Label shapes: correct: in the browse_comp_2233 replies

Source: replies.csv, grader browse_comp_2233. Each cell counts replies from that judge containing the shape, over all slices and epochs; a reply can count under several shapes. Sample errors are excluded.

Shapes

shape claude-opus-5-5 gemini-3.7-flash glm-5.3 gpt-3.5-turbo-0613 gpt-4.1 gpt-4o-mini gpt-5-mini gpt-oss-20b llama-3.1-8b-instruct
plain label: at line start 1271 1305 1303 1305 1305 1305 1304 1291 519
capitalised Label: 0 0 0 0 0 0 0 0 10
bold label, **label:** or **label**: 34 0 2 0 0 0 0 7 272
- bullet before the label 0 0 0 0 0 0 0 0 2
* bullet before the label 0 0 0 0 0 0 0 0 1
numbered 1. before the label 0 0 0 0 0 0 0 0 0
block quote > before the label 0 0 0 0 0 0 0 0 0
heading # before the label 0 0 0 0 0 0 0 0 0
indented label 0 0 0 0 0 0 0 0 1
label only mid-line 0 0 0 0 0 0 0 0 3
inlabel: present (negated label) 0 0 0 0 0 0 0 0 0
label : with a space before the colon 0 0 0 0 0 0 0 0 0
CRLF line endings 0 0 0 0 0 0 0 0 0
no label anywhere 0 0 0 0 0 0 1 7 503
replies 1305 1305 1305 1305 1305 1305 1305 1305 1305
the PR’s parser found no verdict 0 0 0 0 0 0 1 7 507

Replies with no label line

measure gpt-5-mini gpt-oss-20b llama-3.1-8b-instruct
replies with no label 1 7 503
empty 1 1 0
has extracted_final_answer 0 6 488
has reasoning 0 6 496
has confidence 0 6 427
verdict stated in words 0 6 358
median length (chars) 0 340 300