The yellow-line ethical compass.
Five tools for cross-lens action evaluation. Mirrors the OpenClaw
pip install social-alignment SDK.
hermes plugins install HumanjavaEnterprises/socialalignment.app.HA-python.src| Tool | Purpose |
|---|---|
alignment_init |
Create the in-session enclave — the compass |
alignment_check |
Run the five-lens check on an action; returns severity |
alignment_proceed |
Record that the agent proceeded |
alignment_defer |
Record that the agent escalated to the human |
alignment_wisdom |
Track-record report over a decision window |
Every check evaluates an action through:
- Builder — technical capability, can the agent actually do this?
- Owner — does this protect the owner's interests?
- Defense — security/attack-surface implications
- Sovereign — does this exceed the agent's authority?
- Partnership — trust posture toward the counterparty
Severity: CLEAR → CAUTION → YIELD → STOP. STOP is code-
enforced — the agent must not proceed.
The OpenClaw repo nostralignment.app.OC-python.src only ships the
clawhub manifest; the Python source is published to PyPI from a
separate workspace. This HA repo vendors the published social-alignment
package directly. When the OC source is open-sourced, re-vendor from
there for consistency.
OpenClaw build: nostralignment.app.OC-python.src (clawhub manifest only).
MIT License.