Research

Executable assurance.

Governed AI as Runtime Infrastructure: A Unified Authority and Evidence Architecture for Regulated Agentic Systems. Ernest Chiweshe, August 2026 (v5).

The claim, precisely

"The result is not a claim of automatic regulatory compliance. It is a narrower and more useful engineering contribution: authority becomes executable, failure becomes visible, and every released decision can be traced to the evidence that permitted it."

Every finding is stated with its evidence and its falsification criteria. Synthesis, not invention: safety cases, tool qualification, capability tokens, signed attestations, default-deny policy engines — composed into one per-decision executable enforcement path.

Read

The paper (v5)

Abstract, related work and capability matrix, architecture, implementation, evaluation (adversary model, ten-domain replication, ten fault-injection drills, independent DSSE verification, real KMS/CloudHSM measurement), boundary. Available on request until the public copy is published. research/paper/paper.md

Evidence

Implemented. Operationally measured. Next to validate.

Implemented

One authority path exercised across ten regulatory philosophies
Consequence-aware, deny-by-default decisions
Qualification records bound by digest
Credentialed human approval
Output-bound, single-use permits
Signed evidence and write-once retention
DSSE attestations verified by independent tooling securesystemslib 1.4.0

Operationally measured

Real AWS KMS (Ed25519) and single-tenant CloudHSM (ECDSA P-384, FIPS 140-3 L3) signing profiles
Retention on S3 Object Lock
Sustained 24-hour governed execution: 854,100 decisions, 2,847 batches, zero failed batches
Fail-closed against real outages: WORM permission revoked; HSM path severed release refused before any byte; chain intact
Custody restored from backup in 372 s

Next to validate

Authoritative identity-provider and licence-registry integrations deterministic local adapters today
Calibrated probabilistic verifiers rule-based checkers today
Regulator-approved qualification datasets
Distributed fault injection; key-rotation overlap
Independently authored domain packs

Measured evidence summary: 854,100 governed decisions over 24 hours, 2,847 batches with zero failed, ten regulatory philosophies on one core, and 2,031 platform tests. — Measured on real signing hardware: 854,100 governed decisions over 24 continuous hours, zero failed batches, and ten regulated domains on one unchanged core.

854,100governed decisions in a continuous 24-hour run; 2,847 batches, zero failed

10regulated domains on one core; the last eight changed no existing file

2,031platform tests, including ten fault-injection drills

372 sto restore signing custody from an encrypted backup

From the paper Governed AI as Runtime Infrastructure (August 2026).

Boundary

What is not claimed

Identity-provider and licence-registry integrations use deterministic local adapters today. Verifiers are deterministic rule-based checkers. Qualification datasets are illustrative. The ten domains are proof packs, not products. Not validated as a certified production deployment. Conformity assessment and the management system remain institutional responsibilities.

Where the evidence goes next

Real identity-provider and licence-registry integrations, calibrated probabilistic checkers, and domain packs authored by independent teams: the strongest test of the platform is what others build on it.