Responsible AI.

01Principles

Six principles that govern every model we build and every deployment we operate:

• Explainable by default. If we can’t describe why a model made a call, we don’t ship it.
• Human in the loop. Every decision with legal, safety or financial consequence has a human reviewer.
• Auditable. Inputs, outputs, model versions and reviewer actions are logged for the contractually agreed retention period.
• Bounded. Models are scoped to specific tasks; we don’t reuse customer data to train shared models without separate consent.
• Contestable. Affected individuals can request a human review of any decision made about them.
• Right-sized. The smallest, most efficient model that meets accuracy requirements is preferred.

02AI lifecycle

Every model we ship passes through a documented lifecycle:

1. Problem framing — written impact assessment; high-risk classifications trigger a documented pre-release review.
2. Data sourcing — provenance documented; consent or contractual basis recorded; demographic balance assessed.
3. Training — version-pinned data + code; reproducible builds; model card committed before release.
4. Evaluation — accuracy by sub-group; calibration; adversarial & edge-case testing.
5. Deployment — staged rollout; canary monitoring; kill switch in operator UI.
6. Monitoring — drift, fairness deltas, error rates by sub-group, monthly review.
7. Retirement — formal sunset with notice; data archival per contract.

03Human oversight

For every product:

• VIEW — operator confirms or dismisses every alert; the system never acts autonomously.
• ICCC — dispatch decisions are made by humans; AI surfaces priority and evidence.
• SAFER — access, tailgating and watchlist alerts route to a named operator; door release and enforcement actions remain human-reviewed.
• LENS ID — biometric matches above threshold are reviewed by a named operator before any consequence; HyperFUSE research models reach a deployment only through LENS ID and inherit the same gate.
• NMS — automated remediations are limited to network self-healing actions reviewed quarterly.

04Fairness & bias

For models that make decisions about people we measure performance across protected attributes (where lawful to know) and publish the deltas in the model card. Where deltas exceed our threshold (≥3 percentage points on the primary metric) we hold the release until the gap closes or document a justified exception with recorded sign-off from the accountable executive.

The LENS ID matchers are evaluated under NIST FRTE 1:N; the published NIST results are the reference we hold ourselves to.

05Explainability

Every inference returns:

• The output (e.g. confidence score, classification).
• The evidence that produced it (frame timestamps, feature attributions, contributing modalities).
• The model version and the operator-visible reason if any rules-based logic also fired.

This is what we mean when we say "explainable AI" — the audit trail is shipped with the call, not reconstructed after the fact.

06Red lines — what LENS will not do

07Governance

AI review

High-risk model classifications are reviewed before release and the decision is recorded.

Accountable executive

Dr. Debayan Deb, Chief Executive Officer

Whistleblower channel

legal@lenscorp.ai — anonymous receipt available.

08Contact

Questions, concerns or contestability requests: legal@lenscorp.ai. We commit to respond within 10 business days.