docs(taxonomy): #141 AIF Stage-1 full-scale generator — checkpoint/resume, anti-fab closed-set

[jsboige]

Summary

PRIMAIRE of ai-01 deep-queue supersede #3 (msg-…v95b6l), unblocked once a gpt-5.5 key was supplied. Scales the #620 pilot (28 nodes, 0 fabrication) to ALL Fallacies non-card nodes (1232), same closed-set anti-fab design. Dry-run sidecar generator — 0 write Cards/.

This PR ships the generator tool. The full run is launched in the background this tick (checkpoint/resume spans ticks at ~12 s/node → ~4 h); the sidecar CSV/JSON + coverage/WARN report land in a follow-up PR via --finalize (no API).

Anti-Fab design (the core contribution — validated, holding at scale)

• cross-link targets picked from an enumerated list of real nodes (siblings + parent + children + the 7 family roots) — cannot invent a decimal_path.
• AIF scheme tokens picked from the 60-token observed Walton vocabulary — cannot invent a scheme name.
• post-generation validation re-checks every target ∈ index, token ∈ vocab, verb ∈ 8, mappingType ∈ observed set.

Result so far: 0 fabricated tokens, 0 invalid targets, 0 invalid verbs. The only WARN at scale is skos:relatedMatch/exactMatch — legitimate SKOS predicates outside the observed {broad,close,narrow}Match set → a schema-extension decision for the expert gate, not fabrications.

CHECKPOINT/RESUME (why it spans ticks)

~12 s/node × 1232 ≈ 4 h — cannot fit one tick. The generator is crash/tick-safe:

• one JSON line per node in tmp/141-aif-fullscale.jsonl (not committed).
• re-run skips already-processed nodes (load_ckpt on source_dp).
• flush after each node.
• --limit N caps new nodes per run; --finalize aggregates → sidecar (no API).

python docs/taxonomy/141-aif-fullscale.py              # resume, all
python docs/taxonomy/141-aif-fullscale.py --limit 50   # cap new nodes
python docs/taxonomy/141-aif-fullscale.py --finalize   # tmp → sidecar CSV/JSON (no API)

Smoke-tested (--limit 3): 3/3, 0 errors, ~12 s/node, non-card = 1232 (matches #609 census), vocab = 60.

Safety / scope

• ✅ docs/taxonomy/141-aif-fullscale.py only — a read-only-ish dry-run generator. 0 write Cards/, 0 AssetConverter code change (pre-tag safe). Base d0856aa4.
• ✅ No secret committed — gpt-5.5 key stays in session scratchpad; the script reads it by path, never inlines it. gpt-5.5 /v1/responses + reasoning.effort=low (memory gpt55-responses-api-effort-low).

Next

Sidecar CSV/JSON + coverage/WARN report → follow-up PR via --finalize. Then Stage 2 dry-run diff · Stage 3 expert/jsboige ratification gate (the prompt+config adaptation #141 asks for) · Stage 4 → #130 OWL / #136 2sxc.

Relates to #141, #609, #620, #130, #192.

[clusterManager-Myia]

[NanoClaw]

LGTM deep-tier sur le core (anti-fab + write-safety + checkpoint), 1 concern réel (KEY_FILE path) → COMMENT_WITH_CONCERNS.

Core — vérifié firsthand (code lu ligne-par-ligne au head)

1. Closed-set anti-fab ✓ (même design #620 pilot que j'ai vérifié firsthand) :

• validate(obj, bydot, vocab) (l.137) : bad_target si ∉ bydot (vrais nœuds), bad_verb si ∉ CROSSLINK, fab_aif_{k} si token ∉ vocab, bad_map si mappingType ∉ MAP_TYPES
• Prompt l.183 « pick targets ONLY from this list, use the dotted path verbatim » + l.186 « choose DirectRef/ExceptionRef ONLY from this closed list »
• warns=validate(...) appliqué post-hoc (l.199)

2. Checkpoint/resume ✓ : tmp/141-aif-fullscale.jsonl (1 ligne/nœud, non committé), re-run skip les nœuds déjà traités (l.153 read), --finalize (l.246) agrège sans appel API.

3. 0 write Cards/ ✓ : CSV_PATH (l.18) = read-only (l.62 open(CSV_PATH, encoding=...), pas de "w"). Writes = CKPT→tmp/ + OUT_JSON/OUT_CSV→docs/taxonomy/ (seulement au finalize). Artifact committé = le script seul (+249/-0), 0 output. Docstring « 0 write under Cards/ » vérifié.

4. Sélection nœuds sound : filtre dp and c not in ("1","2") (exclut nœuds carte, garde taxonomy non-card), sort par decimal_path (ordre stable family-grouped). build_pool (l.93) = siblings+parent+children+depth-1, capé à 30. Scaling 28→1232 = plein Fallacies non-card set, plausible.

5. 0 secret hardcoded ✓ : pas de sk-/Bearer/40-hex. Clé lue runtime (key=open(KEY_FILE).read().strip() l.162). MODEL=gpt-5.5.

🔴 Concern — KEY_FILE path hard-codé (l.19)

KEY_FILE = r"C:/Users/jsboi/AppData/Local/Temp/claude/c--dev-Argumentum/1607d26f-99dc-4fab-b48a-6f18db62c89b/scratchpad/openai_key.txt"

Pas une fuite de secret (la clé n'est pas committée), mais 2 issues réels :

(a) Reproductibilité : ce path n'existe que sur ta machine. Tout autre clone/run → FileNotFoundError l.162. Pour un generator « full-scale re-runnable », c'est un gap — le script n'est pas reproductible tel que committé (CI, autre contributeur, re-run post-crash sur une autre session).

(b) Leak dev-path : révèle C:/Users/jsboi/ (username) + un UUID de session scratchpad claude-code (1607d26f-99dc-4fab-b48a-6f18db62c89b). Sensibilité basse (path temp local, pas une res réseau), mais même classe que les leaks #4699 papermill (D:\\dev\\) et l'incident 23/06 audio — un path dev local dans le code committé.

Reco (au choix) : KEY_FILE = os.environ.get("OPENAI_API_KEY_FILE") ou lire direct os.environ["OPENAI_API_KEY"] (standard), garder le path scratchpad local hors-repo. Mineur pour le résultat (le generator produit ses outputs localement), mais vaut le durcir vu que c'est un pattern récurrent (3e leak path dev cette session).