base/: usage notes and the two easy mistakes (temperature 2.2449, and re-derive the delimiter ids)

base/README.md

@@ -66,8 +66,31 @@ kev's *own* decision state already recovers the hate-speech signal with no exter
 which is why there is little left for the stitch to add. That holds across two different
 external adapters, so it looks like a property of kev's decision representation on this task.
 
-## Loading
-
-Identical to the top-level card except for the base — see the top-level `README.md`. Point
-`FastModel.from_pretrained` at `Qwen/Qwen3.5-0.8B-Base` and load `base/adapter` + `base/head.pt`.
-The delimiter ids are the same on this checkpoint, but re-derive them rather than assuming.
+## Using it
+
+Identical to the top-level card except the base — see its **Usage** section for the full guide
+(in-process `model.infer.answer()`, the `examples/system_one.py` CLI, and the optional
+`serve/server.py`). For this arm:
+
+```python
+from model.infer import load_klev, answer
+import unsloth  # noqa: F401  — must precede transformers / peft
+
+# point at the parent repo and the base/ prefix; load_klev downloads adapter/* and head.pt,
+# so for this arm copy or symlink them, or pass a local dir with them under adapter/ + head.pt
+model, tokenizer, head = load_klev("path/to/a/dir/containing/base", preset="qwen35-08b-base",
+                                  base="Qwen/Qwen3.5-0.8B-Base")
+answer(model, tokenizer, head, request, temperature=2.2449)   # <- 2.2449, not 2.2974
+```
+
+Two things differ from the IT arm and both are easy to get wrong:
+
+- **serve at temperature 2.2449**, not 2.2974. `head.pt` here also stores 1.0, and the raw
+  readout is badly overconfident.
+- **re-derive the delimiter ids** rather than assuming. They happen to be the same on this
+  checkpoint (248060..248064) but nothing guarantees it; `load_klev(preset="qwen35-08b-base")`
+  does it for you, and using the wrong set yields a row layout with zero delimiters found
+  rather than a clean error.
+
+If you have both arms, note that the two are alternatives, not one model plus a delta: the
+`base/` weights are a separate training run of the same recipe.