A workflow that prioritizes usable output over a long conversational performance.
Define the research contract
A good investigation begins by making the requested decision explicit. “Research this market” is open-ended; “identify the conditions under which this system is cheaper over twelve months” creates a testable contract.
Horizon converts the initial question into scope, evidence requirements, exclusions, and a completion condition. The user can revise that frame before the expensive work begins.
Preserve intermediate work
Search results, extracted claims, contradictions, and discarded evidence remain separate from the final prose. That separation makes revisions cheaper and gives the user a way to inspect how the answer changed.
The final artifact is assembled from those parts: executive finding, evidence ledger, uncertainty, and suggested verification. It can be exported or continued without replaying a long chat transcript.
Know when the work is done
A research run is complete when it satisfies the declared evidence threshold, not when the model has produced a long answer. If the threshold cannot be reached, the correct output is a bounded finding with the missing evidence named.
This completion rule prevents verbosity from masquerading as depth and gives latency and cost a defensible stopping point.