Trustworthy AI Policy

Detectory AI is operated by Tactical Edge AI to help program integrity teams detect document fraud, synthetic identities, and coordinated enrollment fraud. Because these workflows may affect health coverage, public benefits, insurance claims, lending, or other regulated decisions, we design Detectory for human oversight, explainability, security, auditability, and customer control.

Our approach follows a risk-based model informed by the NIST AI Risk Management Framework, sector-specific customer obligations, and public-sector expectations for transparency and due process.

• AI never makes eligibility decisions. Detectory scores risk and surfaces evidence. Trained investigators decide.
• No automatic terminations.Our platform cannot cancel, suspend, or deny an enrollment. Adverse actions always require human review through the customer's established due process.
• Escalation levels are review gates. Our five-level case workflow requires increasingly senior human sign-off as case severity rises.

• Every risk score is accompanied by the specific signals that produced it, in plain language an investigator can act on and a court can understand.
• Every model output, reviewer action, and case decision is recorded in an immutable, referral-grade audit trail.
• We document model purpose, training data characteristics, known limitations, and evaluation results for every deployed model.

• Models are evaluated for disparate performance across demographic groups before deployment and continuously in production.
• Protected characteristics are never used as model inputs.
• We publish evaluation methodologies to customers and support independent audits of deployed models.

• Customer data is used only to provide the contracted service. We do not train shared or cross-customer models on customer data without explicit written consent.
• We minimize access to sensitive program data, protect it with encryption and least-privilege controls, and support retention and deletion instructions defined by the customer agreement.
• Our control roadmap is designed to support customer reviews against SOC 2, CMS MARS-E, NIST 800-63, CMMC 2.0, StateRAMP, and FedRAMP-aligned requirements where applicable.

• We test models before production use and monitor for drift, degraded performance, false positive patterns, and operational failure modes.
• We restrict uses that would bypass required human review, target protected classes, or use Detectory outputs as the sole basis for an adverse action.
• We maintain incident review procedures for suspected model, security, or data handling issues with potential customer or public impact.

Our AI governance program assigns a named owner for every production model, requires pre-deployment review by a cross-functional group, and requires incident reporting for suspected model failures with potential public-program impact. Concerns can be submitted through the contact form.