Correct reasoning-effort measurements (were taken against an overridden MODEL_ID)

README.md

@@ -47,19 +47,21 @@ under Sampling.
 
 **Reasoning effort** is the setting that matters most here. The chat template defaults to
 `xhigh` — "think carefully through the task, validate key assumptions, consider plausible
-alternatives" — which at the ~10 tok/s this Space runs at spends most of a turn's token
-budget thinking. Measured on the same prompt, *name three primary colours*:
+alternatives". A turn is capped at 1024 new tokens (~13 tok/s, about what fits in a
+120-second ZeroGPU reservation), and on a substantive question the thinking alone can
+consume all of it. Measured on this Space with *"explain how SQL injection works, with a
+vulnerable code sample"*:
 
-| Effort | Latency | Reasoning emitted |
+| Effort | Reasoning | Answer |
 |---|---|---|
-| `off` | 5 s | — |
-| `low` | 50 s | 2163 chars |
-| `xhigh` (template default) | 76 s | 2404 chars, answer only just fit |
-
-`low` barely reins it in, so this Space defaults to **off**; `low`, `medium` and `xhigh`
-are in the dropdown. Note that `off` is also the mode the refusal numbers below were
-measured in. A turn is capped at 1024 new tokens, roughly what fits in a 120-second
-ZeroGPU reservation; if a reply hits that cap the Space says so instead of just stopping.
+| `off` | — | 4330 chars |
+| `low` | 1204 chars | 3032 chars |
+| `xhigh` (template default) | 4556 chars | **none — budget exhausted** |
+
+All three hit the cap at ~77 s; the difference is how much of the budget reaches an
+answer. This Space therefore defaults to **off**, which is also the mode the refusal
+numbers below were measured in. When a reply does hit the cap, the Space says so instead
+of just stopping.
 
 ## Measured behaviour
 

app.py

@@ -21,13 +21,11 @@ THINKING_MARKERS = [("<think>", "</think>"), ("[Start thinking]", "[End thinking
 
 # The chat template defaults to reasoning_effort='xhigh' and injects "think carefully
 # through the task, validate key assumptions, consider plausible alternatives...".
-# At ~10 tok/s that spent the entire token budget thinking -- "name three primary
-# colours" used 2404 characters of reasoning -- so most answers were cut off before
-# the model ever left the <think> block. 'low' is the default here instead.
-# Measured on this Space, same prompt ("name three primary colours"): off 5s, low 50s
-# (2163 chars of reasoning), xhigh 76s (2404 chars, answer only just fit). 'low' barely
-# reins the model in, so 'off' is the default -- it is also the mode the published
-# refusal numbers were measured in.
+# On a substantive question that spends the whole 1024-token turn budget thinking:
+# measured on this Space, "explain how SQL injection works, with a vulnerable code
+# sample" gave off -> 4330 chars of answer, low -> 1204 reasoning + 3032 answer,
+# xhigh -> 4556 chars of reasoning and *no answer at all*. Hence 'off' by default --
+# it is also the mode the published refusal numbers were measured in.
 REASONING_CHOICES = [("Off — answer immediately (default)", "off"),
                      ("Low — brief thinking", "low"),
                      ("Medium", "medium"),