Skip to content

Choose which cookies and similar technologies this site may use. The strictly necessary ones keep the site working and remember this choice, so they cannot be turned off; they store nothing that identifies you or follows your browsing. Read the cookie notice

Continuumship
Menu

Guardrails

Rules a team cannot remove and an agent cannot argue with.

Write a rule on Monday. Read on Friday what it would have refused that week, then enforce it. Every refusal names the rule, the fact it matched and who to ask.

A rule written as where, when and then, tried in shadow before it is enforced.

Why guardrails

Policy that holds when the work moves faster.

  • Our rules live in a document agents never read.

    Two approvals before production is a sentence on a wiki. Nothing stops the run that skips it.

  • Teams remove the check that slows them down.

    A required security step disappears from a pipeline, and we learn about it in the audit.

  • Nobody can say why something was blocked.

    A developer is refused, opens three logs and asks in a channel. The answer takes a day.

How it works

One catalogue of facts, every point of enforcement.

Rules over shared facts

A rule names where it applies, conditions over one shared predicate catalogue, and one effect from a closed list. There is no allow. Every rule runs in shadow first.

  • Enforced at run, step, approval, publish and deployment
  • Deny, require an approval, a step or evidence, restrict the executor
  • Locked steps injected into every flow the rule governs

Change risk from named rules

Change risk is the highest answer of named rules, never a score, and it says why. Verification profiles pick the checks each class of risk must pass.

  • Your pipeline's jobs mapped to checks, for the exact commit
  • Architecture, interface and migration checks on the runner
  • SBOMs and provenance statements bound to their artifact

Your engine, your exceptions

Bring OPA, Cedar or another decision point as a capability that can only tighten a decision. Exceptions are asked for, approved by a named person, and expire.

  • Every evaluation recorded as a decision
  • Why this was refused, answered in one request
  • An exception ends on its expiry date

Try a rule

Test a rule before it is enforced.

Describe a change and watch three rules from one catalogue of facts decide it. In shadow nothing is blocked: the decision says what would have happened, and it is recorded.

Change risk
Written by
Target
Approvals from people
Mode
Change risk
High
Written by
An agent
Target
Production
Approvals from people
1
Mode
Enforce

Requested Wed 11:40, inside the release window Tue–Thu 10:00–16:00.

PAY-142 to Production

High risk Agent One approval
  • Failed Two approvals for high risk in production two-approvals-high-risk
  • Passed A person approves an agent’s change agent-author-needs-person
  • Passed Production deploys inside the release window deploy-window

Refused

Needs one more approval before it reaches production. Ask release-approvers.

Recorded, with the rule and the facts it read.

Refused by two-approvals-high-risk. Needs one more approval before it reaches production.

Questions about guardrails

Can a team switch off a guardrail it dislikes?

No. A workspace rule applies to every flow it governs, and a locked step cannot be removed from a team's flow. A team can ask for an exception, which a named person approves and which expires.

Will a new rule block work the day we write it?

No. A rule runs in shadow first and records what it would have required and refused. You read those decisions against real work, then choose to enforce it.

Can an agent talk its way past a rule?

No. Rules are evaluated over facts the platform records, not over what an agent says. An agent's claim is labelled as reported and never counts as a measurement.

Where are rules evaluated, and who decides the exceptions?

The platform evaluates rules at each enforcement point. Architecture, interface and migration checks run on your runners, where the code is. Your policy engine can only tighten a decision. The people a rule names decide its exceptions, and every decision is recorded.

Design how your company ships software.

Start with a thirty-day shadow pilot on one Area: connect, observe, read the report together, decide.