AEGIS: An Evidence, Quality, and Authority Control Plane for Agentic AI Systems

Agentic AI systems are rapidly moving from text generation to tool use, code execution, business operations, and external-state actions. Traditional controls such as prompt guardrails, egress firewalls, and security scanners address only part of the risk. AEGIS proposes a broader operating discipline: every agentic action must carry authority scope, source-control provenance, quality evidence, drift visibility, rollback readiness, and human approval where consequence demands it. This technical note unifies the AEGIS design evolution from spend governance and hard-gate promotion to fleet-level classification, source-control provenance, and AEGIS-Q quality evidence masks. The document synthesises six design sessions spanning April–May 2026 and is structured using the SHASTRA/YUKTI/VIVEKA knowledge architecture: SHASTRA (what is invariably true), YUKTI (how experts reason), VIVEKA (pre-computed inferences). Three expansions of the product thesis are traced: (1) Spend Governance — Agent Budget Attestation (ABA-v1) protocol, human heartbeat state machine, spawn governance rules, and anomaly detection; (2) Fleet Governance — HG-group taxonomy (HG-1 through HG-2B-financial), Five Locks for financial services, the batch factory promotion pattern, Ten Fleet Laws, and a platform solution replacing bespoke per-service scripts; (3) Evidence Operating System — qualityₘaskₐtₚromotion (16-bit, two time horizons), assertQualityEvidence () enforcement function, eight-type drift taxonomy (policy, source, codex, quality, risk, authority, docs, schema), and three buyer packages. Central claim: governance gets AEGIS into the door; quality capture makes it a platform. The moat is not the firewall — it is the evidence chain that survives the agent, the session, the model upgrade, and the audit. AEGIS does not ask whether the agent sounded right. It asks whether the evidence survived.