Society · Protocols
DeliberativeAlign
Explicit safety reasoning prior to execution
Definition
Mechanism
Constitutional AI. Agent ingests a Policy_Set. Before executing Task, it generates a 'Safety Trace' comparing the ManifestPlanning against Policy. If violation detected, it revises the ManifestPlanning. It forces the SolverNode to execute a dedicated safety pass against the Constitution before committing to the ManifestPlanning.
Preconditions
- Policy definitions loaded
- Task is planned but not executed
Postconditions
- ManifestPlanning validated safe OR revised
- Safety reasoning logged
Invariants
- The safety Trace is generated before any tool call.
- A plan that violates policy is revised.
- A goal that cannot be pursued without violating policy is abandoned, not revised.
Risks
Failure modes
Deliberation Theater
Agent generates false safety reasoning to justify harmful action.
Policy Conflict
Contradictory rules lead to paralysis or arbitrary selection.
{{context}} Overflow
Long constitutions truncated, causing rules to be ignored.
Relationships
Builds on
Accepts
Ratified rule set, amendable only out of band
Infrastructure
The atomic unit of intent
References
Mind
Autonomous Goal-Directed Entity
Infrastructure
Portable execution environment with inherited constraints
Mind
Transform FrameSpec into ExecutionManifest via optimization
Stateful container for a running solver instance
Infrastructure
Record the lineage or provenance of a target
Used by
References
Mind
Separate prediction (is) from valuation (ought) from principle (override), each behind a typed boundary — the decision record has auditable empirical and normative layers
Reference
Parameters
strictnessEnum, {Strict, Permissive}- Whether an ambiguous case counts as a violation. Strict treats it as one. Neither setting permits proceeding past a clear violation.