> ## Documentation Index
> Fetch the complete documentation index at: https://docs.gloo.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Governance pipeline

> How Trust Fabric evaluates dispatch through stop controls, identity, resources, policy, and approval.

Trust Fabric evaluates current facts before protected work. The governance engine returns a recorded decision; the platform supplies current identity, role, permission, budget, policy, and approval facts for workflow, Convex-native, and procedure dispatch.

Governance evaluation

Stop controls fail closed first. The engine then proves authority and resources before applying scope, policy, and approval. The first enforced refusal ends evaluation; a required human decision returns a hold.

<ol aria-label="Governance evaluation" />

## Ordered checks

A single fixed gate sequence declares the order. The engine groups it into phases without changing that order.

| Checks | Proof | Result when unsatisfied |
| - | - | - |
| `globalKillSwitch`, `doorSwitch` | Platform and caller class accept new work | Block while stopped |
| `agentStatus`, `factoryStatus`, `identity`, `permission` | Subject exists and the canonical principal has authority | Block; pause can be retryable |
| `concurrency`, `rateLimit`, three budget checks | Dispatch fits operational and financial ceilings | Block, sometimes retryable |
| `autonomyHoldTable`, `nhiEntitlement` | The action fits autonomy and non-human identity rules | Pass, block, or hold by path |
| `scope`, `policyRules`, `approvalRequired` | Scope and targeted policies allow the concrete action or name its release | Block or hold |

There is no gateway-health gate. Workflow launch permission is separate: a launch check decides whether a run may be created, while this engine judges each protected dispatch, inside an admitted run or on another dispatch surface such as chat or a factory.

## Pass, block, and hold

* **Pass:** every applicable enforced check allowed the dispatch; the owning runner may continue.
* **Block:** the first enforced refusal ends this attempt; later applicable gates are recorded as skipped.
* **Hold:** the action is otherwise admissible but needs a human release; the owner creates or reuses an approval human task and parks durably.

A hold is not a soft pass. Completion rechecks the approval's release policy before recording a decision. The release applies to the concrete governed attempt, not later calls.

Policy clauses are parsed at authoring time against whitelisted facts. Exact targets use server-resolved identities, including a run's persisted workflow version. External-delivery policy waits until the concrete operation is classified, then runs before credentials or effects.

Every decision records gate outcomes, actor and role snapshots, applicable limits, and the first cause. `not_applicable` means a check was outside this path; `skipped` means an earlier refusal cut it off. Retries re-resolve mutable authority unless a durable contract explicitly freezes it.

See [Approval lifecycle](https://docs.gloo.com/forge/run/approvals-and-hitl.md#approval-lifecycle) and [Workflow execution](https://docs.gloo.com/forge/architecture/flows/workflow-execution.md).

## Related

* [Trust Fabric concepts](https://docs.gloo.com/forge/concepts/trust-fabric.md): the gate sequence and the four rungs.
* [Approvals and human tasks](https://docs.gloo.com/forge/run/approvals-and-hitl.md): what happens after a hold.
* [Workflow execution](https://docs.gloo.com/forge/architecture/flows/workflow-execution.md): where a run's dispatches meet the pipeline.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.