MiddyMindCleared AI
Turns an app's architecture, data flows, and controls into signable regulator-grade artifacts. Every paragraph carries a citation chain and a calibrated confidence word-label - ungrounded claims are withheld, not guessed.
The composite 'application' entity is resolved across code, docs, and interviews; the regulatory-diff engine requeues only the affected sections when a rule changes.
Step-by-step
4 steps- Step 01
Register an application
You do → Point Cleared AI at a repo, an architecture doc, and a HECVAT draft.
You'll see → The app resolves into one composite entity - components, data flows, third parties, and jurisdictions surfaced as tags.
Under the hood → identity/ merges across sources; sub-threshold matches park in a review queue.
- Step 02
Pick a framework
You do → Choose from FDA SaMD, HIPAA, ONC HTI-1, HECVAT, 510(k), FHIR, or a state AI law.
You'll see → The registry loads 19 tools spanning the six frameworks; the orchestrator plans the artifact.
Under the hood → 6 specialist agents + orchestrator; costs capped per hour by governance/.
- Step 03
Generate + inspect citations
You do → Run the artifact and open any paragraph.
You'll see → Every sentence shows its citation chain and a confidence word ('supported', 'likely', 'insufficient evidence'). Insufficient sections are withheld with a reason.
Under the hood → eval/ grades the run; ungrounded claims never render.
- Step 04
Simulate a rule change
You do → Toggle an ONC HTI-1 clause and re-run.
You'll see → The regulatory-diff engine requeues only the affected sections - the rest of the artifact is untouched, and the audit trail shows why.
Under the hood → memory/ replay + bitemporal diff scoped to the changed rule.
Shipped in this vertical
- 19 registry tools · 6 regulatory frameworks
- 7 seeded agents: orchestrator + 6 specialists
- Regulatory-diff engine + citation-chain viewer