Keeping your Kastel in good shape.

In the console (hosted plans, or self-hosted Enterprise): the “Upkeep” screen, reached from the health card on My Kastel. Open to everyone: each person sees their own scope.

On the command line (free self-host): kastel doctor, kastel lint, kastel ingest failures, kastel jobs ls.

Company context does not decay all at once: it frays. Pages stay started and never finished, others lose their update date, some are no longer linked to anything and become unfindable, a fact reaches its expiry without anyone noticing.

The “Upkeep” screen lists exactly those four things, with the pages concerned: pages to finish, pages with no date, pages never linked, facts expiring within thirty days. Everyone sees the ones in their own scope. When there is nothing, the screen says so; when it could not read, it says that too, and does not show a zero instead.

The same screen carries an “Operational health” section: your Kastel's machinery, in plain terms. Decisions recorded this month, items awaiting approval, budget breaches, syncs stopped by the cap, jobs queued or failed, pages to reindex, age of the last backup. And above all the list of documents not read: those the last syncs could not read, with the reason in plain language - the source did not answer, access was refused, the document could not be read.

Every week, your Kastel asks a few questions of the people who can answer them: “is this still current?”, “is this page complete?”. One question, one click to confirm, or one sentence to correct. A correction goes into approval like any other edit.

Questions arrive in “To approve”; the notification carries only a count, never the content. Everyone chooses to be told by email, by Slack, or not at all. A warning is shown for Slack: a channel is read by everyone in it - prefer a private one.

It is the product's highest-return mechanism and its least spectacular: five minutes a week, spread across the right people, beats an annual clean-up nobody performs.

Your Kastel maintains its own machinery without bothering you: it rebuilds its index when it falls behind, recomputes its indicators, spots answers that cited a page that has since gone. None of that touches any content.

The moment it comes to touching a page, it never does: it proposes. Linking an orphan page to another, adding a missing update date, flagging two pages that say almost the same thing, moving a misfiled section: each of those corrections arrives in “To approve” as a proposal, with its evidence, and waits for a human. You approve in one click, or you ignore it.

This is the promise “your Kastel improves” in its exact version: it improves under your control. It consolidates nothing on its own authority.

Two mechanisms go beyond form and look at substance. They are part of the six functions that only run if you switch them on.

• Contradictions: your Kastel compares overlapping statements and flags the real conflicts. Each conflict arrives in “To approve” as a choice: these two pages disagree, which one holds? On the command line: kastel contradictions report for the history, kastel contradictions scan for the free mechanical check, kastel contradictions run for the full pass.
• Statements: who asserted what, with what certainty, on which page. That is what lets an answer cite rather than assert. The “Who said what” screen lists them and lets you propose one to its destination page - propose, never merge: the edit goes through “To approve”. On the command line: kastel claims list, search, review, promote, and kastel claims extract --quote to learn the cost of a reading pass before launching it.

How solid a statement is shows as a band - strong, medium, weak - never as decimals: that would promise a precision nobody has measured.

The Context Score measures one precise thing: your AIs' ability to retrieve the right company context, on real questions from your business. It is not a satisfaction rating, and it is not a usage indicator.

Three rules make it readable, and they are not negotiable:

• Never a bare figure. The score is always shown with its breakdown and with the number of questions it was measured on.
• A floor of ten questions. Below that, your Kastel shows “not enough data” rather than a figure. A score on three questions is not a score.
• Your questions, not ours. The question set belongs to your deployment. You grow it with your real questions and the pages that ought to answer them.

The measurement runs on your instance, whatever your plan; what hosted plans add is its tracking in the console. On the command line: kastel score to measure, kastel score questions add and kastel score questions list for the question set. The measurement spends your AI keys: it is quoted before each run, and does not start below the floor of ten questions - under it, nothing runs and nothing is spent.