tacit

About

Tacit is an independent EU benchmark for System 1 decision models. It is built and maintained in the EU, written in EU languages rather than translated from English, and covers decisions made under EU rules.

The name comes from tacit knowledge: fast, intuitive judgement that is rarely put into words. System 1 and System 2 are Daniel Kahneman’s terms from Thinking, Fast and Slow. Here, System 1 is the model that answers in milliseconds, and System 2 is the person who gets the case when the model is not sure enough.

Independent. Not affiliated with or endorsed by EU institutions.

Independence and supporters

Tacit is maintained by Greg Brzezinka. All costs (model and API tokens, compute, hosting and the domain) are currently covered by Prosit AS, Greg’s Norwegian-Polish technology consultancy.

Supporters get no access to the hidden test, no preview of results and no say in what is tested or when it is published. Every run so far was done by the maintainer, and no vendor approved its numbers.

Supported by

  • Prosit ASTokens, compute, hosting and domain

Compute and credit donors will be listed here, one line each, with what they gave. The same rules apply to them.

Changelog

  1. v0.1

    27 October 2026

    First release. Polish, 1,254 items in 19 families, 16 models on the board. One dev label (D22) corrected after maintainer review.

Citation

@misc{brzezinka2026tacit,
  title={Tacit: an independent European benchmark for System-1 decision models},
  author={Brzezinka, Greg},
  year={2026},
  version={0.1},
  url={https://tacitbench.ai}
}

Licences, contact and links

Code
MIT
Dev data
CC BY 4.0. Please do not train on it.
Test data
Not distributed. Published as SHA-256 hashes only.
Contact
GitHub Discussions
Code and harness
github.com/tacit-bench/tacit
Dev dataset
huggingface.co/datasets/tacit-bench/tacit-dev
Leaderboard source
LEADERBOARD.md
Explainer
System 1 AI, on agentgreg.ai