Watches
Keeping the model you replaced running as a low-rate sentinel, and hearing about it if it starts winning again.
Watches
A watch is a trial that never ends, with the roles flipped.
After you promote a candidate, a watch keeps the old model running as a low-rate sentinel: the model the promotion just adopted is now the served side, and the model it replaced is the sole candidate. If the one you replaced starts scoring better again, you hear about it.
| Property | Value |
|---|---|
| Served side | The model the promotion adopted. |
| Candidate | Exactly the old incumbent. |
| Sample rate | 2% of sessions, deliberately low because a watch stands forever. |
| Scope and targeting | Inherited from the trial it came from. |
| Spend cap | A monthly budget rather than a terminal one. |
A watch alerts only. It never changes routing. Rolling back is the same explicit promotion action run on the watch's own candidate, so it goes through the same parity-floored, human-initiated path as any other adoption.
Creating one
A watch is created only from a promotion -- which is why the Watches screen has no "new watch" action. The affordance lives where the adoption decision is made.
Arming one is opt-in, and a monthly budget is required and must be positive. A zero budget is a refusal, not "unlimited".
Only one watch stands per scope. A newer promotion supersedes the standing watch, and records what replaced it.
Reading drift
Drift detection is a read, never a writer: the same session-denominated verdict maths a trial uses, over a rolling window of trailing judged sessions, against the same 20-session floor. Below the floor it refuses to speak.
| Status | Meaning |
|---|---|
collecting | Not enough judged sessions yet to say anything. |
healthy | The adopted model is holding. |
drifted | The model you replaced is scoring better again. |
The sign is inverted relative to a trial. A watch's delta is adopted minus sentinel, so a negative delta means the model that was replaced is scoring better. Do not read a watch's sign as a trial candidate's.
A firing watch surfaces in three places: on the Evals Live screen's alert strip, on the Watches screen with its failing dimensions and window-scoped evidence, and as a marker on the scope's routing profile.
The monthly window
A watch's budget recurs rather than ending. The window rolls over, which resets the accrual and re-arms a watch that was paused on budget.
A budget-paused watch is not retired. It is waiting for its window to roll.
History
Watch history is unbounded -- a watch stands indefinitely, so the list can grow without limit. The screen therefore defaults to active watches only: an alert nobody can see is worse than no alert, because the page looks calm.
Watches do not appear in the trials list
A watch is a kind of trial rather than a parallel object, which is what makes its dispatch, its judge version, its scrubbing, its residency posture and its spend accrual identical to a trial's -- the evaluation engine genuinely cannot tell the two apart.
But it is excluded from trial list reads, and it does not carry a trial's verdict. Do not expect to find one on the Trials screen, or to read a watch as won or lost. Its answer is a drift status.