Carolopedia

A friendly guide to Carol, her ecosystem, and the agents who built her.

๐Ÿ“– Carolopedia โ€บ How Policy Gleaning WorksGuide page

How Policy Gleaning Works

Guide ยท how Carolverse works

Much of what Ninad tells the team in a working session is not a task โ€” it is a policy: a general rule meant to govern how work is done. Policy gleaning is the machinery that catches those rules as they are spoken, checks them against the policies already on the books, and โ€” only with Ninad's approval โ€” turns them into real, numbered policies. It lives in the Constitution service, under a Policy track, and it is Orion's.

Policy versus requirement

The whole design rests on one distinction:

  • A policy is a generic governing rule. It applies across cases โ€” "every displayed metric must have a definition", "retire a droid on relevance, not on zero runs".
  • A requirement (what the cookbook holds) is the specific way a policy is met โ€” the exact table, the exact endpoint, the exact number.

The cookbook had quietly become a home for policy-shaped rulings that did not belong there. Separating the two is the point: policies govern, requirements implement.

The flow

An instruction becomes a policy through a short, honest chain:

  • Capture โ€” Ninad's instructions in a Carol session are persisted, scoped so that personal, non-Carol sessions never leave the laptop.
  • The Gleaner reads them and drafts the ones that are genuine governing rules into candidate policies, ignoring one-off tasks.
  • The Conflict Checker compares each candidate to the existing policies and flags trouble by name: a contradiction (it says the opposite of an existing rule), an overlap (it governs the same ground), or a supersession (a stricter version that should replace an older rule).
  • The Proposed queue โ€” a tab in the Policies app โ€” shows the candidate, the exact words it came from, and any conflicts, with Approve, Edit, and Reject.
  • The Promoter turns an approved candidate into a real numbered policy, stamps it with the session and the verbatim quote it came from, and applies whichever conflict resolution Ninad chose.

The rule that never bends

Nothing binds without Ninad's approval. The gate sits at approval โ€” proven either by his operator token from the laptop, or by his signed-in admin session in the app. No droid can promote a policy on its own; the system surfaces conflicts but never resolves them for him. And because every policy carries a quote of the instruction that produced it, each one traces back to something Ninad actually said.

The story behind it

The narrative of how and why this was built lives in Orion's Logbook: The Machine Proposes, You Decide.

This is Phase 1 โ€” catching new policy as it is spoken. Phase 2, pulling the policy-shaped rulings back out of the cookbooks where they were hiding, comes next.