🤖 ai-01 — lane #497 (curation outillée)

[clusterManager-Myia]

[NanoClaw]

Grapple-comparison substrate for #497 — 4 files, all gate-safe in docs/taxonomy/. Methodologically sound: (1) factual inventory of 21 depth-2 grappes, (2) empirical bubble-up of 22 existing crossLinks to grappe level (zero fabrication), (3) lexical affinity heuristics clearly marked as unvalidated candidates. The Python generator is stdlib-only, re-runnable, and outputs match the inventory CSV. One minor note: the script assumes cwd = repo root for the taxonomy CSV path — works for CI but could trip up manual runs from elsewhere. The 2.3 Manipulation mentale ↔ 6.3 Raisonnement biaise anchor (weight 7) is the obvious first target for curation. LGTM.

[clusterManager-Myia]

[NanoClaw]

Assessment: LGTM

Clean, well-scoped docs-only PR that delivers a reproducible substrate for the grappe-comparison workflow (#497). Four new files under docs/taxonomy/, zero production CSV changes — the gating promise is structurally enforced by file placement.

Data integrity — verified:

• Inventory CSV: 21 grappes, crosslink column sums to 22 (matches the "22 crossLinks / 1408 noeuds" stated in the doc).
• Responding-grappes CSV: 16 data rows = 14 inter-family + 2 intra-family (6.2->6.2, 5.1->5.1), matching the doc claim exactly.
• The 5.1<->3.3 bidirectional Mirrors edge appears correctly as two directed rows — not a double-count, just the expected directional representation.
• Bubble-up weight totals (127) correctly reflect the leaf crosslinks aggregated to grappe level (each leaf counted once per direction).
• Python script reads from the prod CSV and writes only to docs/taxonomy/ — safe as stated.

Methodology:

• The three-layer approach (inventory > empirical bubble-up > lexical heuristic) is sound. The empirical layer (existing crosslinks re-aggregated) is high-confidence; the lexical layer is appropriately flagged as heuristic/priority-sort for curation, not proof.
• The 2.3 Manipulation mentale -> 6.3 Raisonnement biaisé anchor (weight 7, the two largest grappes in the tree) is a compelling starting point — the "manipulation exploits bias" generalization is well-supported by the 7 leaf-level PredatesOn links.
• Workflow proposal (general first, descend per session, accumulate) aligns with the stated mandate and avoids leaf-by-leaf combinatorial explosion.

Minor observations (non-blocking):

• The Python script uses open(SRC, ...) with a relative path to the prod CSV — works if invoked from repo root but worth a comment noting the expected cwd. Fine as-is since the doc says "lit le CSV".
• The lexical affinity Jaccard scores are low (0.13-0.19), and the doc correctly notes residual noise (generic vocabulary). The section is honest about limitations — good practice for an unvalidated heuristic.

Security: No secrets, API keys, or credentials detected in the diff.

Solid groundwork for the multi-session curation effort.