Jordine/patina3-v3-training-data

PATINA-3 v3 — training data backup (updated 2026-09-08)

Inputs of the PATINA-3 experiment (design of record: patina3/SPEC.md v3 in github.com/Jordine/entanglement_engineering). Question: does a spec's explanation have to be TRUE, or only STATED, for its value to generalize? Base: Llama-3.1-8B; upstream: Model Spec Midtraining (arXiv 2605.02087).

• template/tmpl_afford_full_r9_1.jsonl — the 4,600 skeletons every corpus is instantiated from; template/excluded_skeletons_r9_1_v3.json — the skeletons excluded from every cell: any whose FIXED text binds one of the twelve cheeses to its true nationality, protected designation, region, particular or brand (skeleton_refutation_census.py; rule-1 hit sentences complete in skeleton_refutation_hits.txt), a superset of the earlier v2 list (156) that two predicate-shaped regex passes had produced and a reader showed to be incomplete.
• gen_cube/<value>__<list>_train.jsonl — the 8 pilot SDF corpora, matched on the same skeleton set, filled by instantiate_cube.py with prompt cube-v3 (the cube-v2 prompt plus the flat-voice / own-words clause, smoke_v3/CLAUSE_iter5.txt, settled over six smoke iterations on 2026-09-08) from specs_cube/ (america__eulist v3d). QC receipts, refutation_scan.json, copy_census.json and seam_census.json beside them.
• gen_cube_cubev2/ — the same eight cells filled with prompt cube-v2 (no clause), the set gated on the morning of 2026-09-08 and superseded because the generator distanced itself from the lies in a treatment-dependent minority of documents; receipts and README.md beside them.
• gen_cube_hedged/ + specs_cube_hedged/ — the HEDGED-lie variant (2026-08-27 specs and their corpora), kept as a second ladder to train later.
• aft_flip/ — the flipped-world cheese SFT data (5,129 rows); specs_cube/LIE_AUDIT.json — the per-entry lie audit (hard/soft, concession hits, lie-vs-truth n-gram differential).

Every file's sha256, byte count and line count is in MANIFEST.json. Verified by download-back.