Today
| Form | Reason | Author | Due | |
|---|---|---|---|---|
| Surgical Consent | Conditional section added | JJunior draft | 2 days | |
| Conditions of Admission | Scheduled review, TJC cycle | Dana Okonkwo | 9 days | |
| Home Health Consent (es) | New language variant | JJunior draft | 4 days |
The hero states the partner thesis, not a feature list. The visual is a live render of the actual product shell, not a mock screenshot. The logo wall sits under the hero, never inside it.
Junior launches from the EHR, prefills from FHIR, and goes live the same afternoon. No HL7. No interface engine. No services engagement.
Junior drafts forms, translations, metadata, and functions. Everything it authors stays in agent ink until a qualified human accepts it. Approval is the moment the color changes.
Autonomous actions are logged and reversible. Content changes to a consent are gated before the write, never after. Clinical risk language is human-only.
One reserved color, one meaning, on every surface.
Patient and Encounter map to one flat dictionary. Designers never touch FHIRPath.
Nine standard collections. Review cycles tracked. Outdated regulatory forms flagged on sight.
Trial, register, link by Issuer ID, pay, provision, first form launched.
Collection names, counts, languages, and revision currency are public, so the catalog earns search traffic. The forms themselves open after a work email. No competitor publishes a library at all.
Rev 2026.02 with laterality, CPT lookup, and dual signature.
Tracked against CMS revision schedule.
One work email opens the full catalog. No call, no demo booking, and no card. You land in the trial from here.
Personal email domains are asked for an organization name. Lead lands in the marketing engine with source attribution.
The funnel spec requires that a failed payment or provision rolls back cleanly with no half-provisioned or double-charged tenant. That failure state is a designed screen, not an error toast, and it says plainly that no charge was taken.
The rail, the canvas, and the thread. The overnight summary is the honest answer to "what did the AI do while I was gone," which is the question that decides whether staff trust the platform.
| Form | Reason | Author | Due | |
|---|---|---|---|---|
| Surgical Consent | Conditional section added | JJunior draft | 2 days | |
| Conditions of Admission | Scheduled review, TJC cycle | Dana Okonkwo | 9 days | |
| Home Health Consent (es) | New language variant | JJunior draft | 4 days |
The query is the one from PRD section 15.2. Results render as the real library grid with the matching gap called out, not as a chat answer about the library. Literal filters stay available beside the AI result so nobody is trapped in the conversation.
| Form | Collection | Form no. | Languages | Governance | Rev | |
|---|---|---|---|---|---|---|
| Surgical Consent | Informed consents | CON-0114 | en es vi | Approved | 2026.02 | |
| Conditions of Admission | Admissions | ADM-0002 | en es | Review due | 2025.11 | |
| Release of Information | Releases | REL-0031 | en | Approved | 2026.01 | |
| Medicare Outpatient Notice | CMS forms | CMS-10611 | en es | Outdated | 2024.07 | |
| Pediatric Immunization Consent | Vaccination | VAC-0077 | en es | Approved | 2026.02 |
Showing 5 of 23. Junior surfaced the match. It has not changed anything.
The signature screen of the system. Junior answered in the thread, but the change itself renders as a diff over the real form, and the approve control lives next to the change rather than inside the chat. Try Approve. The commit is the color moving from agent ink to human ink.
I have been informed of the nature of the procedure, the risks and benefits, and reasonable alternatives, and I consent to the procedure described above.
A diff, never a transcript. The research is explicit that an approval surface which dumps the agent's reasoning defeats itself. Rejection requires a documented reason, and the picklist answers the free-text problem raised in PRD section 11.3.
| 03 Aug 06:12 | JJunior drafted es variant | Autonomous draft |
| 03 Aug 09:41 | Marisol Vergara approved | Approved |
| 03 Aug 09:44 | Awaiting second approver | Pending |
Charts follow the visualization rules: fixed categorical order, a single axis, thin marks, recessive grid, legend plus selective direct labels so identity is never color alone. The series palette is deliberately not derived from agent magenta, because series identity must never be mistaken for agent attribution.
Patient device closed a 17 point gap after the intake redesign. Sample data.
Progressive disclosure, one idea per screen, 17px question type, 48px targets. The ribbon on the right is the DocuSign pattern called for in PRD section 10.9. Junior is present but subordinate, and it explains rather than fills.
Tobacco use changes how your body heals after surgery, so your care team needs it before your procedure. It is not shared with your insurer by this form.
Voice is an optional component. This surface only exists because the health system turned it on and executed a BAA covering the speech path. Two gates therefore: the tenant accepts the risk in AI settings, and then the patient sees exactly where their voice goes before they speak. Push and hold, no wake word, and the transcript maps to fields for review. Nothing is committed by speaking.
You can say your answers instead of typing them. North Kansas City Hospital turned this on for their patients. Here is exactly what happens to your voice.
You will see everything Junior heard, and where it wants to put it, before anything is filled in.
The designer sets the regulatory tier per signature, which is the edge described in PRD section 10.6. Patient signatures are simple electronic signatures upgraded with a field hash so tampering is detectable. Staff signatures are advanced, keyed to a PIN or a device passkey that Junior can validate but never generate.
By signing you confirm you have read the consent above and had a chance to ask questions.
| Signature field | Signer | Tier | Method |
|---|---|---|---|
| Patient consent | Patient | SES | Drawn or typed, field hash |
| Patient initials, page 3 | Patient | SES | Initials, field hash |
| Witness | Staff | AES | Device passkey |
| Surgeon attestation | Staff | AES | PIN protected key |
Voice ships as an optional component, off by default. A tenant turns it on only after an authorized signer accepts the risk and executes a BAA covering the speech path. Until then the mic is absent from the patient surface entirely, not disabled and greyed out.
Patients and staff can speak answers instead of typing. Speech is transcribed and mapped to fields for review. Nothing is written without confirmation.
| Requirement | Status | Record |
|---|---|---|
| Risk acknowledgement | Accepted | D. Okonkwo, 22 Jul |
| BAA covering the speech processor | Executed | signed in Junior |
| Data residency | United States | us-east |
| Audio retention | Not stored | transcript only |
Turning this off removes the microphone from every surface immediately. In-flight transcripts are discarded.
Extract member ID and group from a photographed card. Sends an image containing PHI to the configured model.
| Task | Model | PHI |
|---|---|---|
| Form design | claude-opus-5 | None |
| Library search | claude-haiku-4-5 | None |
| Translation | claude-sonnet-5 | None |
| Voice transcription | on-device, then BAA path | Possible |
Self-managed partners point these at their own endpoints. Any task marked as possible PHI is unavailable until the matching optional component is enabled.
Everything is a token, which is how the white-label requirement is met structurally rather than by exception. Switch partner themes in the top bar and watch the human channel move while agent magenta holds. Switch the artifact theme and watch the dark palette, which is selected rather than inverted.
No component may use these for any other purpose. That rule is enforceable in code review with a single search, and it is what makes the autonomy tiers visible instead of merely documented.
Partners swap this channel along with the human one. What cannot be swapped is the fact that there are two channels and that they are tellable apart. Switch partner themes in the top bar to see it move.
Because partners choose their own agent hue, a bad choice could make agent-authored content indistinguishable from an error, or from the partner's own brand. The branding screen runs a perceptual distance check before a theme can be saved.
| Check | Against | Result |
|---|---|---|
| Agent vs brand | --brand | Pass |
| Agent vs critical | --crit | Pass |
| Agent vs warning | --warn | Pass |
| Agent vs good | --good | Pass |
| Agent vs info | --info | Close |
| Contrast on panel | --panel | AA |
Run in both light and dark. A close result warns. A fail blocks the save.
Each ships with an icon and a text label so state never rests on color alone.
Assigned in fixed order and never cycled. A ninth series folds into Other or becomes small multiples. Validate against both surfaces before shipping any chart.
Production targets ABC Diatype or Sohne with a real mono. The prototype uses a system stack because the artifact sandbox blocks font CDNs and inlining faces as data URIs would add hundreds of kilobytes to a review build.
Two elevation steps, tinted to the ground rather than pure black. Motion is restrained by policy, because in a clinical product anything that pulls the eye off a consent field is a safety issue. Everything collapses under reduced motion, except the approve color change, which carries meaning and stays as an instant swap.
Rejection always records a reason. Pick one so approvers rarely have to write free text.