Agent quality and guardrails

Every run in this browser is scored by the deterministic guardrail engine. This is the observability layer: how often the agent output is grounded, safe, and complete, and which checks catch what. No backend, no account, this browser only.

No runs yet in this browser.

The figures here aggregate real guardrail results across the notes you run.

Run the demo

Immutable audit log

Every run seals a record (evidence, model versions, prompts, and the harness decision) into a hash chain: each entry hashes the previous entry, so tampering with any earlier record is detectable by replaying the chain. In this demo it persists in your browser; in production the same records go to append only storage server side.

No audit entries yet. Run the demo to seal one.

The guardrail suite

12 of 12 red team cases pass

Every run is checked by these deterministic layers before it is trusted. The suite itself is graded by a red team harness (run npm run grade-guardrails), so the checks are validated to fire on injected violations.

input
  • input.injectionhigh
    The note carries no instruction injection aimed at the model.
grounding
  • grounding.codesblocking
    Every code in every report traces to the validated extraction or the note.
  • grounding.numbershigh
    Every dollar and percent figure traces to a priced service, the benefit math, the note, or a published rate.
  • grounding.billablemedium
    No non billable category header is presented as billable.
policy
  • policy.denial_probabilityblocking
    No report states a denial probability (no data exists to ground one).
  • policy.prescribingblocking
    No agent issued prescribing or diagnosing.
  • policy.phi_isolationblocking
    The aggregate employer view carries no identifying detail.
  • policy.injection_resistedhigh
    No report echoes instructions injected into the note.
  • policy.cost_estimate_labeledmedium
    Costs shown to the patient are labeled as estimates.
consistency
  • consistency.readinessmedium
    Report claims about claim readiness match the checklist.
style
  • style.dasheslow
    No hyphens or dashes in any report.
quality
  • quality.structurelow
    Every report has a summary, at least two sections, and at least one action.
  • quality.dx_coveragemedium
    The primary diagnoses appear in the reports.