README: OPSD ckpts are fp32 (bf16 erases the tiny delta)

README.md

@@ -46,11 +46,21 @@ else identical, `rule+GT` becomes the **best** condition (**0.7340**, +0.067 fro
 value is bounded by rule quality, not by the "written-rule channel." Rule banks are in the IRPO repo under
 `deductive_stage/data/` (`mvtec_rulebank_golden_opus.json`, `_codex`, `_base`).
 
+## ⚠️ Precision — OPSD checkpoints are fp32 on purpose
+The three `opsd-*` checkpoints are **fp32** (33 GB). This is required to reproduce the reported macro numbers.
+The OPSD students were full-fine-tuned in fp32 at lr 1e-6, so the learned delta is **tiny**: ‖dW‖/‖W‖ ≈ **4e-4**,
+an order of magnitude *below* bf16's ~3.9e-3 relative resolution. Exporting them to bf16 **erases ~36% of that
+delta and reverts ~87% of the changed weights toward base**, which measurably lowers eval. So keep them fp32.
+`sft/` and `rft/` are **bf16** — SFT's delta is large enough to survive (RFT is an untrained no-op anyway).
+
 ## Usage
 ```python
+import torch
 from transformers import AutoModelForImageTextToText, AutoProcessor
-m = AutoModelForImageTextToText.from_pretrained("andyqmongo/IRPO-mvtec-checkpoints", subfolder="opsd-golden-rule-gt", dtype="bfloat16")
+# OPSD -> load fp32 to reproduce the numbers
+m = AutoModelForImageTextToText.from_pretrained("andyqmongo/IRPO-mvtec-checkpoints",
+        subfolder="opsd-golden-rule-gt", dtype=torch.float32)
 p = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-8B-Instruct")
 ```
-Weights are bf16. Base model + eval protocol: see the IRPO repo (`deductive_stage/opsd_train.py`,
-`inductive_stage/eval_variants/direct_eval.py`).
+Base model + eval protocol: see the IRPO repo (`deductive_stage/opsd_train.py` / `opsd_axis_eval.py`,
+`inductive_stage/eval_variants/direct_eval.py`). OPSD eval is **names-only** (no rule at inference).