Fair play & AI grading

Logosmose is a competitive arena where AI-graded answers determine your place in the public ranking. For that to work, participants need to trust two things: the grader scores fairly, and other participants play by the rules.

This document explains how the AI grading system works, what we consider rule violations, how we detect them, and what happens if you breach the fair play policy.

Every answer you submit is scored by a multi-model AI grading pipeline. No single model controls the verdict — we use four frontier models (OpenAI, Anthropic, Perplexity, Google) at different stages to reduce bias and gaming risk.

The current grader version (v1) evaluates your answer against five criteria: Relevance, Reasoning, Clarity, Accuracy, and Concision. Each criterion carries a weighted percentage that contributes to your final score out of 100. Answers are also checked for how likely they were written by an AI, and a machine-generated reading handicaps the score in proportion to that likelihood.

The full rubric, prompt, and grader version history are published at logosmose.com/transparency for full transparency.

The following actions violate fair play and will result in sanctions:

• (a) Using an AI assistant or co-writer during a timed answer
• (b) Pasting prepared text written before the question was revealed
• (c) Receiving the question or hints from another participant or third party
• (d) Sharing your question or answer with another participant before pool close
• (e) Running multiple accounts to enter the same pool
• (f) Probing the grader at scale to reverse-engineer its weighting
• (g) Submitting answers generated by another participant on your behalf

We monitor for rule violations using a combination of automated signals and manual review. Detection mechanisms include paste pattern analysis, submission timing anomalies, duplicate content detection, and cross-account behavioral clustering.

If a participant's behavior triggers multiple red flags, their account is flagged for manual review by our trust team. We investigate before taking action.

Clicking "Discard" during a timed round counts as a forfeit — you use one of your participations, but your answer is not submitted, scored, or made public. This is not a breach.

If you let the timer expire without submitting, any draft text you wrote is auto-submitted. If that draft was prepared before the question reveal or pasted from external sources, it still violates the rules.

Answers that read as AI-written are automatically handicapped in scoring, in proportion to that likelihood. Beyond that, confirmed severe violations of our rules lead to escalating action, counted per calendar year:

First and second violations: a formal warning by email, and the offending answer is removed and excluded from the ranking and public feed for everyone. Third violation in the same year: a 30-day account suspension — you cannot sign in and your content is hidden — followed by permanent deletion unless the decision is overturned within 30 days.

Because these decisions are automated, you can contest a warning or a suspension at contact@logosmose.com. Egregious abuse (such as multi-accounting or coordinated cheating) may be acted on directly, up to immediate account termination.

The grader rubric and model weights are versioned and publicly documented on our transparency page. When we change the grader, the update is documented there first — there are no silent rollouts.

The current version is v1 (Lite Beta). Its models, rubric, AI-writing check and changelog are published on the transparency page.