AI you can put your name on

This is the invariant the whole system is built around. An agent can structure your spoken data into trials, draft a SOAP note, surface a mastery or phase-change proposal, assemble a reauthorization packet, or answer a question about your caseload. It can never sign a note, decide a proposal, or submit a claim. Every clinical or billable record reaches the permanent record only through an explicit human confirmation, and every one of those promotions is logged.

The fastest way for an AI to be wrong in a clinical note is to invent a number. Cadence guards against it directly: the data behind a note is computed deterministically from confirmed records, the agent is instructed to reuse those figures verbatim, and a data-anchor check runs on the draft — any percentage that doesn’t match a committed measurement is flagged for the clinician to verify before they sign. When you ask Cadence to revise a note, that same check re-runs on the revision, so a rewrite can never smuggle in a figure that isn’t in your data.

Each agent is graded against a labeled corpus of real ABA scenarios — capture accuracy on narrated sessions, and note quality on committed data — with the anti-fabrication guard asserted as a regression test. And because clinical work can’t depend on a model being reachable, the core paths degrade gracefully: eyes-up tap capture and a deterministic note template work with no AI and no connectivity at all.

Model calls are made with a de-identified token in place of the child’s name — you see the real name in the app, the model never does. Protected health information is encrypted at the field level, our AI provider operates under a signed HIPAA business associate agreement, and analytics are kept off every page that shows PHI. Every agent action — what it read, what it proposed — is written to a tamper-evident, append-only audit trail, so the record of what happened is always reconstructable.

• Sign a note, report, or assessment — that attestation is yours.
• Decide a clinical proposal (mastery, phase change, target change) on their own.
• Submit a claim to a payer, or invent a unit, code, or authorization number.
• State a measurement that isn’t in your committed data.
• See a client’s name, or act anywhere off the audit trail.
• Automate a clinical decision — even the one low-risk action Cadence can take on its own (archiving a plan whose targets are all complete) is a configurable, off-by-default, non-clinical convenience.