The Sovereign Capability Envelope (SCE): Law-First Permissioning, Conformal Risk Control, and Auditable Governance
Year: 2025
Type: article
Abstract: We present the Sovereign Capability Envelope (SCE), an inference-time governance layer that first decides whether a candidate action may be executed at all (permissioning), and only then selects the most useful option within the permitted set. SCE combines (i) a ranked constitution Φ, (ii) vector risk budgets b, (iii) calibrated multi-head risk with conformal upper-bounds, (iv) a lexicographic principle-law first, utility inside the law, (v) a fallback ladder when budgets are exceeded, and (vi) a Decision Trace with numeric, span-level, and counterfactual reasons (L2–L5). We add a legislative feedback loop (LEL) and Impact Analysis Report (IAR) separating regulatory budget tuning from constitutional rule amendments with canary/rollback/sunset. Experiments show improved safety–utility Pareto and bounded latency via a cascaded design: rules → multi-head risk → calibration → conformal lookup → permission → utility. ContextThis work is part of the QWRZ research line on Governed Behavioral Intelligence, succeeding the paper GBI: Governed Behavioral Intelligence (DOI: 10.5281/zenodo.17267857). It defines the inference-time governance layer within the broader architecture of safe, interpretable, and sovereign behavioral AI.
Author: Hussien, Jaffer
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7102408647
FWCI:
Cites:
Cited by:
Related to: