A working prototype. A research instrument to qualify.

Three distinct things you can evaluate.

• The backend prototype: source-bound plan interpretation, policy checks, decision receipts, and observation-recording components exist in code. That does not establish their fitness for a clinical study.
• The protected professional evaluation: an invitation-only workflow records identified reviewers’ judgments of synthetic cases. It is separate from patient research and remains gated.
• The public Maya demo: a browser-only illustration with fixed synthetic records. It does not call the backend engine, run randomization, deliver an intervention, or collect adherence reports.

The continuity prototype is a candidate foundation for the proposed JITAI instrument. The connection still needs integration and verification against a study protocol. Maya illustrates the intended research record; it is not evidence that this complete pipeline operates today.

Component status: implementation is not validation.

Implemented component means present in the backend code, as reported by the project lead. This inventory is not an independent scientific or security review. Demonstrated means inspectable with synthetic examples. Proposed study capability means not established as a qualified end-to-end research capability. These labels do not certify reliability, clinical safety, or effectiveness.

How the engineering supports the question.

1. Represent the plan. Source-bound compilation separates documented goals, permitted support, prohibitions and handoff conditions. An interpretation is not permission to act.
2. Check a proposed action. Explicit policy gates check authority, operating mode, plan boundaries and relevant constraints before producing a disposition.
3. Preserve the decision. Receipts carry versions, hashes and reasons so a reviewer can inspect the basis of a proposal. Reconstructing any historical decision requires complete source history and the relevant versions; a receipt alone cannot supply missing history.
4. Keep observations distinct. The learning-ledger code distinguishes assignment, exposure, observation and reward. Its application-level protections reject edits and deletion of evidence rows; this is not a claim that a database administrator cannot alter storage.

These components are candidates to reuse and test. They do not establish a novel architecture, a tamper-proof record, a validated adherence measure, or an operational micro-randomized trial.

Where generative AI is used.

The public Maya and four-case walkthroughs use fixed text and deterministic browser behavior; they make no model calls. The separate backend includes an optional OpenAI-assisted synthetic conversation path for structured scene interpretation and grounded language composition, alongside deterministic rules and fallbacks. Its use depends on configuration.

Model interpretation can influence a proposal; it can be wrong. Policy checks constrain proposed actions, but their existence is not proof of safety. Evaluation must assess both interpretation errors and boundary enforcement. AI-generated language is not an independent source of clinical authority.

Records and operating boundaries.

Potential inputs include documented nutrition interventions, appointment history and other authorized records. Available integrations and source fidelity must be verified for each use; this is not a claim of a finished placement in every practice platform.

In shadow mode, proposed actions are for review. Nothing is automatically sent to patients. Prior silence never increases KeyRD’s authority to contact a patient. A visit limit is context for professional review, not permission to contact someone or infer coverage.

As of October 21, 2026, KeyRD does not process real patient records. The public demonstrations and protected synthetic evaluation use fictional cases. Any future real-record evaluation or research would require its own scope, permissions, data agreements and applicable ethics review before data use. Data permissions and accountability remain separate.