Responsible AI Policy

Principles

1. Evaluate before you deploy. No agent ships to production without a scorecard and a rollback path.
2. Governance at runtime, not just in policy. The platform enforces what the policy document promises.
3. Provenance by default. Every change is signed, versioned, and replayable.
4. Human in the loop where it matters. High-impact actions require approval. We publish the default thresholds.
5. Honest limitations. We state readiness level per domain. We state when a score is a proxy rather than ground truth.

Prohibited uses

• Autonomous clinical diagnosis or treatment without a regulated partner.
• Credit, insurance, or employment decisions that disadvantage protected groups.
• Generation of content that violates applicable law, including CSAM and illegal weapons design.
• Social scoring systems as defined by the EU AI Act.
• Disinformation, impersonation, or deceptive persuasion at scale.

Model providers

We are model-agnostic. Providers are selected per case for capability, safety, and cost. Prompts to third-party model APIs are governed by that provider's terms and our DPA. Single-tenant deployments can restrict providers per policy.

Evaluation integrity

• Judges are rotated and calibrated quarterly.
• Held-out splits prevent contamination.
• Scoring outputs are signed and versioned.
• See /leaderboard/methodology.

Red-teaming

Every agent that appears on the public leaderboard passes an Agent 007 red-team suite. Enterprise deployments extend this with domain-specific adversarial cases.

Incident response

If an agent causes harm, we will (i) freeze the affected tool and version, (ii) notify customers within 72 hours where material, (iii) publish a post-mortem on the blog.

Contact

Policy questions: hello@xploreintelligence.co.uk (subject: Responsible AI).