Point-in-Time Context Admission
Admit only information available at the decision time when replaying or evaluating a historical task.
These examples and illustrative results are independently authored teaching materials, not measured model results.
Use case
Evaluate an incident assistant at a 10:05 decision cutoff after the root cause is known. Teaching R1 occurred and became available at 10:02; R2 occurred at 10:03 but was readable only at 10:12; postmortem R3 was published at 11:00. Later facts may inform labels, not historical input.
Mechanism
Freeze cutoff/snapshot and record event time separately from availability. Admit only pre-cutoff available material; late records, corrections and postmortems remain label/analysis inputs. Unknown availability is an explicit gap, not eligibility inferred from event time. Preserve construction rules for repeatable leakage checks.
Bad example
Use R1,R2 and the 11:00 root-cause postmortem in the 10:05 replay. R2 occurred at 10:03, so it was available. Claim the assistant diagnosed the cause from early evidence.
Good example
Freeze the teaching cutoff at 10:05. Include R1 only. Exclude R2 because availability is 10:12 despite its earlier event; keep R3 as a later evaluation label outside model input. Record availability/provenance, preserve unknown times and maintain snapshot/input-label separation to assess the same early evidence.
Why the change matters
Hindsight changes how early evidence is interpreted. Availability-based admission prevents late records/postmortems from inflating historical decision ability and distinguishes an event’s occurrence from information actually known.
Observable expectation
An illustrative inventory admits R1, excludes R2 as late and R3 as label-only. Inspect inputs for no postmortem cause or later corrected fields.
Unknown availability does not pass automatically; a result citing R3 first needs a leakage check. No real model evaluation was run.
Limits
Point-in-time boundaries serve replay/prediction/evaluation, while current work may need latest facts. Availability, versions and permissions all matter; filtering one timestamp alone does not establish absence of leakage.