Transparency
Real numbers behind the claim on the About page: ranking here is provably tested, not just described that way. This reads from the same tables the ranking system itself runs on, nothing curated or delayed.
The three arms, right now
| Arm | Role | Queries this window | Blended score | Status |
|---|---|---|---|---|
| A | Actively tuned | 5 | 1.0000 | Last changed 2026-09-06 |
| B | Last promoted winner | 10 | 1.0000 | Last promoted 2026-09-06 |
| C | Permanent frozen control | 6 | 1.0000 | Frozen since 2026-08-19 |
Rolling window ending 2026-09-09 17:13 UTC. Blended score combines click position, reformulation rate, and dwell time into one number, higher is better. "Not enough data yet" means this window's sample size is too small to trust, not that the arm is performing poorly. The exact weights each arm uses aren't published here: it's the one thing kept back, since publishing them would hand anyone a blueprint for gaming the ranking.
Automatic tuning
Currently testing a change to multi_engine_bonus on arm A, started 2026-08-20.
Track record: 0 candidates tested, 0 committed, 0 discarded.
Scale
| Searches, total | 637 |
|---|---|
| Documents in the persistent index | 9180 |