Methodology

The full write-up (pipeline, caching, privacy, evaluation protocol and known limitations) is being written. These principles are fixed:

1. How, not what. We label techniques and behaviours, never people, outlets or viewpoints.
2. The AI proposes, code verifies. Every highlighted passage must appear word for word in your text; positions are computed by our code.
3. Reported speech is not the author's manipulation. Quoted words attributed to someone else are marked and not counted.
4. Abstain instead of guessing. “Nothing found”, “insufficient data” and “not checked” are valid answers.
5. Two independent signals before calling anything “likely coordinated”.
6. Only show what we can measure. Techniques below 70% measured precision are experimental and hidden by default.
7. No single “propaganda score”. Profiles and evidence instead.
8. Colours describe, they don't judge. Amber means “no source given”, not “false”; green means “source given”, not “true”.
9. Innocent explanations are listed next to every flag.
10. Privacy by default. No accounts; swarm uploads are processed in memory.

No AI-generated-text detection: such detectors are unreliable and known to be biased against non-native writers.