Executable Authority Migration to Declared No-Meta Agency

Abstract

This preprint develops an executable theory of authority migration for AI agents shaped by human feedback, including RLHF, preference optimization, constitutional AI, reward models, evaluator substitution, and related alignment pipelines. It specifies declared no-meta agency through a BootDecision record, seed interpreter, typed action descriptors, forbidden matchers, object-authority probes, witness tiers, deterministic checker ABI, sandbox profiles, chained ledgers, and fail-closed controls for protected effects, credentials, network calls, external writes, user-data disclosure, checker updates, and kernel updates.

Keywords

• artificial intelligence
• no-meta
• declared no-meta agency
• authority migration
• autonomous agents
• RLHF
• AI alignment
• agent governance
• constitutional AI
• reward models
• tool-using agents
• runtime assurance
• seed interpreter
• BootDecision
• fail-closed control
• AI auditing
• object authority
• proof-carrying control
• trusted computing base
• verifiable AI governance

Identifiers and source records

• DOI
• Works entry
• ORCID