| # Standing Measurement Protocol |
|
|
| 1. Re-run the identical five queries monthly. Same wording, character for character. `data/queries.csv` is the fixed instrument. |
| 2. The primary instrument is a fresh Chrome Incognito session, logged out of all services, one system at a time, screenshots captured. Automated index checks (search APIs, agent tools) are supplementary and must be labeled as a different instrument in `runs.csv`. |
| 3. Never test inside an existing conversation. An earlier test of a predecessor term was contaminated when the model referenced a prior turn. Contaminated measurements are worse than none because they look valid. |
| 4. Record `domain_appears`, `cited`, and `attributed` as separate columns. A system can do any of these without the others. |
| 5. The 31 July 2026 baseline is singular and unreproducible. Never overwrite it, never call a later run a baseline. |
| 6. File each run as a new set of rows in `data/runs.csv` with `run_date` and `t_plus_days` (site launch = day 0, 2026-07-31 baseline = -1). |
| 7. First attribution: capture the screenshot, record the row, preserve the raw output. That artifact paired with the baseline is the payoff of the entire protocol. |
|
|
