Tracking status
Drift from baseline
Both signals measure how far this endpoint's behavior has moved from its reference period (its first two weeks). LT tracks the average change in the returned tokens' log-probabilities (nats); B3IT tracks the total-variation distance of the border-input votes (0–1). Both read 0 when the endpoint matches its baseline and rise — staying elevated — after a real change. Dashed lines mark detected changepoints; σ is one of the metrics feeding LT's change-detection criterion, separate from how far the signal actually moved.
Detected changes
| Date | Method | Drift reached | σ |
|---|
LT Logprob tracking
Primary signal where the provider returns logprobs. The drift is the mean absolute change in the returned tokens' log-probabilities versus the reference period (in nats — natural-log probability units). The detector also reports σ, one of the metrics feeding its change-detection criterion, which can be high even when the drift is small.
B3IT Border inputs
Black-box fallback for endpoints that hide logprobs. Tracks how far the top-token votes on a ranked set of near-tie prompts have moved from the reference distribution (total variation, 0–1). Corroborates LT where both run on the same endpoint.