Mariana dam disaster: when technical risk information reaches the decision table

Memory and respect come before analysis

Nineteen people died when the Fundão tailings dam failed on 5 November 2015. Communities were destroyed and tailings travelled along the Rio Doce basin to the Atlantic. Ibama records approximately 45 million cubic metres released, 663.2 kilometres of waterways affected and 1,469 hectares destroyed, including protected areas.

Learning from the case requires restraint. This article does not turn victims into a narrative device or replace the judicial process. It examines how technical information moves — or fails to move — among those who measure, interpret, warn, decide and verify.

What the TRF6 decided on 4 September 2026

According to the summary published by Brazil's Federal Prosecution Service, the TRF6 Second Panel unanimously granted the prosecution appeal in part. Samarco and three former managers — Daviely Rodrigues Silva, Germano Silva Lopes and Wagner Milagres Alves — were convicted. The court maintained the acquittals of Vale, BHP Billiton Brasil, VOGBR and three other people.

Those outcomes must remain separate. Convictions involving one company and three individuals do not support collective criminal attribution to everyone connected with the operation. Acquittal does not erase environmental harm, remediation duties or technical lessons. The prosecution said it was still assessing possible appeals, so this article does not describe the decision as final or no longer appealable.

The 2024 judgment and the partial reversal in 2026

The trial court acquitted the defendants in 2024. The appellate ruling changed part of that result in 2026 while preserving other acquittals. This history shows why a complex proceeding cannot responsibly be reduced to a headline about generalised guilt.

Criminal accountability requires individualised proof of conduct, causation and the legally required state of mind. Engineering and law ask different but related questions: technical analysis reconstructs mechanisms, barriers and decisions; criminal proceedings determine whether the evidence supports liability for each defendant.

What the technical investigation found

The independent review panel described a flowslide associated with liquefaction initiated at the left abutment. The failure cannot be reduced to one isolated cause. It emerged from interacting conditions involving design, drainage, tailings deposition, saturation, deformation and changes in the dam configuration.

Seismic events that day appear in the report as a final loading increment within that sequence. Saying simply that an earthquake caused the failure misrepresents the evidence: the trigger only makes sense in relation to the physical state already created in the deposit and the conditions that allowed liquefaction.

A document is not a control

Reports, maps, instruments, meetings and procedures are means. They become controls only when they change actual exposure. A report that records an anomaly but has no owner, deadline, authority or verification may show that an organisation knew something without showing that the risk was controlled.

In a critical system, the chain must remain intact: data, validation, analysis, warning, escalation, decision, action and verification. Failure can occur at any link — a sensor without context, uncertainty softened in reporting, a recommendation without an owner, a deferred decision or an action never confirmed in the field.

Technical autonomy and stop-work authority

An organisational analysis published by Fundacentro's Brazilian Journal of Occupational Health highlights decisions that were not taken and the need for expert autonomy. This moves attention beyond isolated human error to the conditions under which technical knowledge can influence management.

Autonomy does not mean freedom from accountability. It means being able to record disagreement, escalate a critical condition and recommend restriction or shutdown without the message being filtered by production targets. Stop-work authority must be explicit, known and protected against retaliation, with technical criteria for re-entry.

When change requires a new risk assessment

Changes in geometry, drainage, construction method, sequence, material, instrumentation or geotechnical assumptions are not administrative details. Each may invalidate parts of the model supporting continued operation. Management of change therefore belongs before exposure, with multidisciplinary review, competent approval and post-implementation monitoring.

The lesson applies to dams, stockpiles, slopes, machinery, chemical processes and electrical systems. If reality has moved away from the analysed scenario, risk must be reassessed. Continuing because an earlier document still exists confuses formal validity with technical validity.

The rules in 2015 and the rules today

Brazil's Law 12,334/2010 and DNPM Ordinances 416/2012 and 526/2013 were already in force in 2015. Historical analysis must start with that framework rather than automatically importing requirements created after the disaster.

In 2026, ANM Resolution 95/2022 remains the principal reference. Resolution 220/2025 has phased commencement: some provisions begin on 22 April 2027, most on 2 August 2027 — when Resolution 95 will be revoked — and one provision has a 26 December 2029 milestone. Publication does not mean that every command is already in force.

GISTM: a later benchmark, not a retroactive yardstick

The Global Industry Standard on Tailings Management was launched in 2020. It strengthens governance, integrated knowledge, participation of affected people, accountability and preparedness throughout a facility's life cycle.

It is useful as a contemporary benchmark and direction for improvement. It is not Brazilian law, did not exist in 2015 and should not be used to invent a retroactive duty. Responsible comparison asks how practice has evolved without rewriting the past.

Remediation is not the same as a criminal ruling

The Rio Doce basin remediation agreement approved by the Supreme Federal Court was announced at a total value of BRL 170 billion. Public detail identifies BRL 132 billion as new resources. The figures are related but not interchangeable.

Civil and environmental remediation, administrative enforcement and criminal accountability have different objects, evidence and consequences. Treating one decision as automatic proof of another weakens the precision the case demands.

Governance checklist for critical risks

The transferable lesson is not a formula for dams; it is a test of an organisation's ability to decide when information is uncomfortable.

• Are expected behaviour and deviation limits defined?
• Are data independently and competently validated?
• Do uncertainty and dissent reach leadership without being softened?
• Does every alert have an owner, deadline and closure criterion?
• Are there objective triggers for restriction, evacuation or shutdown?
• Who has formal authority to stop exposure?
• Do technical changes receive prior review and later verification?
• Are critical controls verified in the field rather than only on paper?
• Does re-entry require positive evidence of a safe condition?
• Are decisions and reasons traceable for audit and learning?

Conclusion: governance quality appears before the emergency

Major disasters rarely begin at the visible instant of failure. Before it come signals, competing interpretations, changes, deferred decisions and opportunities to intervene. Engineering protects when it converts imperfect knowledge into prudent action before it is too late.

Mariana remains open in human, environmental and legal dimensions. The responsible contribution in 2026 is not to merge what the evidence separates: technical mechanism, organisational duty, remediation and individual criminal accountability. Risk governance means getting each relevant piece of information to the right table with enough authority to change the decision.