autonomous-investigation

Use a protocol with a hard question budget (usually 3), a search-plan gate that auto-continues, and labeled assumptions for unanswered gaps. This lets investigations run as scheduled agent tasks that degrade gracefully instead of stalling when nobody responds.

Require every key claim to carry a Fact, Inference, or Assumption label, with real checkable URLs for facts. Define a domain-specific do-not-invent list naming fabrication risks like pricing, market share, and customer wins, and put unfound items in an explicit gaps list.

Confidence stacking grades findings by independent channel agreement: one channel is a watch item, two channels form a working hypothesis, and three or more channels make actionable intelligence. Conflicting channels trigger a dig rather than averaging, since conflict often signals deception.

Yes, if the investigation uses a stable output schema where section order and structure never drift between runs. Diffable schemas enable delta monitoring, so teams read what changed since last quarter instead of re-reading full reports.

Collection must be legal, open-source work: anything published, filed, posted, or publicly observable. It excludes pretexting, soliciting NDA-protected information, or scraping against accepted terms, following the SCIP Code of Ethics stage test.