truth-check-template

Companion repository to the article "Ein Truth-Check-Protokoll für AI-Forschungs-Output - wie wir bei myBytes nichts veröffentlichen, das wir nicht verteidigen können" (mybytes.com/research/truth-check-protokoll).

This repository contains the templates and a reproducible demo notebook that anyone can use to apply the seven-step truth-check protocol to their own AI/ML research output, vendor evaluations, or PoC reviews.

What you get

• templates/claim_map.csv - empty template for Step 1 + 2: atomar claims of an article or pitch, each typed (T1-T7).
• templates/anchor_mapping.csv - empty template for Step 3: per claim, the external evidence anchor with tier classification.
• templates/reproducibility_bundle.md - checklist for Step 4.
• notebooks/truth_check_demo.ipynb - reproduces Plot 3 of the companion article from the example claim-map.
• docs/protocol.md - the seven steps in detail, as authored.

The seven steps in one line each

1. Claim Extraction - break the text into atomar claims.
2. Claim Classification - type each claim (T1 Stylized Fact, T2 Methods, T3 Causal, T4 Forecast, T5 Regulatory, T6 Remote Sensing, T7 Market Mechanism).
3. Anchor Mapping - for every claim, at least one Tier-1 or Tier-2 source. Tier-3 never stands alone.
4. Reproducibility Bundle - code, seed, data version, CITATION.cff.
5. Steel-Man Counter-Argument - the strongest opposing view, addressed in-text.
6. Limitations - where would this claim be false?
7. Independent Review - second reviewer signs off against 1-6.

Quick start

git clone https://github.com/myBytesResearch/truth-check-template.git
cd truth-check-template
python -m pip install -r requirements.txt
jupyter notebook notebooks/truth_check_demo.ipynb

Adapt for your own publication

1. Fork this repo.
2. Copy templates/claim_map.csv and templates/anchor_mapping.csv into your article's working directory.
3. Fill them out as you write. The discipline of doing it inline prevents the most common Truth-Check violations.
4. Use the notebooks/truth_check_demo.ipynb as a starting point for your own claim-map visualisation.

Tier hierarchy at a glance

Tier	Definition	Examples
T1	peer-reviewed scientific literature	Nature, Science, Econometrica, Journal of Finance, Remote Sensing of Environment, SSRN/arXiv with caveat
T2	institutional, authoritative	IMF, World Bank, FAO, ICCO, USDA, EU Commission, ESA, NASA, MIT NANDA, Gartner Research
T3	industry-respected (never stands alone)	Mintec, S&P Global Commodity Insights, Argus, Reuters, FT, Bloomberg, Risk.net
T4	own research with full reproducibility	public repo, fixed seed, dated data version, license-compliant

Licence

• Content (Markdown, documentation, templates): CC BY 4.0
• Code (Python, notebooks): MIT

Both licences explicitly permit forking, adaptation and commercial use. Attribution to myBytes GmbH appreciated, not required.

Citation

See CITATION.cff for machine-readable metadata. To cite in text:

Winger, G. & Pianowski, M. (2026). The myBytes Truth-Check Protocol for AI Research Output. myBytes GmbH. https://mybytes.com/research/truth-check-protokoll

Issues, PRs, criticism

Please open issues or PRs. The protocol is explicitly open for critique - we want it to be stronger after release than before.