Hallucination detection
Measured classifier accuracy on internal benchmark
View evaluation noteSafety
Safety as architecture
Safety is embedded at every layer of the stack — from routing and input handling to runtime enforcement and post-hoc audits. Claims on this page are evidence-first: any numeric claim links to an Evaluation Note that documents dataset, protocol, sample size, version, date, and limitations.
Principles
- Evidence-first: every public claim must reference an Evaluation Note that includes methodology and limitations.
- Least-privilege default: systems default to conservative behaviors for sensitive inputs and require explicit opt-in for higher-risk behaviors.
- Human-in-the-loop: critical decisions include clear human escalation paths and explainable traceability.
- Audit-readiness: logs, traces, and versioned artifacts are retained to support institutional review.
Pre-inference
Pre-inference controls and practices that reduce risk before, during, and after inference.
- Routing and intent classification with risk-aware thresholds
- Dataset provenance and versioning captured in evaluation notes
- Opt-in/offline modes for sensitive workloads
Runtime
Runtime controls and practices that reduce risk before, during, and after inference.
- Hallucination detection at 94% measured accuracy (see evaluation note)
- Confidence calibration with 0.73 correlation to correctness (see evaluation note)
- Constraint runtime and policy enforcement in Mavaia
Post-hoc
Post-hoc controls and practices that reduce risk before, during, and after inference.
- Trace capture via ARTE and ACL pipelines
- Evaluation notes with limitations for every public metric
- Audit-ready logs for institutional governance
Evidence & evaluation notes
Key metrics are surfaced here and link to Evaluation Notes that record dataset, protocol, sample sizes, limitations, and source artifacts.
Confidence calibration
Correlation to correctness on held-out tasks
View evaluation noteLatency (deployment)
Representative median P95 for on-prem profiles
View evaluation noteIf a metric is important to your evaluation, request the full protocol and reproducibility notes via the contact form.
Governance, audits, and next steps
Institutional deployments follow a governance checklist: threat model, compliance mapping (e.g., HIPAA, FedRAMP), audit log retention policy, and a reviewer-approved Evaluation Note set for any numeric claims.