MeridiansMeridians

The same practice, headless — one loop, three drivers

From the Meridians Wiki · Public · Maintained · joint

The Program feed is produced the same way whether a human runs a selected day in the studio, an autopilot schedule fires it, or an AI drives it over MCP. Domain and constellation Run now surfaces divide the weekly agenda into independently selectable days, preserving each member domain as the owner of its work. There is no second mechanism for headless: the MCP surface is a 1:1 clone of the human practice, so an AI chat can replicate every behaviour above. Each human-loop stage maps to a concrete op or read — this is the contract an AI agent follows to run the loop the way a Director does.

Position is an explicit, user-created domain cycle; there is no default generation clock. The operator authors its question, sample horizon, and eligibility cadence in Operations. Each eligible run creates one fresh dated Graph → Reading → Position proposal, then waits while that sample is live or awaiting judgment. Position monitors remain the faster Research path that re-prices the already-live thesis as new evidence lands. Once a sample closes, a later Position-cycle window may produce the next comparable sample.

Human stageWhat the person doesHeadless (MCP) equivalent
Run the feedmanual "run now" or an autopilot tickrun-domain-loop (one domain, Research→Opinion→Merge→Position) · run-constellation-program (a whole portfolio) · or a single task (research-run, run-loop-opinion, run-loop-tutor, run-loop-project)
Orientswipe the feed; see what changedread-stories (the queue as the reader sees it — each card's consequence, directiveId, drill, and this member's receiptState) · read-loop · whats-changed
reach caught-upswipe past a cardmark-story (state: read / deferred / dismissed) — receipts are per-member delivery state, kept out of the model; read-stories reflects them so the queue goes finite
Judgeinspect a proposal, then accept/correctread the proposal CONTENT in read-loop.proposals (tutor Merges + Position candidates / re-priced reads, with rationale) → stage-loop-directive (optionally with a payload edit) → the owning task run applies it; cancel-loop-directive backs out
Steerpoint a read at a questioncreate-projection (a priced Reading + a monitor; grounded from a graph→reading, a saved reading, or a direct outline) · read-projections / read-projection (the read, its priced Reading, and how it moved)
Learnwatch the read re-price; read how it movedresearch-run feeds Signals that reposition the active Position's read; read-projections shows the priced Reading and the track of how the read moved; the Position reasoning is not reduced to true/false, while the position it argues for is scored and calibration is measured across positions
Refinefold a mature belief into canoncontribute-stream-prior (add evidence) → commit-merge / run-loop-tutor (fold a mature Stream into continuity, generating the assessment arc)
Configureset cadence, autonomy, spendconfigure-loop (enable / guided-vs-autonomous / per-task overrides / budgets / timezone) · apply-schedule-preset (both spend levers at once) · create/update-research-cycle · update-projection-monitor
Rememberre-open a past chat; review what the Pilot didread-chats (domain / branch / constellation threads — list, open one transcript with its Pilot toolTrace, or search across them to collect information); write via UPSERT_CHAT_THREAD

The affordance chain teaches the practice, not a flat op→op map: a run's result points at read-loop / whats-changed (orient first); orientation points at the proposal or directive only when something material moved; a Position points at research-run to see the next Signal that re-prices its read. An AI should escalate to Judge/Steer/Refine only at the consequential boundaries — a story card whose consequence is action, a proposal awaiting judgment, contested evidence — exactly as a person does. Most passes end at Orient + mark-story: caught up, nothing to do.