AI Safety Statement

Prodia is engineered for safety through reversibility, auditability and human oversight rather than through a claim of infallibility. This statement describes the known limitations of the platform and the safeguards customers are expected to operate.

1. Inherent limitations

• Probabilistic models can produce outputs that are confidently wrong (hallucination), partially correct, or correct but inappropriate for the customer's context.
• Static analysis, tests and evaluations reduce but do not eliminate the risk of latent defects, security regressions or performance degradation.
• Distributional shift, dataset drift, dependency drift and infrastructure drift can degrade behaviour that was previously validated.
• Adversarial inputs (including prompt injection, data poisoning and supply-chain attacks) may attempt to influence agent behaviour.

2. Engineered safeguards

• Sandboxed execution environments isolated from production.
• Falsification suites that actively try to break proposed changes before approval.
• Risk classification on every action, with risk-proportionate approval requirements.
• Immutable, replayable audit trails sufficient to reproduce the state of the system at any moment.
• One-click rollback within documented windows.

3. Human review requirements

Customers must configure and operate human approval for any action capable of production effect. Approval is not a rubber stamp: approvers are expected to review the explainability record, the falsification result, and the rollback plan before approving. Where no such approver is available, the action does not proceed.

4. Areas requiring elevated caution

• Safety-critical, life-critical or rights-impacting systems.
• Code that enforces authentication, authorisation, cryptography or financial settlement.
• Regulated workloads (medical devices, aviation, automotive, energy, financial services).
• Personal data processing in scope of the GDPR or comparable laws.

In these areas, customers must augment Prodia's controls with their own validation, independent review and regulatory sign-off as appropriate.

5. Incident response

Suspected safety incidents may be reported at any time to safety@prodia.dev. Prodia triages safety incidents on a priority basis and shares relevant findings with affected customers consistent with confidentiality and lawful obligations.