ogous/llm-judge-version-change

LLM judge scores across two Claude Sonnet upgrades (SummEval coherence)

The judge scores behind the paper LLM-as-a-Judge Under Version Change: Recalibrate or Re-judge? (Ogous Chan Ali, 2026; arXiv: coming soon). Three versions of one judge, claude-sonnet-4-6, claude-sonnet-5 and claude-sonnet-5-5, rated the coherence of the 1,600 SummEval summaries of 100 articles (16 systems × 100 articles) on a 1-to-5 scale, under one prompt and an output schema, with each item judged six times. The study was pre-registered (OSF y6qm8) and collected on 2026-10-01 through the provider's batch API.

This is the data part of the study package (doi:10.5281/zenodo.23358200), which also holds the registration, both analysis implementations, the collection harness and every analysis output. The code that reads these scores as eval files is anchorset.

What is here

File	Rows	One row is
draws.jsonl	9,600	one judge version, one set, one system, one article: scores holds that set's three draws
sentinels.jsonl	144	a fixed probe item (one per system) scored three times during a judge version's collection: wave is start, middle or end
attempts.jsonl	29,232	one API attempt: ids, times, requested and served model id, stop reason, token counts, cache use, and the sha256 of the request and of the response
decoding.json		the request settings of each judge version (max_tokens, thinking, the output schema)
flags.json		per version, the draws flagged for emitted reasoning or truncated output (none)

Fields of draws.jsonl and sentinels.jsonl: model is s46, s5 or s55 (the three versions above, in that order); set is 1 or 2, two independent sets of three draws on the same item (set 2 is the re-ask); system is the SummEval system id (M0 … M23); article is the SummEval article id; scores are integers from 1 to 5.

claude-sonnet-4-6 and claude-sonnet-5 ran with thinking disabled. claude-sonnet-5-5 does not accept that setting and ran at its lowest one (between_tools). claude-sonnet-5 and claude-sonnet-5-5 reported zero thinking tokens on every attempt; claude-sonnet-4-6 reports no count (null). Every attempt in attempts.jsonl returned a valid score.

What is not here

No article text, summary text, prompt or response text, and no expert score. The expert coherence scores are SummEval's own (Yale-LILY/SummEval, MIT): fetch model_annotations.aligned.jsonl from the SummEval repository's link, which the study package's README gives with its sha256, and join its id to article and its model_id to system. An item passes the study's expert gate when the three experts' coherence scores sum to 12 or more. The study's split into 80 scored and 20 calibration articles is s1/manifest.json in the study package.

Licence and citation

Released by the author under CC BY 4.0; the scores were returned by the provider's models to the author as its customer (the study package's RIGHTS.md, rows 6 and 7). Cite the paper:

@misc{ali2026judgeversion,
  title        = {{LLM}-as-a-Judge Under Version Change: Recalibrate or Re-judge?},
  author       = {Ali, Ogous Chan},
  year         = {2026},
  howpublished = {\url{https://ogous.dev/publications/llm-judge-under-version-change}},
  note         = {Preprint}
}