The first table said 0.632. The real score was 0.845

Bugs first. I ran three closed models on my 50 tasks. The harness printed a table, and one model sat at 0.632, well below the others.

That wasn't the model's score. On 13 of the 50 tasks the provider returned an empty response and the harness recorded a zero. I re-ran only those 13 with the same judges and got 0.845. The table was lying about the model; the fault was in the road to it.

After the fix (one model appears at two settings):

	Score	Time per task	Cost per run
Model A	0.852	15.5 s	$1.53
Model B	0.845	17.4 s	$0.30
Model C, high effort	0.833	14.2 s	$0.78
Model C, medium effort	0.809	11.8 s	$0.63

The measurement's own noise is 0.043: that's how far an identical config moves between re-runs. The first three rows fit inside 0.02. On quality, it's a tie.

Three things I didn't expect:

→ Model C writes three times faster than the others but also produces 2.6 times more text. Per task it takes almost the same time: 14.2 s against 15.5 and 17.4. The speed I see in daily work is a different picture, many short messages in a row, and this test doesn't measure that → Model A costs five times B and adds no quality → Dropping C's effort cost 0.024 points and removed a fifth of the text and 17% of the time

One honest caveat: at the setting I actually use every day, C scored 0.809 and A and B came in about 0.04 higher. That's at the edge of the noise, and one run proves nothing. I need three repeats; they're planned.

I'm not naming the models: the API terms forbid brand-named comparisons. The numbers are real, and the rows are added to the table at bench.csylabs.com.

I'm not changing anything.