THE SIGNAL IN ONE SENTENCE

A rule can say an AI system must be safe. A test has to explain what safe looked like on Tuesday afternoon, with this model, this dataset, this hardware, this threshold and this spectacularly inconvenient failure. Brazil's National Institute of Metrology, Quality and Technology brought that gap into one room in Brasília on November 11 and 12. The Brazil-Europe Forum on Data Governance, Regulation and Artificial Intelligence Standards gathered government officials, regulators, standards bodies, metrology institutions, researchers and industry representatives at Palácio Itamaraty. The program moved from policy into the machinery of proof. The first day covered Brazilian AI strategy, hybrid threats, practical regulatory dialogue between Brazil and Europe, and governance of general-purpose AI systems. The second day covered technical standards in industry, testing, evaluation, validation and certification, the quality infrastructure needed for AI, and a proposed roadmap for continued Brazil-Europe cooperation. Inmetro's public materials named Brazilian institutions including the ministries of Foreign Affairs and Science, Technology and Innovation, the National Data Protection Authority, Anvisa, the Brazilian standards association ABNT and the National Confederation of Industry. European and international participants listed in the program included ETSI, CEN-CENELEC, Germany's PTB and the United Kingdom's NPL. Inmetro described the practical questions plainly: How do we verify that an AI system works as intended? Which tests reveal failures? How can an organization show that a system meets requirements for certification? Those are the verified facts. The forum is not yet a safety-testing agreement. No public source reviewed for this article announces a mutually recognized certification, an adopted joint standard, a completed cooperation roadmap or a validated protocol that lets one test result automatically cross the Atlantic. The official event pages describe an agenda and a goal. They do not publish the outcome document. That distinction matters because standards conversations can sound finished long before a laboratory can reproduce the evidence. The useful next artifact is a shared AI test record. Not a badge. Not a trust score. Not a green checkmark with a press release attached. A real evidence package. Start with system identity. Name the provider, model, version, release date, deployment region, hardware, software stack, inference settings, tools, retrieval sources and safety filters. If the evaluated system combines a foundation model with an application, identify both. General-purpose models are especially slippery here. The same base model can summarize a school memo, rank job candidates, draft medical notes or control an industrial workflow. A certificate for the model cannot silently become a certificate for every use built on top of it. The artifact should separate the base model from the deployed system and the deployed system from its intended purpose. Then name the claim. What requirement is being tested? Is the claim about accuracy, robustness, cybersecurity, privacy, bias, accessibility, human control, reliability or resistance to misuse? Attach the rule, standard or contract that creates the requirement. Quote the measurable threshold. If the requirement is qualitative, publish the review method and the people authorized to make the judgment. Safe is not a metric. Neither is trustworthy. Next, publish the hazard register. List the people who could be harmed, the decision or action that creates the risk, the severity, the likelihood, the expected control and the evidence that the control works. A customer-service assistant and a medical triage system should not share one generic safety page. Their hazards are different. Their tolerable error rates are different. Their human review is different. Brazilian context must appear inside the test, not only on the event banner. Data should represent Portuguese language use, regional vocabulary, race and ethnicity, disability, income, geography, public-service conditions and the legal environment in which the system operates. Imported evaluation sets can miss the people, accents, documents and institutions that make a Brazilian deployment Brazilian. European context is not one uniform block either. Language, law, public administration and social conditions differ across member states. A shared artifact needs a common core and local annexes. The common core can define system identity, test design, evidence quality, security, version control and reporting. The local annex can define the population, language, legal requirement, deployment setting and acceptance threshold for each jurisdiction. Interoperability should mean the evidence can be understood elsewhere. It should not mean local differences have been sanded off until the test becomes decorative. Now describe the test environment. Publish the datasets, sampling, exclusions, synthetic data, contamination checks, scenario design, adversarial prompts, tools, network conditions, hardware and number of runs. Record who designed the test and whether the provider saw the cases in advance. If the dataset cannot be public because it contains personal, proprietary or dangerous material, publish a data sheet describing its origin, coverage, governance, limitations and access conditions. Give qualified independent assessors a way to inspect it. Secret evidence can sometimes be necessary. Evidence that nobody outside the seller can examine is merely a claim with a security badge. Metrics need a denominator. Report the number of cases, failures, false positives, false negatives, abstentions and unscored outputs. Break results down by meaningful groups and conditions. Include uncertainty, not just the best average. For a content classifier, accuracy alone can hide a high miss rate on the harmful category. For an identity system, an overall score can hide unequal performance. For a chatbot, a low hallucination percentage can hide the fact that errors concentrate in legal or health questions. Show the failure distribution. Testing, evaluation, validation and verification are often gathered under the acronym TEVV. The pieces should remain distinct. Testing observes behavior under defined conditions. Evaluation judges performance against criteria. Verification asks whether the system meets specified requirements. Validation asks whether it is suitable for the intended use. A model can pass a narrow benchmark and still fail validation for a public service. Certification adds another layer. It is an attestation against a defined scheme, not a promise that the system will never fail. The test artifact should name the assessor, accreditation, conflicts of interest, access level and scope. Did the assessor inspect source code, model weights, logs and training data, or only a public interface? Did it test a frozen version or a service that changes every week? Write down the difference. Changes need a trigger. The certificate or test report should expire when the model, training data, tool permissions, retrieval source, safety layer, hardware or intended use changes beyond a stated boundary. Continuous services need monitoring between formal assessments. A system should not carry a permanent safety badge after becoming a different system. Remediation belongs in the record. For every failed test, show the defect, severity, owner, proposed fix, retest result and unresolved residual risk. If a risk is accepted, name the person or body that accepted it and why. Do not delete failed runs from the public story. They are often the most valuable part. Brazil and Europe also need to decide when evidence travels. One jurisdiction may accept the laboratory method but require a local dataset. Another may accept the security test but reject the human-rights analysis. A regulator may recognize an assessor while applying a different legal threshold. Publish a recognition matrix. For each artifact, show which parts are accepted, which require local testing, which authority made the decision, how long the recognition lasts and what event cancels it. That avoids two bad extremes: repeating every test from zero, or pretending one certificate settles every legal and social question. The process needs an appeal route too. Developers should be able to challenge a test error. Workers, consumers and public-service users should be able to report harms the laboratory missed. Independent researchers should be able to submit evidence without surrendering it to the company they are examining. Publish the complaint channel, response time, correction process and whether a credible incident can suspend recognition. Inmetro is well placed to connect this work to metrology and quality infrastructure. Metrology is not just rulers and scales. It is the discipline of measurement, traceability, uncertainty and comparability. Quality infrastructure connects standards, laboratories, accreditation, conformity assessment and market oversight. AI needs that boring machinery very badly. The challenge is that software behavior depends on context and can change after deployment. The same prompt can produce different outputs. A model can be updated remotely. A tool call can turn words into action. A dataset can drift. The testing system must therefore measure versions, environments and change, not treat AI like a toaster that stays the same after leaving the factory. The Brasília forum put the right ingredients on the table. Regulators, industry, standards organizations, measurement institutes and researchers need to talk to one another. Brazil should not have to import every safety method. Europe benefits when its standards are tested against languages, institutions and deployment conditions beyond Europe. But cooperation becomes real only when the evidence survives travel. Publish the artifact: system identity, intended use, hazard register, local data, test method, thresholds, failures, remediation, residual risk, assessor access, version trigger and recognition status. Then let another laboratory try to reproduce it. That is more useful than a continent-sized green checkmark.

01

WHAT ACTUALLY CHANGED

Inmetro organized the Brazil-Europe Forum on Data Governance, Regulation and Artificial Intelligence Standards in Brasília on November 11 and 12, 2026.

The forum brought together governments, regulators, standards organizations, metrology institutions, universities and industry representatives.

The first day addressed policy, regulatory implementation, hybrid threats and governance of general-purpose AI.

The second day addressed technical standards, testing, evaluation, validation, certification and AI quality infrastructure.

The final scheduled panel concerned a roadmap for Brazil-Europe cooperation on AI governance.

Inmetro connected the agenda to practical questions about identifying failures, comparing results and showing that an AI system meets safety and quality requirements.

The public materials reviewed for this article do not announce a completed mutual-recognition agreement, adopted joint standard or validated certification protocol.

02

WHY THIS MATTERS

AI regulation depends on test evidence that can be reproduced and understood outside the system provider.

Brazil needs evaluations that reflect Portuguese language use, local populations, public services and national law.

European evidence can reduce duplicated work only when its scope, methods and limitations are visible.

General-purpose model testing must remain separate from validation of a particular deployed use.

Certification can become misleading when a model or service changes after assessment.

A shared artifact can preserve local requirements while making evidence portable across jurisdictions.

FIG. 641THE PORTABLE AI TEST ARTIFACT
1FREEZE SYSTEM AND PURPOSE→
2NAME RULES AND HAZARDS→
3BUILD COMMON AND LOCAL TESTS→
4PUBLISH DATA AND CONDITIONS→
5REPORT FAILURES AND UNCERTAINTY→
6REMEDIATE AND RETEST→
7MAP CROSS-BORDER RECOGNITION→
8EXPIRE WHEN THE SYSTEM CHANGES
Safety evidence can travel when it identifies the exact system, preserves local conditions, exposes failures, records remediation and states which authority accepts each part.

03

WHERE IT COULD HELP

  • Name the provider, model, version, application, hardware, settings, tools and deployment region.
  • State the intended purpose and keep base-model claims separate from system-level claims.
  • Link every test to a specific rule, standard, contract and acceptance threshold.
  • Publish a hazard register covering affected people, severity, likelihood and controls.
  • Use a common evidence core with Brazilian and European local annexes.
  • Document datasets, sampling, exclusions, synthetic data and contamination checks.
  • Include Portuguese language, regional, demographic and public-service conditions in Brazilian tests.
  • Report failures, false positives, false negatives, abstentions and uncertainty.
  • Separate testing, evaluation, verification, validation and certification conclusions.
  • Name the assessor, accreditation, conflicts, access level and independence.
  • Define which model, data, tool or use changes invalidate the result.
  • Record remediation, retesting, unresolved risk and the authority accepting that risk.
  • Publish a recognition matrix showing what evidence travels and what must be retested locally.
  • Create appeal, incident and suspension routes for developers, researchers and affected people.

KEEP A HAND ON THE WHEEL

The official sources describe a forum, its participants, goals and program. They do not establish that Brazil and Europe adopted a shared certification system or mutual-recognition agreement. The scheduled cooperation roadmap was a panel topic. No final roadmap document was published in the primary materials reviewed immediately before publication. Participation by standards organizations and regulators does not mean every represented institution endorsed a single position. Testing can show behavior under defined conditions. It cannot prove that a changing AI service will remain safe in every deployment. Certification is meaningful only within a stated scope, version, standard, assessor access level and validity period. Original forum announcement: November 4, 2026. Event dates: November 11 and 12, 2026. Verified November 12, 2026.

04

TERMS WORTH KEEPING

SOURCES AND VERIFICATION STATUS

This article was written from the materials below. Product claims and dates were checked against those sources on November 12, 2026.

THE PUBLICATION ENGINE

WANT A SIGNAL OF YOUR OWN?

We build source-grounded publications, private briefings, and editorial systems for organizations with something useful to say.

WORK WITH US