sovereign-agent

Durable non-goals

A non-goal is removed only by a recorded authorization and a matching CHANGELOG.md entry. It is not removed by a convenient pull request.

Read with roadmap.md (sequence) and API.md (what the 0.x line promises). Historical v0.3 refusals that still hold for 0.x are also in v0.3-non-goals.md.

The 1.x educational line is authorized by ruling 2026-08-25 educational reset. That ruling does not move tag v0.7.0 and does not rewrite the 0.x contract.


Through v0.7 (0.x line)

These items remain normative for sovereign-agent<1.

No Sandcastle

No Sandcastle dependency, adapter, service, or invocation path. The v0.3 prohibition in v0.3-non-goals.md remains in force through v0.7.

No governance decisions in this package (0.x only)

On the v0.5–v0.7 line, Sovereign Agent does not decide organizational authority, accept SOWs, or interpret Zero Employee policy. It executes under evidence and admissions it is given. ZeoCore also must not import Zero Employee governance code.

This non-goal is superseded for Sovereign Agent 1.x by holding 2 of the educational-reset ruling: 1.x may contain the minimum governance necessary to teach and run an end-to-end outcome. It remains in force for published 0.x releases.

No second reusable capability schema

Reusable actions on the 0.x line are ZeoCore capabilities. This package does not grow a parallel capability type system. Runtime evidence types (RuntimeCapabilityAssertion, RuntimeCapabilityManifest) are not a second authoring schema.

No generic workflow language

No graph DSL, BPMN, or general-purpose workflow product. Session directories, tickets, and runtime commands remain the 0.x execution model.

No Kubernetes, Nomad, or cloud autoscaler

v0.7 may add production worker and fleet control on hosts the operator already runs. It does not add a cluster scheduler or cloud autoscaler.

No multi-region control plane

One control identity per deployment. Multi-region failover is not authorized through v0.7.

No general secrets or object-storage product

Credentials stay in operator-owned env/files. Artifacts stay in session directories. This is not a secrets manager or an S3 competitor.

No multi-repository atomic execution

A governed execution targets one configured repository identity. Cross-repo atomic commits are out of scope.

No silent isolation or network downgrade

If a caller requested a sandbox minimum or network policy, the runtime fails closed rather than silently running weaker. Isolation and network claims must be ENFORCED. DockerWorker is a real digest-pinned backend; absence of an engine or digest is still a refusal, never a silent downgrade.


Sovereign Agent 1.x (educational line)

Authorized 2026-08-25. See the SOW metadata and migration guide.

1.x does include minimum local governance (outcomes, SOWs, rulings, authority, acceptance) in order to teach a complete organization.

The 2026-09-07 always-on teaching ruling authorizes one Telegram adapter, a bounded MCP client, local versioned skills, an explicit OS-supervised service, and an optional isolated container tool in this distribution. These are educational implementations; ZeoCore remains optional. The exclusions below apply outside that bounded teaching scope.

1.x still refuses:


How to change this list

  1. File authorization (1.x holdings are recorded in this repository under docs/rulings/; corpus filing follows when sow_repo is set).
  2. Record the reversal in CHANGELOG.md under the release that reverses it, and edit this file in the same change.
  3. Sandcastle remains refused through v0.7 without a further reversal. 1.x also does not add Sandcastle.
  4. Tag v0.7.0 is never moved.