LLogosmose
Grader v1 · Lite Beta

The grader, in the open

No single model decides the verdict. Four frontier models contribute at different points of the pipeline. Every version we ship is documented here.

Models in this version

OpenAI
GPT-class · reasoning step
30%

Decomposes the answer into claims and assesses logical sequencing.

Anthropic
Claude · rubric scorer
30%

Scores each of the five criteria against the rubric.

Perplexity
PPLX · factuality
25%

Cross-checks factual claims against fresh sources.

Google
Gemini · style & concision
15%

Audits prose quality, hedging, and concision.

Fallback: Adopts the same merged-contribution approach. No single vendor controls all the grading.

Rubric · five criteria, one number

Relevance
25%
Reasoning
25%
Clarity
20%
Accuracy
20%
Concision
10%

AI-writing check

Beyond the five criteria, every answer is assessed for how likely it was written by an AI. The more an answer reads as machine-generated, the more its score is handicapped: no penalty below a confidence floor, then a continuous, linear reduction up to the full handicap. Logosmose is a contest of your thinking — the AI-likelihood, your raw score and your final score are all shown on your verdict, in the open.

Changelog

v1Lite Beta

Initial grader: merged scoring across OpenAI, Anthropic, Perplexity and Google, plus a continuous AI-writing detection handicap. Nothing has shipped to production yet.

How we ship a new version

No silent rollouts — any model swap is documented on this page first.