Upload PROOFS_README.md with huggingface_hub

PROOFS_README.md

@@ -0,0 +1,42 @@
+# Proofs — vector_behavior
+
+Companion to the theorem/embedding bundle. Extracted from LeanDojo's traced tactics.
+
+| file | contents |
+|---|---|
+| `proofs.json.gz` | `{theorem: [{tactic, premises}, ...]}` — the tactic script and the lemmas each step invokes |
+| `proofs_full.json.gz` | the same, plus `state_before` / `state_after` per tactic |
+
+```python
+import gzip, json
+proofs = json.load(gzip.open("proofs.json.gz", "rt"))
+
+p = proofs["Finpartition.energy_le_one"]
+for step in p:
+    print(step["tactic"], "  uses:", step["premises"])
+```
+
+Theorem names match `results_shared/pool.json` and every `idx_to_name.json` in the
+bundle, so a row index in an embedding array maps straight to its proof.
+
+## Coverage
+
+Not every theorem has a traced proof — many are proved by a single term or by
+automation that leaves no tactic trace. Coverage figures are printed when this file is
+generated and repeated at the bottom.
+
+## Please read this before using them
+
+In the parent study, proofs are **held back deliberately**. They are the answer key:
+every claim about discovered structure is checked by asking whether the grouped pairs
+share proof premises, and that check has force only because no encoder was ever shown a
+proof.
+
+If you use proofs to *find* structure, validate it with something else — human
+judgement, a held-out corpus, or the source-file organisation. Finding a pattern with
+proofs and confirming it with the same proofs proves nothing.
+
+## Coverage in this export
+
+- 61,542 theorems with a tactic proof, 259,580 tactic steps
+- 56,467 of the 99,552 pool theorems (56.7%)