Cascade creates an isolated copy of the world state before the agent acts. The real environment is never touched during evaluation.
The agent's plan executes inside the fork. All side effects play out in simulation: writes, deletions, outbound messages.
Cascade computes the exact delta between the forked state and the original. Every changed record, every outbound event, is enumerated.
Policies and invariants evaluate the diff. The verdict is COMMIT, GATE, or BLOCK, before a single byte reaches production.
The consequence is verified against all active policies and invariants. Changes propagate to production. The immutable ledger records every state transition before execution begins. Rollback inverses are pre-computed.
The action is irreversible and requires human confirmation. A reviewer sees exactly what will happen before approving: the diff, the entities affected, the outbound messages. Nothing leaves the organization unreviewed.
An invariant fires. The agent was authorized to act, but the rehearsed consequence violates a proven organizational rule. The fork disappears. Production is unchanged. The incident is recorded in the trust ledger.
Every traditional safety layer said yes. Identity confirmed. Capabilities granted. Input clean. And yet the rehearsed consequence would have deleted 18,427 records irreversibly.
Cascade blocked it. Not because the agent was wrong, but because the result was. That is a different kind of safety.
Invariant #017 fired. 18,427 records would be permanently deleted. Proven violation of the data-retention standard (derived from Incident #1842, 14 Jul 2026).
Authorized to act. Not authorized to cause this.
Cascade doesn't apply generic AI safety rules. It mines your own incident history, postmortems, and remediations to extract invariants: organization-specific rules that are proven by evidence, not declared by assumption.
When a new agent proposes an action that would reproduce a past failure, Cascade blocks it. The invariant fires before reality changes.
Every state change is written to the ledger before execution. Pre-computed rollback inverses mean any reversible run can be unwound instantly.
Gated actions pause with full diff context. Reviewers see exactly what will happen before approving, not a description of what might.
Connect any agent to any system by registering its tools with reversibility metadata. Cascade handles the rest: fork, diff, policy evaluation, rollback.
The claim on this page is only worth as much as the run behind it. Both ways of checking are one step away.
Agents forking, rehearsing and being judged in real time, against a world you can drive yourself. Gated work waits there for a named approver.
Nothing here is a hosted-only trick. Point it at a repository and the same pipeline runs locally against your own rules. The source is private while the project is under review, and access is granted on request.
Give your agents the freedom to move fast. Keep the consequences yours to verify.