-
Notifications
You must be signed in to change notification settings - Fork 0
World Manifest
sv-pro edited this page Apr 30, 2026
·
1 revision
The World Manifest is the formal specification of what exists in an agent's universe. It is closer to a constitution than a configuration file: it defines the action space, trust model, capability matrix, and non-overridable invariants that govern execution.
| Area | Role |
|---|---|
| Base ontology | Actions, schemas, and parameters that can ever exist |
| Trust channels | Mapping from sources such as user input, email, memory, or tools to trust levels |
| Capability matrix | Which trust levels or actors may use capabilities such as read, write, memory, or external side effects |
| Taint rules | How contamination propagates through data and across sessions |
| Invariants | Laws such as taint containment that cannot be overridden by runtime agent reasoning |
At design time, the manifest is compiled into deterministic runtime artifacts:
- schema validators
- capability projection tables
- taint propagation structures
- policy lookup tables
- provenance requirements
Runtime enforcement uses those artifacts directly. The LLM may help author or refine the manifest, but the LLM is not on the critical enforcement path.
The manifest lets the system distinguish between:
- actions that are allowed
- actions that are explicitly denied
- actions that require human approval
- actions that are absent from the current world
This is the foundation for deterministic execution governance.