Project Keystone · Internal reference
The knowledge layer, the conductor, the specialist assistants, and the rails they all inherit — laid out end to end. Every claim here is quoted from a source document and carries the state that document gave it, so you can tell a ratified decision from a slide nobody has approved yet.
Four different kinds of thing, not four rungs of a ladder. Select any part to see what the source documents actually say about it.
Three models were on the table at scoping. Two were rejected by name, and knowing which ones matters — an assistant that only reports rebuilds the model the team threw out.
The ratified roadmap, transcribed with its own dates and statuses. The assistants on this site are one row of it.
The roster is not settled, and the scoping document says so about itself. Three of the five named assistants carry something the source marked UNVERIFIED, the naming question is open and unowned, and not one of them has a PRD yet.
That's fine — it's early. What isn't fine is a deck that reads as though it were all decided. So this library never flattens the difference. Every claim shows its state, quotes its source, and links to the page it came from.
Four ways in, depending on what you came for.
Eleven entries — five named in Confluence, six concepts proposed here against documented workflows that have no assistant assigned. Each carries a full evidence ledger.
Open the roster →Fourteen chat use cases — checking coverage, working a denial, quoting what a patient owes — each with illustrative prompts and the workflow behind them.
Open the chats →Every rail an assistant inherits: identity, access tiers, the provider-held ceiling and revoke right, per-task permissions — plus the ones still open.
Open the rails →The AIX v1 roster, targeted at Q4 2026 beta. Six further concept entries fill the documented workflows that currently have nobody assigned to them.
Stated up front, because finding out later is worse.