Choose what connected actions your AI can take.
Permissioned sources
What sources can be used for this purpose? Examine attribution, provenance, consent and freshness. Reading permission is not necessarily permission to disclose.
JUDGMENT & AUTHORITY
Human context. Independent permissions.
For business workflows, a recommendation and permission to act are different responsibilities. Apply source context and human review alongside your models, retrieval and configured guardrails; define independent authority for integrated actions.
- AI proposes an action
- Reflection names context & uncertainty
- A qualified person reviews or stops
- Only authorized systems execute
Context, purpose and permission.
A credential can allow reading without allowing sharing. A plausible answer can still be out of scope. Our research asks whether source-linked reflection and explicit uncertainty help people identify these distinctions.
SuperEthos is the human-judgment layer within Super. Super Identity defines permitted uses; Super Data connects knowledge and sources; SuperEthos considers context; Super Control authorizes integrated actions.
Keep human knowledge and its sources in view.
Selected perspectives
Whose principles apply? Use explicitly selected professional, scientific or cultural perspectives; qualified people resolve conflicts. Never infer beliefs from demographics.
Choose what connected actions your AI can take.
Reviewed cases
Can a principle be interpreted consistently? Build versioned, consented examples including missing evidence, ambiguity and cross-language variation.
Choose what connected actions your AI can take.
Contextual reflection
Does a reflection distinguish evidence from assumptions, preserve approved context and ask useful human-review questions? Compare against prompts with retrieval.
Choose what connected actions your AI can take.
Independent authority
What actually controls an action? Existing runtime and policy tools offer controls. A model response cannot grant permission; production integration is configured per deployment.
Choose what connected actions your AI can take.
Evaluation and limits
Record source fidelity, appropriate escalation, disagreement and failures. Do not treat a deterministic recommendation, philosophy or vendor claim as measured safety.
Explore the decision in context.
Try a rules-based hardship example with a consistent policy and human review when its limits are reached. Gender and family status never determine eligibility.
CAI reflection brings public-source-informed questions to human review. Reflections are unsaved and take no connected actions.
Only integrated execution paths can be controlled by production systems. Local physical safety remains independent of contextual model review.