How to run a research session

A research session is goal-driven: the orchestrator runs toward a measurable session goal — a verifiable completion condition — not an open-ended prompt (design spec §2, §6b).

The engine agents and commands referenced here are delivered in Milestone 3 (Engine). This guide describes the intended flow and is the contract those commands implement.

1. Author the goal

Turn your raw ask into a measurable goal with the goal-writer command. The goal declares the decision the research must enable, what is in and out of scope, and the checks that gate “done”.

2. Start the session

The orchestrator owns phase management. It spawns parallel dimension-analysts (one per configured dimension), each researching independently and emitting MIF-backed findings validated against schemas/findings.schema.json.

3. Falsify

Exactly one adversarial gate runs: the falsification-analyst treats each finding as a hypothesis, searches for disconfirming evidence, and assigns an ordinal verdict (falsified / weakened / survived / inconclusive). Falsified findings are quarantined; weakened ones have their confidence lowered.

4. Synthesize and publish

report-synthesizer consumes the surviving findings. Outputs render through the typed findings→artifact contract — blog is first-class; book and other channels arrive via optional channel packs.

In the same phase the orchestrator reconciles the topic’s navigation README.md so its counts, dimensions, key findings, and report table stay current — see Maintain topic READMEs.

Continuity

Progress is written to a progress file on every phase transition, so a session can be resumed after interruption.