A historical decision can only use information that was available when the decision was made.
Keep two clocks
An observation can describe an earlier event while entering the system much later. Those timestamps answer different questions. Observation time says when something happened. Availability time says when the system could have known it.
Backfilled records do not become historical knowledge just because their observation dates are old. A replay must respect the ingestion and completion evidence that establishes actual availability.
Freeze the boundary
Choose a decision cutoff, preserve the eligible input set, and identify the policy or code that operated on it. Keep later information outside that boundary, including corrections and classifications introduced after the fact.
Separate the questions
Reproducing an artifact establishes that a particular process can produce the same output from the same sealed inputs. It does not establish that the output predicted a market outcome or supported an executable trade.
Predictive evaluation needs its own protocol: a declared universe, frozen decisions, defined outcome windows, explicit missingness, and enough evidence to support the conclusion.
Leave uncertainty visible
An inconclusive study is a result worth preserving. Missing outcomes, sparse observations, and unmeasured execution costs should remain visible rather than disappearing into a cleaner story.
Cavrino’s replay and research work motivates these boundaries. This note describes the method; it does not claim validated predictive performance.