A trusted kernel forms the final acceptance boundary of a proof assistant. More complex tools may parse code, search tactics and generate candidate proofs, but the kernel applies a much smaller set of rules to decide whether the resulting proof term is valid.
Keeping the kernel small reduces the amount of software that must be trusted for the logical guarantee. A defect in an untrusted proof generator should produce a rejected proof, while a defect in the kernel could undermine every accepted theorem.
ELI5
A trusted kernel is the small final checker inside a proof system. Other programs can search for a proof, but the kernel decides whether every step follows the allowed logical rules.
For example, a complicated AI might suggest a proof of a math statement, while the trusted kernel checks it step by step before accepting it. Keeping this checker small makes it easier to inspect and reduces the number of bugs that could allow a false proof through.
Why should a proof kernel be small?
A small kernel is easier to inspect, test and reason about, reducing the trusted computing base responsible for accepting every proof.
Does the trusted kernel search for proofs?
Usually no. Tactics and external tools search for a proof, while the kernel checks the completed proof term against the logical rules.

