Research that can be inspected, challenged and reproduced.
Independent research on AI governance, traceability and autonomous systems — with methods, limitations and artifacts visible alongside the conclusions.
Current work
Independent technical reports and preprints. Publication status is shown explicitly; preprints are not peer reviewed unless stated otherwise.
Vertical AI Governance Stack v1.1 (Draft Specification)
Pre-registered v1.1 draft specification, 22-requirement matrix, dual conformance paths (Path A vs Path B), adversarial falsification, and frozen human inter-rater validation protocol (External Evaluation Pending).
Applied Audit 02: OpenAI Agents SDK
Experimental conformance audit of OpenAI Agents SDK runtime primitives, guardrail hooks, tracing spans, and handoff authorization under Vertical AI Governance Stack v1.0.
Applied Audit 01: OpenAI Swarm
Framework-native governance controls versus externally required controls under the Vertical AI Governance Stack v1.0. Baseline case study in minimal orchestrators.
Vertical AI Governance Stack v1.0 — First Applied Validation
Governance assessment of PrintPriceOS Control Plane exposing execution drift, asynchronous CVA bypasses, and framework candidate refinements.
Vertical AI Governance Stack v1.0
Canonical reference architecture for Semantic Sovereignty, Invariant Matrix (L1-L6), and Executable Governance controls.
Implementation Conformance Checklist v1.0
Operational audit tool for evaluating L1-L6 compliance, Governance Proof Obligations, and the Autonomy Ladder (A0-A4).
Vertical AI Manifesto: Semantic Sovereignty and Executable Governance
Models will be commodities. Operational knowledge will not. The foundational architectural thesis of Vertical AI.
The Oracle Blind Spot
Re-adjudicating automated telemetry assertions with transport-level evidence. Empirical study of verification-oracle defects.
Vertical AI Governance
Pre-registered research progression on semantic sovereignty, executable governance, and verifiable autonomous systems.
“Models may change. Operational knowledge must remain sovereign.”
People behind the work
LatinFlash Research publishes work with named authorship, explicit affiliations and versioned research records.
Evidence before narrative
We separate measurement from inference, preserve negative results, disclose methodological corrections and publish boundaries.
Traceability
Important claims should be traceable to data, protocol, literature or a clearly labelled inference.
Adversarial review
We actively search for alternative explanations, ceiling effects, false confidence and disconfirming evidence.
Reproducibility
Where licensing and privacy permit, scripts, ledgers, frozen outputs and reconstruction instructions accompany the work.