Trust but verify: the ledger, stats, health, and privacy
Cognition is autonomous by default and inspectable on demand. Nothing it does is a black box. Here are the four ways to look inside whenever you want, and why being checkable, not asking for faith, is how good automation earns trust.
Automation that runs in the background has a trust problem to solve: if you cannot see what it did, you either trust it blindly or not at all, and neither is good. Cognition's answer is that everything it does is auditable from inside your agent, in plain English, at any time. You do not earn trust by being told to relax; you earn it by being checkable. There are four windows in, and they are worth knowing all of.
The attribution ledger: what it actually did
Ask "what did Cognition use?" and you get the call ledger for the session: every activation, in order, with the trigger that fired it, the sources it loaded, the action it took, the result, and a timestamp. This is the ground truth of what happened, not a summary of what should have happened.
The ledger records the fast no-matches too, the moments it deliberately stayed quiet and spent nothing. That is deliberate: silence is logged, so you can always answer "why did it do nothing here?" instead of wondering whether it was even running.
Stats: the value receipt
Ask "show me my stats" for tokens saved, hours saved, hit rate, and skill-tree value, each with its basis shown rather than asserted. Accept the follow-up to render it: an ascii dashboard that works in any host, or a mermaid chart for markdown contexts. The basis matters, a number you can trace beats a number you have to believe.
Watch these over a week of real work rather than a single session. Value concentrates in repeated, known-territory tasks, so any one day is noisy; a week shows the real trend.
Health: is it really persisting?
Ask "is Cognition healthy?" to confirm the backend is genuinely recording, not just simulating. It checks that migrations and ledgers are in place and that behavior is persisted. This is the meta-check: it tells you whether the stats and receipts above are real records or safe-but-not-persisted placeholders. If you are going to trust the other three windows, this is the one that backs them.
Privacy: redaction before anything is saved
Before any memory that might carry secrets or PII is stored, ask "is this safe to save?" It runs a redaction-first review: it shows only redacted previews and lets you decide, per flagged item, whether to redact it, suppress it, or explicitly keep it. The default leans safe, sensitive values are redacted unless you consciously choose otherwise.
Raw secrets, keys, and credentials are redacted before saving by default, but you are the backstop. If you ever see a sensitive value sitting in a draft, stop and run the privacy review rather than approving it as-is. The brain is shared; a leaked secret in a skill is a leaked secret for everyone on the org key.
The five-second trust ritual
Do this once and the black box is gone for good: run the ledger, then the stats, then health, back to back. You will see what fired, what it saved, and that the backend is actually recording it. Five seconds, and "trust the automation" becomes "I checked the automation," which is the only kind of trust worth having.
next steps