LuminaBench
Reasoning benchmark / Conceptual Reasoning

ACCoRD

Lumina VerifiedLumina checked the published sources, versions, model labels, score extraction and conversion. External evaluations were run by their publishers; Lumina has not independently rerun or reproduced them.

Normalized constraint violation

Version CRI-public-components-20261106Verified 6 November 2026155 published resultsOfficial source

ACCoRD

LuminaBench
Show
20 of 155 models · best variant per model
03006009001,2001,120Claude Opus 5.5Max1,114Claude Sonnet 5.5Max1,113Claude Fable 5Max1,102GPT-6 Sol1,100Claude Opus 5Max1,099GPT-6 AstraMax1,098Claude Haiku 5.5Max1,095Claude Fable 5.1Max1,095GPT-6.1 Sol1,093GPT-5.51,093Claude Opus 4.61,091GPT-5.41,090Claude Opus 4.71,088GPT-5.21,087Claude Opus 4.51,086GPT-5.6 Sol Pro1,084Muse Spark 1.11,080Kimi K3.1 Pro1,080Gemini 3.8 FlashHigh1,079Claude Opus 4.8

Original average USD per LMCA/DTBench response as documented in the source data README; absent costs stay unavailable. No time or token measurements inferred. Missing metrics are omitted from their graph; quality scores remain available.

Results table

Method & coverage

Original task component from public cri-scores.csv. Preserve native efforts and assisted/partial qualifications. Parent CRI aggregate is not imported. Primary normalized linear-margin constraint violation, lower is better; no mixing with rescaled concord score or separate 60-sample arena experiment.

Original average USD per LMCA/DTBench response as documented in the source data README; absent costs stay unavailable. No time or token measurements inferred.

How the Lumina conversion works

1,000 − 75 × ((original score − 0.2921109677419355) / 0.12116131585763587). The result is rounded once. Scores may tie.

Frozen initial observed text-only cohort; one mean per exact publisher release/API, balancing repeated settings/routes within that release. Original score and protocol remain separate. No inferred variants or future refitting.

The source metric is lower-is-better; its standardized difference is negated before adding the baseline.

Calibration
accord-z75-v1 · experimental
Reference mean
0.2921109677419355
Population SD
0.12116131585763587
Coverage
Standalone benchmark.
Identity, exclusions & attribution

DTBench rescaled uncertainty is not attached to raw decision accuracy; ACCoRD rescaled uncertainty is not attached to normalized constraint violation.

Text conceptual comparisons, decision problems and logical consistency constraints; no visual or audio inputs.

Standalone text benchmark; no new combined index. Source identity/effort/protocol qualifications remain in Data and CSV.

Official source · Publication terms / evidence

Publisher terms. Scores and resource figures are attributed to the original publisher.