How We Test and Grade Evidence
How we prompt AI models, interview Guzhen craftspeople, measure physical samples, and grade every disagreement from A to C.
1. How we query AI models
Before or right after filming a workshop question in Guzhen, we ask current frontier models the exact same plain-language question in a clean session with no leading instructions. Each prompt is run three times per model so we can distinguish stable claims from one-off hallucinations. Every disagreement page logs:
- The model name and exact version string;
- The date queried and number of runs;
- The verbatim prompt text;
- A plain summary of what the model claimed.
When major model updates ship, we re-run the identical prompt and append the new result to the page’s Re-test log rather than overwriting history.
2. How we record craftspeople
Workshop masters are interviewed in Mandarin or Cantonese on location in Guzhen, Zhongshan (and neighboring Foshan, Jiangmen, and Shunde workshops). Every craftsperson signs a bilingual release choosing either to be credited by name and workshop or to appear anonymously by trade (for example, Polisher, 18 years). We publish their original Chinese words alongside our English translation.
3. Evidence grades (A, B, C)
Neither an AI model nor a veteran craftsperson wins an argument by sounding confident. Every verdict carries an explicit evidence grade:
| Grade | Definition | Examples |
|---|---|---|
| Grade A | Measured with a calibrated instrument or proved by a primary standard / regulatory document | Gemological refractometer reading (n = 1.516 vs. 1.545), handheld XRF lead ppm check, integrating sphere CRI/CCT/flicker report, thermocouple 24V temperature rise, UL Product iQ file verification |
| Grade B | Repeatable physical workshop demonstration on camera without lab-grade instrumentation | Side-by-side prism dispersion projection at fixed distance, scratch/file check on solid H65 brass vs. plated steel, drop-pack survival test |
| Grade C | Experienced practitioner judgment or shop economics where no single physical test settles the question | Yield-loss estimates on complex hand-cut crystal shapes, realistic small-batch tooling trade-offs |
4. Verdicts
Each row resolves to one of four outcomes: AI was right, The master was right, Both were wrong, or Unresolved (when our test setup could not isolate the variable cleanly yet).