Evidentiality Framework for AI
Download .md

Build on it

How to Take This and Build Something Better

This is a framework, not a finished product. The idea is for people to take it and make something much better with it.

Part 1 Build on It

  1. Take it. Text is CC BY 4.0; scripts are MIT. Rename the labels, change the instructions, build tools on it, ship it. Credit it. To cite: Zesbaugh, J. (2026). Evidentiality Framework for AI (version 0.7, early findings). https://evidentiality-framework.org/ On GitHub, the repository’s “Cite this repository” button gives the same citation in APA and BibTeX.

  2. Pick something to build. Some ideas we haven’t built:

    • A browser extension that colours the labels in any AI chat.
    • A chat interface that shows each label as a badge you can click to see its source.
    • A gate for a real agent pipeline, with a checker that isn’t the writer.
    • A swarm dashboard that follows one claim from agent to agent.
    • Versions of the labels for other languages, or for fields like medicine, law or journalism.

  3. Start from the working version. The version we use every day is the most complete. The tested version is the one with evidence behind it.

Part 2 Good First Projects

  1. Measure how often the labels are right, at scale. Our checks are one run per model, or a few runs per condition. Nobody has a rate yet.

  2. Build a checker that isn’t the writer. An application or a second model that applies or verifies “checked”.

  3. Separate the labels from the instructions. Run the spoke and wheel test with the instructions but without the labels.

  4. Turn examples into rates. Twenty or more runs per version, other model families, scored by someone who doesn’t know which is which.

  5. Fix the known issues. Models label their own sums as (u); a conclusion’s closing tag doesn’t say whose conclusion it is; plans drift into reports of progress over several rounds; a received (m) is passed on as checked; the full version’s document format sometimes appears on other tasks; the gate’s gaps.

Part 3 Tell Us What You Made

  1. Show us. Open an issue in the repository with a link to what you built.

  2. Report a result. Model and version, which instructions, which test, how many runs, what you counted, and quotes for every failure. A result that shows it failing is as useful as one that shows it working.

A more formal write-up: the working paper (a draft, not peer reviewed). Thank you. Source and updates: the GitHub repository.