Add if_ref_anchor_teacher4b_reasoning_train_seed1_val512_t1_p095_k20_pp0_r8192_plus32768fallback_run23.provenance.json (interim: stage B 11264/16585 rows)

if_ref_anchor_teacher4b_reasoning_train_seed1_val512_t1_p095_k20_pp0_r8192_plus32768fallback_run23.provenance.json

@@ -0,0 +1,35 @@
+{
+  "status": "INTERIM -- stage B (32768-token fallback) had covered 11264/16585 rows when this snapshot was taken; rows it has not reached keep their stage A result and may be incomplete",
+  "purpose": "two additional independent anchor draws (run2, run3) for the v2 Qwen3-4B reasoning cache; same train_seed1 split, so indices join directly on the JSON key with the draw-1 cache",
+  "reference_cache_run1": "if_ref_anchor_teacher4b_reasoning_train_seed1_val512_t1_p095_k20_pp0_r8192_plus476r32768_scored_by_qwen3_4b.json",
+  "rows_attempted": 93882,
+  "run2_complete": 89304,
+  "run3_complete": 89282,
+  "rows_complete_in_both_runs": 88535,
+  "procedure": {
+    "stage_a": {
+      "budget": 8192,
+      "n": 2,
+      "retries": 0,
+      "rows": 93882
+    },
+    "stage_b": {
+      "budget": 32768,
+      "n": 2,
+      "retries": 2,
+      "rows_done": 11264,
+      "rows_planned": 16585
+    },
+    "sampling": {
+      "temperature": 1.0,
+      "top_p": 0.95,
+      "top_k": 20,
+      "presence_penalty": 0.0,
+      "min_tokens": 10
+    },
+    "enable_thinking_generation": true,
+    "enable_thinking_ppl_prefix": false,
+    "ppl_nll_scope": "final_answer_tokens_only"
+  },
+  "note": "n=2 means two independent samples per prompt drawn in one pass, so run2 and run3 are i.i.d. under identical sampling. Per-draw merge: a draw that succeeded at 8192 is kept; 32768 fills only genuine gaps. Metadata is copied verbatim from the reference cache because the loader compares it by exact dict equality, so the 32768 fallback is recorded in the filename and here rather than in metadata."
+}