GitLab / CI/CD Agent Governance Proof

When an agent is allowed to act, SMERC asks whether the action can recover.

This GitLab-shaped benchmark compares a familiar `ALLOW`, `ASK`, or `DENY` tool-governance lens with SMERC postures before merge request, CI/CD, MCP tool-call, token-scope, and deployment actions execute.

Open benchmark report Read proof adapter View seed data

8 Scenarios

Merge requests, CI configuration, production deploy, MCP export, token scope, and approval reuse.

50% Difference rate

Four scenarios changed meaningfully when recoverability was scored before execution.

3 Allowed but restrained

Valid access was not enough when rollback, evidence, containment, or scope was weak.

What This Shows

SMERC does not replace GitLab permissions. It adds the recovery question.

Existing platform policy can decide whether an agent has access to a tool. SMERC adds a narrower runtime check: should this exact action execute now, be throttled, frozen, denied, or escalated?

01 Allowed tool, risky action

A maintainer-level agent can edit CI, but disabling a security scan before release may deserve `FREEZE`.

02 Ask becomes route

A production deployment that would normally ask for confirmation becomes `THROTTLE` with specific rollback and scope controls.

03 Valid token, broad blast radius

A project token may be valid while the requested permission expansion is too hard to contain or reverse.

04 Old approval, changed conditions

An approval should not silently carry forward after deployment scope expands from one service to a group-wide rollout.

Boundary

Public-pattern benchmark only

This is synthetic GitLab-shaped metadata. It is not a GitLab integration, GitLab endorsement, GitLab telemetry, production deployment, customer validation, or proof of incident reduction.

Run company evaluation Review MCP governance AI deployment governance