In the console (hosted plans, or self-hosted Enterprise): Settings, then “What runs on its own”. Switching on and off is reserved for owners and administrators; everyone else sees the state without being able to change it.
On the command line (free self-host): kastel automations list, kastel automations on, kastel automations off.
This screen is conditioned by no plan, and never will be: we do not sell the right to stop a spend.
Nothing switches itself on
No function that spends your AI keys starts without an explicit gesture from you. All six are off on delivery, and stay off until someone at your company switches them on.
The reason is sovereignty, not thrift. These functions cost little; that is not the argument. The argument is that an AI re-reading your context on its own initiative must have been told to, by you - otherwise the question “what does your tool do with our documents when nobody is watching?” has no good answer.
The flip side is stated just as plainly: while these functions are off, your Kastel stays heavily degraded. It learns nothing from what your AIs do, watches for no contradiction, and its quality figure goes stale. At the end of the build, your Kastel recommends switching all six on, and the most important is lesson compilation - that is the one that makes it improve.
The six
| Function | What it does | What it spends |
|---|---|---|
| Weekly contradiction probe | Every week, re-reads pages that might contradict each other and flags the real conflicts before a colleague runs into one | your AI keys |
| Check on every sync | As soon as a connector brings in new content, checks straight away that it contradicts nothing existing, instead of waiting for the weekly pass | your AI keys |
| Statement attribution | Shows, inside answers, who said what and with what certainty | nothing by itself (see 04) |
| Deterministic contradiction scan | A mechanical check that spots obvious conflicts of value, type or date. Works even with no AI key configured | nothing, ever |
| Lesson compilation | Distils what your AIs have learned into reusable lessons. The function whose absence degrades your Kastel most | your AI keys, outside the monthly cap |
| Weekly Context Score measurement | Re-measures the Context Score every week so the figure on screen stays fresh | your AI keys, quoted before each run |
This list is closed. A seventh function that would switch itself on and spend your keys would be a decision of its own, not a convenience added in passing.
Switching on, off, and when it takes effect
Every toggle is recorded in your audit log, with its author, and sealed. Changing a setting is a governance decision: it is traced as one.
The effect is not instant everywhere, and your Kastel says so on each line rather than letting you believe otherwise: most functions take effect on the next upkeep pass, within the hour at the latest; the check on sync takes effect at the next connector sync; statement attribution, at your instance's next restart.
Switching off and back on restores the standard cadence, not necessarily the exact value you might have set by hand on a self-hosted Kastel. That value is in the audit line, so it is recoverable.
Three traps, stated plainly
- The free scan follows the probe. The deterministic scan does not run while the weekly probe is off, and your Kastel refuses to switch it on alone rather than write a setting that would be a lie on screen. Switch the probe on first. It never switches the probe on for you: the probe is the one that spends.
- The check on sync, however, stands alone. You can therefore have the weekly probe off and the targeted check on: that is a legitimate state, not a bug.
- “Statement attribution” extracts nothing. That toggle opens the display of attributions inside answers; it launches no reading and spends nothing by itself. Reading the statements is a button, pressed by an owner or an administrator, after its estimated cost is shown. Nothing fills itself in.
What the checks found
The same screen shows what these functions spent and what they returned: how many questions hit a contradiction on the last pass, how many conflicts went into “To approve”, what share of comparisons was re-read from cache and therefore not paid for twice.
A rate is never shown on its own. It always comes with its 95% confidence interval: “X of N questions hit a contradiction, between A and B% in reality”. On a small sample the interval is wide and the screen says so - it narrows pass after pass. A figure without its interval is not a figure you can act on, and your Kastel refuses to present it as one.
You can also run a check on demand, without waiting for the weekly pass. The estimated cost is shown first, as a range, with what is left of your monthly cap; if the upper bound does not fit, nothing starts. The mechanical check runs without a quote: it costs nothing, ever.
What these functions produce - contradictions to settle, attributed statements, the Context Score - is read and worked on from your Kastel's day-to-day upkeep.
Read the Keeping your Kastel in good shape page