Lattice Health

The most interesting thing about Lattice Health is what it refuses to build. Every clinical AI vendor ships a dashboard; Lattice's homepage makes the case that leadership reads email, not SaaS, and ships one PDF per role at 07:00 instead. That is an editorial decision about the audience, and it shapes everything else. The IT Director's report carries alerts fired and acknowledged in the last 24 hours, latency p95 versus target per model, silent-update detections and posture changes. The CMIO/CIO report carries drift (PSI) by model with baseline comparison, HHS §1557 subgroup fairness disparity, a radiologist thumbs-up/thumbs-down rollup, and — notably — a selection-bias caveat in plain language, which is the sort of honesty most vendors' reporting layers omit. The Compliance Officer report carries continuous posture against ARCH-AI, Assess-AI, ONC HTI-1 and FDA PCCP envelope status with automatic breach detection, plus chain-of-custody evidence packets.\n\nThe trust architecture is where it stands apart from generic model-observability tools. Each institution holds its own private signing key; the public key is yours to publish. Any signed artifact verifies offline with standard open-source tooling, with no Lattice account, server or internet connection, which matters when a report travels three forwards deep into counsel's inbox. The chain-of-custody packets are transparency-log anchored. Public design records separate what is enforced cryptographically from what is enforced procedurally — an unusual and welcome distinction, and the reason its standards claims are readable rather than decorative.\n\nThe privacy model follows the same logic. Lattice works with signals, not patient records, ingesting HL7 and DICOM feeds with IHE-aligned UPS-RS work-list status, AI-Results and ERA classification shapes that PACS and EHR teams will recognise. A PHI redactor runs in strict-mode default and a failure blocks the entire batch before egress. That trades throughput for safety in a way a hospital compliance office will accept and a data-science team may find frustrating. Radiologist feedback is aggregated to model level only, so the 'readers are pushing back' finding never becomes surveillance of an individual reader.\n\nWhere it does not fit: it is a monitoring layer, so it does nothing in an environment without deployed clinical AI, and it will not satisfy a team that wants live streaming telemetry rather than a daily artifact. The per-role reports are shaped for leadership rather than for data scientists tuning models — there is no modeler-facing workbench described anywhere on the site. Deployment is on-prem or private cloud, so network, key-management and air-gap operations land on your team; the vendor notes deterministic narration is a first-class path so no external API is required for the core workflow, which helps in air-gapped environments but puts the operational burden squarely on hospital IT. Alerting goes out over email, Microsoft Teams and