Adaptation time
Measure correction receipt to first verified improvement, with resolved and pending evidence kept visible.
Adaptation time measures how long a correction took to produce its first verified improvement. The clock starts when the correction is received. It ends at verification; serving the resulting Model is a separate milestone.
The calculation
Tᵢ = max(0, first verified improvement timeᵢ − correction received timeᵢ)
summary adaptation time = median(Tᵢ for resolved corrections i in R)R contains resolved corrections with actual training exposure. With an even number of durations, the median averages the two middle values. Times are compared in the same unit.
A receipt alone does not resolve a correction. Verification requires:
- actual training exposure to the correction;
- a failed parent check;
- the candidate passing every declared check under the frozen correction-verification plan.
Batch reservation or consumption alone does not prove that the optimizer used the correction. Training completion, acceptance and serving are distinct from the first verified improvement.
Read the median alongside pending corrections
| Value | Meaning |
|---|---|
| Median adaptation time | Receipt-to-first-verification time among resolved corrections with training exposure. |
| Resolved / tracked corrections | How much of the scoped correction population contributes resolved durations. |
| Pending corrections | Scoped, non-withdrawn corrections without a retained verified-improvement time. |
| Oldest pending correction | Age of the oldest pending correction at the end of the window. |
Illustrative calculation, not an OpenPond measurement: resolved durations of 1 hour, 3 hours and 8 hours have a median of 3 hours. If two other corrections remain pending and the oldest is 12 hours old, report those two pending cases and the 12-hour age beside the median. Their exclusion does not mean they adapted instantly.
No resolved durations means Not measured, not zero seconds. A fast median with many old pending corrections is incomplete evidence.
Keep the evidence scope fixed
Read the workspace, Model, run, exact checkpoint and artifact, correction-receipt window and observation time before comparing results. Withdrawn corrections are excluded. A result for one checkpoint does not establish the behavior of its base, a later candidate or every serving target.
Open the exact run's Correction outcomes, then follow its correction/task policy and evaluation links. Inspect blockers when exposure or verification is pending. The compact A badge shows measured adaptation time; missing values may omit the badge. These reads summarize retained evidence and start no new inference or training.
Adaptation time is separate from the optimizer's learning rate, replay fraction, training frequency and repeat rate. See Research and evaluation for broader evaluation questions and How it works for admission and candidate gates.
Contract reference
The public correction-metrics contracts define retained observations, nullable measurements and scoped windows. UI availability depends on retained evidence and the running release. The formula above explains the calculation; it is not a new measured result.
Compatibility
Requires retained correction, actual training-exposure and verification evidence for the exact Model/run/checkpoint/artifact. Missing measurements are not zero.