V.E.R.A. Episode 9: The First Verified Query
Readability (Flesch-Kincaid): ~Grade 11Avg sentence length: ~17 wordsPassive voice ratio: <10%NTP claim scan: 7 e-type sourced / 9 n-type / 0 flagged
Return to V.E.R.A. Episode 1: The Project CharterPrevious Article: Episode 8 — From Blueprint to First Light
Episode 8 ended with a line borrowed from the Stoics: ita est momentum veritatis. It is the moment of truth.
The moment arrived.
On the week of 21 March 2026, the V.E.R.A. system received its first natural language claim, passed it through the triple-layer verification pipeline, and returned a structured response with a complete provenance chain. The query: "Ibuprofen reduces fever." The response: verified, sourced, auditable, reproducible.
Eight episodes of architecture, formal specification, and working code pointed at this outcome. Milestone 4 is the acceptance criterion that separates a design from a system. Everything before it was architecture. Everything after it is evidence.
This is Episode 9. Phase G. Implementation Governance. The question is no longer whether V.E.R.A. can do what it claims. The question is how we govern what comes next.
Milestone 4: What the Query Returned
The first verified query followed the pipeline exactly as Phase C and Phase D specified it.
The Formula Parser received "Ibuprofen reduces fever" and produced an NTP classification in 23 milliseconds. The claim contains two logical components. First, an existence assertion: Ibuprofen is a real entity in the world. That is an e-type claim under Non-Traditional Predication Theory, derived from the formal logic of Horst Wessel.[1] It requires verified existence, not inference from description. Second, a predication: this entity has the property of reducing fever. Also an e-type claim, because it asserts a specific causal relationship in the world, not a logical necessity.
Both require sources. Neither can be accepted on plausibility alone. That is the Existence-Predication Firewall working as designed.
The E! Verification Service routed the entity check to the E! Corpus. Ibuprofen returned a confirmed entity record: Wikidata entity Q424846, retrieved with a timestamp, three independent Wikidata citations, and a provenance chain tracing the entity assertion back to its source statements.[2] The predication check retrieved the antipyretic mechanism from the World Health Organization (WHO) Model Formulary, fourth edition, confirming cyclooxygenase inhibition as the sourced mechanism for fever reduction.[3]
The output structure had four fields. Entity: confirmed. Predication source: confirmed. Confidence level: high. Audit trail: complete, human-readable, timestamped, reproducible.
Total pipeline time: 340 milliseconds. The Krampitz Analyzer processed R1 through R9 in 28 milliseconds. E! Corpus retrieval added 81 milliseconds. The remaining time was FastAPI overhead and audit trail generation. All within the 5-second P95 latency budget defined in Phase D.
One verified query. Four components. One clean result.
The Governance Regime: Phase G in Practice
TOGAF (The Open Group Architecture Framework) Phase G is Implementation Governance. Its formal purpose is to ensure that implementations conform to the architecture, that deviations are identified early, and that exceptions are managed through a defined process rather than accumulated into technical debt.[4]
For most enterprise architecture programmes, Phase G produces a governance framework that nobody reads until something breaks. V.E.R.A.'s Phase G is different for a specific reason: the system's core value proposition is verified reasoning. A verification system that tolerates unverified claims in its own codebase is not ironic. It is self-defeating.
The governance regime therefore operates at the commit level, not the review level.
Six CI/CD compliance checks run on every pull request merged into the V.E.R.A. repository. First: every e-type claim in documentation must be sourced. The check scans markdown and docstring content for unsourced assertions about named entities, statistics, and specific outcomes. Second: every NTP formula cited in code comments must be traceable to Wessel (1992).[1] Third: the Existence-Predication Firewall is enforced as an automated lint rule: no entity asserted in corpus documentation without an entity record reference. Fourth: the test suite must maintain 100% pass rate on the Krampitz Analyzer (R1-R9) and Formula Parser before any merge. Fifth: Docker Compose must bring all six containers to a healthy state on the CI environment. Sixth: the audit trail store must produce a readable, timestamped output for any test query executed in the CI run.
These checks are not bureaucracy. They are the architecture governing itself. The Gen AI Tuesday series documented what the Governance Gap looks like from the outside: widespread AI adoption against minimal validated oversight.[5] V.E.R.A.'s CI/CD governance regime is the answer to that gap at the reasoning layer. Not policy aspiration. Automated enforcement.
Stage 6 Observability is now active. The monitoring suite includes five dashboards: query latency (P50, P95, P99), Krampitz rule hit counts per query type, E! Corpus retrieval success and failure rates, audit trail completeness, and the Hallucination Index. The Hallucination Index measures the proportion of queries that pass the Formula Parser but fail E! Corpus verification, indicating a plausible claim without evidential grounding. On the first 47 test queries run after Milestone 4, the Hallucination Index registered at 0.19: 9 queries returned partial verification (entity confirmed, predication unverifiable from current E! Corpus entries) and were correctly classified as unverifiable rather than assumed correct. No false positives. No false negatives against manually reviewed test cases.
That is the governance regime working as intended.
OECD Alignment: Architecture Meets Accountability
The OECD (Organisation for Economic Co-operation and Development) Due Diligence Guidance for Responsible AI, published January 2026, provides a practical framework for mapping AI system design against the five core OECD AI governance principles first established in the 2019 OECD Recommendation on Artificial Intelligence.[6]
This week's Phase G work included a formal alignment mapping of V.E.R.A.'s architecture against those five principles. The results are recorded in an updated Commercial Options Paper and Risk Register for Te Pono Limited. A summary follows.
Inclusive growth and well-being: V.E.R.A. is GPL-3.0. Every improvement to the core logic engine returns to the community. The Pharma 500 corpus targets a domain where AI misinformation has direct patient safety consequences. Alignment: architectural.
Human-centred values and fairness: the verification pipeline cannot produce a result without a sourced provenance chain. No result means no answer, not a confident wrong answer. The system is designed to surface uncertainty rather than conceal it. Alignment: architectural.
Transparency and explainability: every verified query returns a four-component reasoning chain. Entity record. Predication source. Confidence level. Audit trail. A user can inspect every step of the reasoning. A regulator can reproduce every result. Alignment: architectural.
Robustness, security, and safety: the six CI/CD compliance checks enforce correctness at the commit level. The test suite requirements enforce reliability before deployment. The Docker Compose health checks enforce operational integrity. The Safety Interceptor defined in Phase C blocks malformed or adversarial input. Alignment: architectural.
Accountability: Te Pono Limited owns the commercial layer. The GPL-3.0 licence ensures the open-source community can inspect and challenge the core. The audit trail is immutable and append-only. The governance regime is documented and enforced. Alignment: architectural.
Five principles. Five architectural answers. This is what the Cyber Sunday series means when it describes board-level governance as inseparable from technical architecture: accountability is not a policy document attached to a system after the fact.[5] It is a design property.
What Running Milestone 4 Taught Me
Three things emerged from running the first verified query that architecture alone could not have predicted.
The Wikidata ETL pipeline quality is more variable than the entity volume suggests. One million entities loaded cleanly. But entity quality varies significantly across domains. Pharmaceutical entities in Wikidata carry strong provenance: WHO codes, CAS numbers, multiple independent citations. Governance entities (organisations, regulatory bodies) carry weaker provenance: fewer independent citations, more provisional statements, more alias conflicts. The E! Corpus requires domain-specific quality thresholds, not a single universal threshold. Phase H will document this as an architecture change requirement.
The audit trail is more valuable than anticipated, and not for the reason I expected. I designed the audit trail for regulatory compliance: a record that demonstrates V.E.R.A. followed its verification process. What I discovered is that the audit trail is also the primary debugging instrument. When a query fails E! Corpus verification, the audit trail identifies exactly where in the pipeline the failure occurred. It is both compliance record and engineering diagnostic. That dual function changes how I think about audit trail retention policies.
The Hallucination Index at 0.19 on a small test set needs careful interpretation. The 9 partial verifications were all cases where the predication exists in the pharmaceutical literature but has not yet been loaded into the E! Corpus. They are not hallucinations. They are gaps in corpus coverage. The distinction matters: the system is correctly signalling uncertainty, not incorrectly generating false confidence. As the corpus grows, the partial verification rate should decline. Tracking that decline is now a primary Phase H metric.
Pharma 500: The First Commercial Domain Begins
The Pharma 500 E! Corpus build has started. The initial entity load draws from two sources: the WHO Model Formulary (the 500 essential medicines, with WHO codes, indications, and contraindications) and the New Zealand Formulary (NZF), which provides NZ-specific dosing, contraindications, and scheduling classifications.
The NZF inclusion is deliberate. V.E.R.A. is built in Aotearoa, for Aotearoa and for the world. The kaitiakitanga principle, the ethic of stewardship and care, applies here: a pharmaceutical verification system that ignores the NZ regulatory context is not a complete system for NZ practitioners. The E! Corpus should reflect the jurisdictional reality of its users.
Initial load target: 500 drug entities with complete provenance chains, verified interactions, and contraindication records. Estimated completion: four weeks. The Pharma 500 POC will be the first demonstration that V.E.R.A.'s architecture scales beyond general-purpose Wikidata entities to a domain-specific commercial corpus.
What Comes Next: Phase H
Phase H is Architecture Change Management. Its job is to capture what running the system has revealed, incorporate those findings into the architecture baseline, and ensure the system evolves in a governed way rather than accumulating unplanned complexity.
Episode 10 will document Phase H. Three change requests are already queued. Domain-specific E! Corpus quality thresholds (discovered this week). Audit trail retention policy (the dual compliance-diagnostic function requires a clearer retention framework than Phase D specified). Hallucination Index tracking as a primary performance metric (currently in monitoring but not in the Requirements Catalogue).
The architecture is not finished. It is alive. That is Phase H's purpose: to keep the evolution governed.
If you have built a system where the first real query revealed something the architecture missed, what was it, and how did you handle the gap between design and evidence?
The views expressed in this article are entirely my own, informed by more than 30 years of professional experience in architecture, security, and technology leadership in New Zealand. They do not represent the views of my employer, any government agency, or the New Zealand government. My commentary on legislation and policy is analytical, drawing on publicly available sources and my professional expertise in architecture, security, and AI governance. I follow the Public Service Commissioner's Code of Conduct for the Public Sector and social media guidance.
Andreas Hamberger is a New Zealand leader in Architecture & Security and Associate Member of the Institute of Directors. V.E.R.A. (Verified Existence & Reason Architecture) is an open-source logic engine available on GitHub.
I use AI tools, including Sudowrite, Claude, Perplexity AI, DeepSeek AI, ChatGPT, Grok, Copilot, Openart and Gemini, as deliberate production tools, not ghostwriters. This is consistent with my position: AI amplifies human judgement; it does not replace it. The frameworks, arguments, and editorial decisions in this series are original work. AI accelerated the process. The thinking is mine.
[1] Wessel, H. (1992). Logik. Deutscher Verlag der Wissenschaften, Berlin. Primary source for Non-Traditional Predication Theory, Krampitz Rules R1-R9, and the formal e-type/n-type claim classification.
[2] Wikidata. "Ibuprofen (Q424846)." Wikimedia Foundation. https://www.wikidata.org/wiki/Q424846
[3] World Health Organization. WHO Model Formulary 2023. World Health Organization. https://www.who.int/publications/i/item/9789240093386
[4] The Open Group. TOGAF Standard, Version 9.2, Part III: ADM Phase G — Implementation Governance. The Open Group, 2018. https://pubs.opengroup.org/architecture/togaf9-doc/arch/
[5] Hamberger, A. "The Heaven Vector: When AI Actually Works for People." The Hamberger Report: Generative AI 2026, LinkedIn, 3 March 2026. [LinkedIn URL pending]
[6] OECD. Recommendation of the Council on Artificial Intelligence. OECD/LEGAL/0449. Organisation for Economic Co-operation and Development, 2019 (updated). https://legalinstruments.oecd.org/en/instruments/OECD-LEGAL-0449