All experiments

Experiment 02 / Logosmose / 2026

Can AI-assisted evaluation earn trust?

Write for five minutes, get scored by four AIs, see where you rank.

FounderProduct designSolution architectureLead full-stack development

01 / Hypothesis

Evaluation can be structured without pretending it is objective.

Inconsistent human evaluations made me question trust in grading. The product makes a narrower promise: assess a response against a public rubric, not a person’s intelligence.

02 / Smallest loop

Make the test brief and inspectable.

  1. 01

    Reveal one difficult question.

  2. 02

    Write for five minutes, within 2,000 characters.

  3. 03

    Evaluate the response across five public criteria.

  4. 04

    Return the score, criterion breakdown and rank.

An AI evaluation architecture you can trust

One answer reaches four specialist evaluators. Two share relevance and reasoning, one reads clarity and concision, one checks factual accuracy. Their weighted contributions remain visible alongside the result.

Answer; M1; M2; M3; M4; Aggregate; Relevance; Reasoning; Clarity; Accuracy; Concision.

  1. Your answer goes to four specialist AIs
  2. Each AI grades only the criteria it owns
  3. The judgments combine into one score
  4. Five dimensions stay visible — the verdict can be checked

03 / Division of labour

AI contributes. People remain accountable.

AI leverage

Where models contribute

Four providers split a weighted five-criterion rubric: OpenAI and Anthropic share relevance and reasoning, Gemini grades clarity and concision, Perplexity checks factual accuracy. No model controls the verdict alone.

Human judgment

Where responsibility stays

I set the claim, rubric, limits, aggregation rules and wording. Users decide whether the result helps them.

04 / Product evidence

The constraint and the verdict are visible.

Live screens from free participation and training. They show the interface, not adoption or grading validity.

Logosmose answer screen showing a five-minute timer, a revealed question and a 2,000-character response limit.
Production UI / timed response with explicit limits
Logosmose result screen showing an overall score, five-criterion chart and rank.
Production UI / score, criteria and rank returned together

Reality check / the business model changed

I removed the financial dimension after legal concerns, gambling associations and payment-processor friction. The free product now tests trust and repeat participation before adding commercial complexity.

Next test

Do people choose to return?

Visit Logosmose