Skip to content
Control

AI can propose. Zof determines whether it has earned the right to act.

The Control Plane combines change, system context, risk, evidence, and enterprise policy into an enforceable decision, applied where actions actually happen.

Enforceable, not advisoryHuman approval where policy requires itWorks for any proposer
How a decision is made

Five inputs. One enforceable decision.

For actions routed through Zof, the Control Plane evaluates the proposed change, system context, risk, evidence, and enterprise policy.

Sample decision · illustrativeInputs → Decision
Inputs
Change

What is being proposed, by whom or what, and in what scope.

System context

What the System Graph says the change can affect, and how critical that is.

Risk

Criticality, blast radius, and the incident history of the affected paths.

Evidence

What the Verification Engine has proven, and what is still missing.

Enterprise policy

The rules your organization set: required evidence, approvers, change windows, data boundaries.

Decision
Require approval

Technical requirements satisfied. The Payments Tier 0 policy requires a named human approver before release.

Allow · Deny · Require evidence · Require approval
Decision vocabulary

Four decisions today. Finer-grained control as policy matures.

Decisions are explicit and machine-enforceable. The initial vocabulary covers the cases most enterprises need first; advanced decisions shape how an allowed action runs.

Initial decisions
Allow01

Evidence and policy are satisfied. The action may proceed.

Deny02

Policy forbids the action, or evidence shows it should not run.

Require evidence03

Specific verification is missing. The action waits until it is produced.

Require approval04

Technical requirements are met, but policy demands a named human decision.

Advanced decisions
Limit01

Allow with constraints on scope, time, or environment.

Canary02

Allow a bounded rollout and observe before widening it.

Rollback03

Reverse an action whose observed outcome violates policy.

Remediate04

Propose a fix, which then earns its own right to run.

Enforcement

Not another dashboard. Enforced where actions happen.

A control decision is only worth something if it can stop an action. Zof applies decisions at the boundaries where consequential actions actually run.

  • CI/CD pipelines

    Merge gates on GitHub.

  • Deployment infrastructure

    Not available today.

  • Autonomous agent tooling

    The tools coding agents and autonomous systems use to act, so proposals wait for a decision.

  • Other action boundaries

    Not available today.

Governance inside the plane

Policy, roles, and approvals are the primitives.

RBAC, SSO, named approvals, autonomy levels per environment, data boundaries, and audit trails are how your organization expresses policy to the Control Plane.

Explore the governance layer
  • Define distinct responsibilities for proposing, verifying, and approving changes.
  • Route governed actions and rollbacks through configured policy decisions.
  • Keep decision records linked to the policy version used.
Why this matters more over time

The smarter the model, the more independent control matters.

As frontier models and coding agents take on more consequential actions, the question stops being whether they can act and becomes whether they should. An independent layer with system context, evidence, and policy is how enterprises answer that question.

Read the Reliability Thesis

In practice

Separate verification from permission to act

Verification describes what the checks establish. Control evaluates whether the proposed action satisfies the configured policy. The distinction matters when tests pass but an approval is missing, or when evidence is incomplete for a consequential action.

  1. 01

    Describe the action precisely

    Identify the change, target environment and requested action. A permission to review or propose a patch should not be treated as permission to merge or deploy it.

  2. 02

    Evaluate the applicable requirements

    Compare the available evidence and authorization with the configured policy. The decision vocabulary includes allow, deny, require evidence and require approval. Confirm the enforcement points available for your workflow.

  3. 03

    Resolve the reason before retrying

    If evidence is required, complete the missing verification. If approval is required, route the decision to the appropriate reviewer. Reassess when the change or target has changed since the original decision.

Illustrative scenario · not a customer result

Passing tests, outstanding approval

A sensitive service change passes its selected checks. The release policy still requires an authorized reviewer. The team must resolve that approval requirement before proceeding through the configured control point; the passing checks answer a different question.

Questions to resolve before you begin

How is Control different from Governance?

Governance defines responsibilities, rules and approval expectations. Control applies those rules to a particular action. The Governance page helps your team decide what the rules should require.

Does publishing a report block deployment?

A report alone does not enforce a gate. Review how the decision connects to your actual merge or deployment path, and test the failure and bypass cases during implementation.

Next step

Put an independent decision in front of consequential actions

Book a Reliability Sprint to see the loop on one of your systems, or talk to Zof about your control model.

  • Human approval where policy requires it
  • Connect policy gates to your CI/CD and agent tooling
  • Private deployment options
01Zof Console

One surface for posture, operations, and what needs attention next.

The authenticated home that engineering, QA, and SRE teams open every day: quality posture, in-flight runs, coverage by module, and what needs attention next.

OPERATIONAL KPIs

  • Runs
  • Coverage
  • Risk

Live across every environment you ship to.

WORK SPINE

  • Specs
  • Tests
  • Schedules

From specification to scheduled regression.

GUARDRAILS

  • RBAC
  • SSO
  • audit

Every action attributable to a named human.

LIVE/console
Zof home command center showing 12 runs at 94% pass, 3 open critical issues, 84% coverage, four module traceability bars, the specification pipeline, upcoming schedules, and recommended next actions with an active-runs sidebar.
Example console home · Checkout Service · Staging.