Changes
What changed, and when.
Newest first. One line per change. Corrections are published rather than quietly fixed.
27 October 2026
- What's New updated. Three entries: the Meaningful AI Act discussion draft, Med Management (IVs Are Not Enough), and Ami. What's New.
- Documentation adds a section for the bill (web page and PDF, hosted at danielpettus.com), the Ami privacy policy and Med Management. Documentation.
- The bill's HTI-5 finding is current. HHS finalized the removal of the HTI-1 source-attribute requirements on 26 October 2026; the draft, Rev 3, says so, and its clinician protection now has enforcement. The draft.
16 October 2026
- New page: Ami privacy policy. What the iPhone app collects (nothing), what leaves the phone (requests to the FDA's public openFDA service, without your medicine names), and what TestFlight testing adds. Privacy policy.
12 October 2026, afternoon
- New section: Standing Watch. A tenth box on the front page and a menu item. The section.
- The ICU watch board. Five synthetic beds and a simulated clock. Every finding is carried by a mediator (recipient and escalation chain, criticality and reason, reminder and escalation timers, respond, defer or reject, audit trail) and shown as FHIR R4 DetectedIssue, Task, CommunicationRequest and AuditEvent. A button sends a bed's chart, with its simulated pump feed, to the live reasoning; warning and critical cards are wrapped in a mediator, information cards are shown and sent to no one. Watch board.
- One patient, five rooms. ICU, ward, discharge, home and after day 30, with one audit trail, downloadable as a FHIR Bundle. Scenario findings, labeled as such. Five rooms.
- The bridge. A going-home packet encrypted in the browser (JWE, dir, A256GCM) and carried as a SMART Health Link with a QR code; Ami opens it holding a pseudonymous ID only. In this demonstration the blind server is the visitor's own browser storage, and the page says so. The bridge.
- The design pitch is public (Rev 4, 13 slides, links to the new demonstrations), with a new Standing Watch datasheet. QR codes are drawn with qrcode-generator by Kazuhiko Arase (MIT license), served from this site.
12 October 2026
- One Formulary becomes Smart Formulary and Smart Orders. The front-page box, the menu and the section page are renamed. The open-source tools keep their names: One Formulary (Apache 2.0) and Library Match are the Smart Formulary half. The section.
- New: Smart Orders. A demonstration that drafts one individualized IV order (propofol) for a synthetic patient from the published Schnider pharmacokinetic model, shows the model's range and uncertainty, writes the order as a draft FHIR MedicationRequest, and asks the live reasoning to review it. The slow-circulation adjustment is labeled illustrative, not validated. Smart Orders.
- Credit where it is due. The section now explains where Smart Orders comes from: Dr. N. Ty Smith's circulation and uptake models (1972) and his BODY simulation, with sources.
- Smart Orders is not open source. It is patent pending and all rights reserved, and its footer says so, so nobody mistakes it for part of the Apache 2.0 tools.
30 September 2026
- AI MedAgent is a registered Oracle Health app. It signs in through Oracle's authorization server, reads the chart from Oracle's FHIR R4 sandbox servers, and takes device data from an OpenBedside gateway now running in the cloud. What is real, what is simulated, and how to check it.
- The medication screen reads the whole laboratory record, not only renal results, and checks it for physiologically impossible values before any clinical content. A card says so when it finds them.
- Device data reaches the reasoning. A bedside gateway can send the pump's state with each request. The patient match is checked by rule, and the model is told a pump rate is not a dose.
- The reasoning card uses the coverage counts it is given and states lab defects as proportions. It no longer recounts, and no longer generalizes one bad test to all labs.
2 September 2026
- The access code is retired. The demonstration is open to anyone. A daily allowance, counted per day and per visitor, replaces the gate; when it is used the page says so and labels anything shown afterwards as a note written in advance, not generated live.
- The evidence is on the page. Every alert now shows the federal data the reasoning read: RxNorm class, FDA adverse-event counts with the data date, DailyMed label status, ClinicalTrials.gov study count, each with the URL that produced it, and a line saying how many of the source queries answered.
- Every alert is labeled "generated live" or "validated note, not live." The status line shows how many live runs remain today.
- New page: the method in plain language. New page: this one.
- Documentation reordered, method first, with a "For:" line under every document. A two-hundred-word summary added to the top of the EHR evidence page.
31 August 2026
- PCD-E Sandbox Ingest Specification v0.4: section 10a documents the conformance block every response carries; section 8.3 covers the two incompatible encodings of keep-vein-open seen in the field. PDF.
- The home page names MEDITECH alongside Epic and Oracle Health as a live record source.
29 August 2026
- Infusion Pumps on Hospital Networks, version 1.0: evaluation and wireless test criteria with scoring, open access, vendor neutral. Supersedes the 2020 draft. PDF, 13 pages.
27 August 2026
- Site restructured around one navigation: AI Reasoning, EHR Integration, Med Device Sandbox, Documentation, About, Contact. Four pages added.
26 August 2026
- Five Minutes and a Reason, the memorandum behind the device sandbox, published. PDF, 20 pages.
- Sandbox keys mint themselves at
POST /device/key with an organization name. No approval, no email.
25 August 2026
- EHR Integration Datasheet v0.3 and The connection was the easy part, the same findings in plain language, published.
24 August 2026
- MEDITECH Greenfield sandbox connected. Live records now read from three vendors.
21 August 2026
- CDS Hooks service live at /cds-services. Every call reports how many of the eight standard screening checks it could perform.
Earlier history is in the dated deploy notes kept with the source. Corrections to published claims are recorded on the pages they correct, not removed.