# Auditable AI architecture > Markdown (GEO/AEO) version of . Author: **Mauricio Yokoyama Issei** · en-BR · Published: 2026-07-23 · Updated: 2026-10-10. ## In short Three systems — **institutional memory without hallucination**, **specification as contract** and **continuous resilience mat** — presented in two readings of the same truth: the executive, measure in risk and return, and the engineering, measure in coverage and latency. - **GraphRAG and ontologies** — the AI's response comes from the company's memory, with a path from provenance to node. - **Spec-Driven Development** — what has been approved is exactly what comes into production; the change goes back to spec, never straight to the code. - **DevOps Salesforce** — single and supervised delivery path, with known and exercised rollback. - **Determinism, auditability, traceability and reversibility** — four principles declared verifiable, not decorative. ## The three layers The page is organized in increasing depth: Surface (what the business earns, without untranslated jargon), Mechanism (as the pieces connect) and Core (the technical density, reserved for those who actively choose to go there). Density is always a visitor's choice, never an imposition. ## Two readings of the same truth A persistent selector alternates between the Executive and Engineering perspective. Phrase pairs maintain the same number of sentences, the same score and the same metrics, varying only the vocabulary — a restriction that makes the transition a transmutation rather than a content exchange. ## Frequently Asked Questions **What makes an AI architecture auditable?** The ability to reconstruct, for any output, the evidence that sustained it: what sources were consulted, under which version of the specification, with which approval. Auditability is not a report produced afterwards — it is the byproduct of a single decision and delivery path. **What is GraphRAG and why reduce hallucination?** It is the recovery of context from a graph of knowledge with typed relations, rather than just similarity of text. As each response carries the nodes covered, there is verifiable provenance; and since the absence of data is treated as explicit uncertainty (open-world), the system prefers not to respond to invent. **What is Spec-Driven Development in practice?** The specification is the reference of the intention and the basis of the tests: acceptance criteria have seen tests, and any behavior change passes first by spec. Spec is not infallible — when it, the tests and the code differ, someone with authority decides what to correct. A criterion is only worth failing — "improvement performance" is not criterion, "p95 below 400 ms" is. **What governance principles are declared?** Determinism (same entry, same output, with isolated and testable non-determinism), auditability (all decision holds re-executable evidence), traceability (from production behavior to the specification line that required it) and reversibility (no change enters without back-execised path). **Why does the page have two perspectives?** Because the same architecture needs to be defended in two vocabularies: risk, governance and return to executive decision; ontologies, contracts and latency for engineering decision. Phrase pairs maintain identical syntactic structure and metrics — it is the same truth told in two ways, not two different statements. ## Glossary - **Auditability** — Every automated decision keeps the evidence that sustained it, in a format re-executable by third parties. - **Traceability** — From the production behavior it is possible to reach the specification line that required it and the approval that released it. - **Reversibility** — No change enters without known and exercised way back. Untested rollback is not rollback. - **Provenity** — Record of nodes and sources traveled to produce a response, node by node. *© 2026 Mauritius Yokoyama Issei. Citable content with assignment (fair use educational).*