Deliberative Programming: A Versioned Deliberation Protocol for Human–AI Collectives

Metacognition in AI is usually pursued inside an architecture: a system that monitors and controls its own reasoning. We report a working alternative — institutional metacognition — in which those functions operate across a human–AI collective, through a shared versioned record: a further layer over whatever metacognition the participants already carry, not a replacement for it. The practice is deliberative programming: the deliberation itself — questions, alternatives, reasons, confidence with grounds, revision conditions — is authored as a first-class, versioned input to the work, colocated with the artifacts it governs. Its primary unit of record is an addressable decision question; decision-ready questions carry the ORRCF protocol (pronounced "orr-SEF"): Options, a Recommendation, a Rationale, a Confidence stated with its reasons, and a Falsifier. The record holds the persistent meta-state: what was decided, on what grounds, and under what conditions to revisit. Monitoring and control happen around it — agents read the record before acting, committed scripts check it, and a human gate rules on what crosses outward — supported by a toolchain over markdown and git that runs today, from an authoring template to a phone client for accepting or overriding recommendations. We then report two case studies: a public repository built in a few hours under this convention by a different vendor's model on different infrastructure, onboarded through repository artifacts and governed by eight questions answered from the author's phone (n=1 project, one day, founder-guided); and a retrospective reconstruction assembling decision grounds dispersed across a package's commits and pull-request discussions. We report the loop's own failures alongside its uses, grading every claim as implemented, observed, or open.