only fields that cannot see what comes after

Before a decision stream is allowed to feed anything live, it has to pass a registration gate. The interesting part is not that the gate checks the output is well formed. It is that the gate tries to show the output could have been produced at the moment it claims.

The strongest check is truncation invariance. Produce the whole stream from the full series, then cut the series at many different points and produce the stream again from each truncation. Every decision that appears before the cut must be identical, in the same order, with nothing lost and nothing gained at the boundary. If a decision changes because more data now follows it, the producer was reading forward, whatever its own description says.

That check forced a second decision: what counts as the same decision. The comparison keys had included a field derived from how long the series runs, which is null past the head of the archive. Comparing on it would have measured the length of the record, not causality. Equivalence keys were restricted to fields that cannot depend on later data by construction, and the cases that legitimately differ at the tail were separated out and swept explicitly instead of being folded into the total.

The cost is that the gate yields a narrower claim than "correct". It says the stream is reproducible, ordered and causally honest — and it has to be re-run whenever the inputs move.

What it proves: the engine can show, by construction rather than assertion, that its decisions do not depend on information it would not have had.