File size: 1,754 Bytes
a79b199 b54a239 a79b199 b54a239 a79b199 b54a239 a79b199 b54a239 a79b199 b54a239 a79b199 | 1 2 3 4 5 6 7 8 9 10 11 12 | # Current CrossSearch grading protocol
Version: gpt61-20261106. Scope: 4,650 existing answers (3,300 native-search and 1,350 uncapped search comparisons).
Judge: GPT-6.1 Sol through the authorized inference gateway, high reasoning effort, temperature 0, 16,000 maximum response tokens. The semantic system prompt is [CrossSearch-GPT61-Judge-Prompt.txt](CrossSearch-GPT61-Judge-Prompt.txt). The judge received the original full task context, rubric, reference answer and candidate answer, without the generating model identity. No generator model was rerun.
Each rubric ID must appear exactly once. Each item has a Boolean pass/fail decision, a rationale and verbatim evidence when supplied. The validator checks all quoted evidence against the unchanged answer and requires a normal completed response. Invalid or incomplete responses do not become grades; distinct completion attempts retain the same input and all previous observations are preserved.
Weighted score = earned rubric weight divided by total rubric weight. Full credit requires score 1. Pass@1 is the mean of full-credit indicators across three attempts per task. Pass^3 is the share of tasks with all three attempts receiving full credit. Confidence intervals use 20,000 task-bootstrap resamples, keeping each task’s three attempts together. Search differences use paired resampling of the same 50 tasks.
Original grades remain in the unchanged rollout shards and in run-index-original-20261105.jsonl.gz. Current grades are in run-index.jsonl.gz and grades-gpt61-20261106.jsonl.gz. Do not mix judge versions in a single reported metric. The independent human study used question-only context; these benchmark regrades additionally include full task context and reference answers.
|