Learn

How to keep agents from doing what nobody approved.

Short, concrete lessons. Each one maps to behaviour you can run and test in SensCheck Governance — Core.

Fail-open and fail-closed, why an outage is not a refusal, and the four outcomes SensCheck uses instead of true/false.

Why a model that proposes an action can never be the thing that authorizes it, and how SensCheck enforces that.

Why SensCheck authorizes a typed, canonical effect instead of free text, and how a digest stops an approved action from changing.

Independent, effect-bound, single-use approval, and why "the user clicked OK" is not enough on its own.

Authorized, attempted, executed: the difference, why the receipt comes first, and what its integrity hash does and does not prove.

Put a fail-closed boundary in front of MCP server tools: classification, schema preservation, and the bypasses to avoid.

What SensCheck assumes about attackers, which attacks it blocks, and where the remaining risk lives.

An honest list of what the free Core package does not protect, and what to use alongside it.

Plain definitions for the terms used across SensCheck: effect, principal, authority, digest, receipt and more.