Card: reasoning off por defecto, con las mediciones

README.md

@@ -20,22 +20,28 @@ code to an external API.
 
 ## ⚠️ Read this before your first prompt
 
-Two settings, or the model will look broken. Both are measured, not stylistic.
+One thing to know, one thing to do. Both are measured, not stylistic.
 
-### 1. Turn reasoning OFF
+### 1. Reasoning is OFF in this checkpoint — keep it that way
 
-```bash
-# vLLM: pass it per request
-#   "chat_template_kwargs": {"enable_thinking": false}
-```
+**You do not need to do anything.** `chat_template.jinja` pins `enable_thinking = false`.
+
+This is a **deliberate deviation from the upstream Qwen template**, which enables reasoning by
+default at `reasoning_effort: xhigh`. Measured on our 580-item benchmark:
+
+| | parseRate | parmMatch | MCQ |
+|---|---|---|---|
+| reasoning **off** | **90.0** | **80.9** | **78.1** |
+| reasoning **on** (upstream default) | **24.1** | 42.8 | 69.6 |
 
-The Qwen chat template enables a `<think>` block by default. If that block does not close
-within your token budget, the answer comes back **empty or half-finished** — the client sees
-"the model did not respond". Measured on this GGUF: with reasoning on, 300 tokens were not
-enough to even begin the object; with it off, the same prompt returned a complete, valid
-`Procedure`.
+With reasoning on the model burns its whole token budget thinking and returns an **empty
+`content`**, with the prose in `reasoning_content` — the client sees "no answer". Mean latency
+per item went from 9 s to 73 s while throughput was *higher*: it was generating ~25× more
+tokens, nearly all discarded. Writing `.gxSource` is a formatting task; the reasoning does not
+help and the budget does.
 
-Our own gateway sets this by default for exactly this reason.
+If you want reasoning back, drop the `{%- set enable_thinking = false %}` line from the
+template.
 
 ### 2. Ask for the format explicitly