A Research Run Knows When to Stop

The platform’s standing agents never stop. They watch, score, discard, and surface — cycle after cycle — because the domains they cover never stop moving. But not every question is a domain. Some questions are singular: asked once, answered once, then closed. For those the engine takes a different shape. It stops being a watchman and becomes an investigator. And the most telling thing about an investigator is that it knows when the investigation is over.

§ I   Two shapes of the same engine

The standing loop and the research run are the same machinery aimed differently. The standing loop is continuous and keyed to a domain; it runs because the domain keeps changing. A research run is episodic and keyed to a single question; it runs because the question was asked, and it ends when the question is answered.

A research run does to its question what the architecture’s conclave principle prescribes for anything that matters. It decomposes the question into independently testable sub-questions. It sets specialist analysts on them in parallel. A validator checks each claim against the evidence the analysts produced, and a supervisor decides whether the chain proceeds, goes to debate, or is killed. Where a standing agent watches one source indefinitely, a research run assembles a whole chain around one question — and then dissolves it.

§ II   Knowing when to stop

A research run is iterative. It does not answer in a single pass; it runs cycles, each one sharpening the last. The hard question is when to stop. A system that halts after a fixed number of cycles is not doing research — it is filling a quota.

The platform’s research engine halts on a different basis. It stops when its confidence in the answer is high and further cycles have stopped reducing uncertainty. When an additional cycle would only restate what the last one found, the run is over. A cycle cap exists as a backstop — a question that refuses to converge does not run forever — but the cap is the exception, not the design. The design is convergence. That is what separates a research run from a chatbot’s answer: the run is built to recognise the moment its own additional effort has stopped paying, and to stop there.

§ III   A capability, not a service

The research engine is real, and it runs — from the command line, and on a schedule when a standing question warrants periodic re-investigation. What it is not is a product. No external party commissions research through it. There is no intake, no queue, no billing. Its output lands in the platform’s own research files, and from there it can feed a publishing arm or simply inform the operator.

What would make it a service is not more capability; the capability is built and exercised. It is the surface around it — a way for someone outside to pose a question and receive the run, with the isolation and accounting that an outside engagement requires. That surface does not exist yet. What exists is the engine itself: aim it at a single question, and it will investigate, and it will know when it is done.

The watchmen and the investigator run on one engine. The only difference is whether the question keeps moving or holds still. A domain gets a loop that never ends. A question gets a run that ends on purpose.

Drafted with AI assistance under operator supervision; substantive claims are operator-authored or operator-approved.