Chapter 22 — Adaptive Cadence and Event Detection
Choose the sampling schedule
Fixed mode records at a configured interval. Adaptive mode can record more densely when the instrument detects change and relax toward a less frequent schedule at other times. These are sampling choices, not declarations that the model has learned, generalized, or failed.
Use the Bench controls or supported iris.cmd commands. The run must reach
should_extract() or extract() for staged commands to be read.
echo "interval 500" > iris.cmd
That example pins the interval at 500 training steps. Choose a value suited to your workload and the temporal resolution you need; it is not a universal recommendation. Percentage settings depend on the declared total-step count.
Adaptive bounds and sensitivity
cutback <fast> <slow> sets the fast and steady-state bounds used by adaptive
capture. In the Bench, review both interval boxes before committing a change.
sensitivity <S> changes how readily the instrument records denser evidence:
higher means more frames, lower means fewer. Its relationship to the number
of frames is approximate, and a quiet run can remain quiet.
These controls affect measurement work and storage as well as temporal resolution. Measure overhead on the actual model and hardware. A denser capture is useful only when the observations it records help your question.
What a marker establishes
An event marker records an instrumentation decision or reported event. It does not identify a grokking event, causal circuit, optimizer failure, or generalization mechanism. Representation-motion telemetry and trainer outcomes remain different measurements. Their alignment is an observation to test, not a universal prediction rule.
Changes you make to the run can produce real changes in the readings: checkpoint resumes, optimizer changes, reinitialized weights, or a changed data pipeline. IRIS™ does not know the meaning of an intervention merely because it observed motion afterward. Keep a run note and interpret the event against that record.
Timing claims need cadence
When comparing a marker with accuracy or loss, report the relevant step labels, the sampling intervals, and the source of the outcome metric. Sparse captures bound the timing you can resolve. A visible animation between frames is not another measurement, and the mean-accuracy curve is not a substitute for per-domain evaluation.
Inspect the tracker log and the frame coverage when an event is recorded without a nearby plotted value. Do not invent a missing point or relocate a marker to make a proposed explanation look cleaner.
