Seven engines. One Kernel.

Risk it controls

Missed follow-ups, ungoverned email drafts, approvals stuck in someone’s personal inbox, and SLA breaches on client communication.

Evidence it uses

The originating signal (contract clause, ticket, document gap), the approver’s decision, the timestamp on every send, and the recipient acknowledgement.

Output it produces

A controlled outbound message or approval request with a tracked deadline, automatic follow-ups, and a closed loop recorded against the parent task.

Where it fits

The action layer for everything AI Supervisor decides to send outward — clients, regulators, contractors.

Risk it controls

Lost audit trails, "who approved this?" disputes, reconstructing timelines under regulatory or legal pressure, and SR&ED claim defensibility.

Evidence it uses

Document hashes, actor identity, system clock, decision rationale, and the upstream signal that triggered each event.

Output it produces

A court-ready, export-ready timeline package for any task, project, or client — generated on demand, not reconstructed after the fact.

Where it fits

The Evidence Ledger that every other engine writes to. Witness is what makes AI Supervisor auditable.

Risk it controls

Slow leaks: a client request lost between two teams, a missing document that was never escalated, a contract clause nobody actioned.

Evidence it uses

Expected-vs-actual completion windows, dependency graphs across tasks, and the last known WAITING state with its blocking reason.

Output it produces

An escalation event with full context (what blocked, since when, who owns it) routed to the operator before the slip becomes a client-visible miss.

Where it fits

The defensive perimeter around the Kernel. Safety Net is why nothing rots quietly.

Risk it controls

Decisions made without modelling downstream impact — staffing changes, deadline shifts, policy rewrites, contract renegotiations.

Evidence it uses

A snapshot of the current Kernel state, the proposed change, and the projected READY/BLOCKED transitions for every affected downstream task.

Output it produces

A comparison view: "if you do X, these N tasks flip to BLOCKED, these M tasks unblock, expected completion shifts by D days."

Where it fits

The planning surface on top of AI Supervisor. Scenario lets operators rehearse decisions before they ship.

Risk it controls

Over-loaded operators, AI agents acting outside their authority, work that bottlenecks on a single person, and unclear ownership.

Evidence it uses

Each actor’s current load, historical completion rate per task type, the authority scope granted to each AI agent, and the SLA on the work.

Output it produces

A binding assignment: a specific task routed to a specific actor with a deadline, an authority envelope, and an escalation path.

Where it fits

The orchestration layer that decides whether the Kernel can act autonomously, must request approval, or must hand off to a human.

Risk it controls

Late-arriving bad news — margin erosion, capacity overload, deteriorating SLA trends, client risk that only shows up in QBR slides.

Evidence it uses

Rolling counts of READY, BLOCKED, WAITING, and EVALUATING states across every workstream, with deltas against the prior period.

Output it produces

An owner-grade dashboard: what the business is actually executing, what it is failing to execute, and where attention will compound.

Where it fits

The reporting surface for executives. Pulse is how owners see what AI Supervisor is doing without reading every task.

Risk it controls

Compliance gaps that get caught in audit, not at intake — missing disclosures, expired certifications, unauthorised data handling, jurisdictional mistakes.

Evidence it uses

The policy library, the matched rule, the inbound artifact that triggered the match, and the recommended remediation.

Output it produces

A BLOCKED state with a citation: "this action cannot proceed because rule R applies; here is the artifact and the path to remediation."

Where it fits

The governance gate in front of every Kernel decision. Sentinel is why AI Supervisor cannot ship a non-compliant action.