FRAME — Forensic Repository Audit, Measurement & Evaluation — is the forensic layer of the PatientsUseAI Radar. The catalog (home) shows what exists. The audits show what holds — and what doesn't — when you press on the architecture, the safety posture, the evidence contract, and the path to non-technical patients.
Five repos. Eight dimensions. Three tiers. The foundation for patient-directed AI is converging on self-hosted local-first architecture with PolyForm Noncommercial licensing, but the access gap — non-technical patients, caregivers, non-English speakers — is still total.
First FRAME batch: five repos audited across eight dimensions. Three FOUNDATIONAL (Tula 23/24, OwnChart 22/24, OpenKP 21/24), one FUNCTIONAL (UHR 12/24), one HISTORICAL (YouBase 5/24). Three convergences emerged: self-hosting + local-first as the privacy standard, PolyForm Noncommercial as the emerging license norm, and critical AI health literacy as the philosophical anchor.
Most governance-mature patient AI agent in open source — frontier patterns, working evals, self-hosted, open-core commercial path. Missing explicit AI evidence contract.
Best patient agency philosophy in batch — evidence contract, immutable sources, user correction as canonical, no telemetry by architecture. Alpha state, single maintainer.
Cleanest CLAIM implementation in batch — patient-directed MCP surfacing how the institution characterizes patient engagement. Local-only, KP Northern California, 24 tools, 567 tests.
Principled consent-first design with best scope discipline in batch — but design phase only, no code, no FHIR/HL7, no AI layer. Blueprint not product.
2014-era founding document in patient data sovereignty tradition — individual ownership, third-party subscription model, BIP32/IPFS architecture. Draft 1, appears dormant. Historical ancestor of current FOUNDATIONAL projects.
Analysis Date: 2026-05-28
Repos: realactivity/tula, nickpdawson/OwnChart, hugooc/OpenKP, Aaqib-bashir1/unified-health-record, YouBase/white-paper (via unclenate)
Analyst: FRAME v1.0 / Synambix
| Repo | PAI | AI | TM | SDP | IS | CA | SM | SH | Total | Tier |
|---|---|---|---|---|---|---|---|---|---|---|
| realactivity/tula | 3 | 3 | 3 | 3 | 3 | 2 | 3 | 3 | 23 | FOUNDATIONAL |
| nickpdawson/OwnChart | 3 | 3 | 2 | 3 | 3 | 3 | 2 | 3 | 22 | FOUNDATIONAL |
| hugooc/OpenKP | 3 | 3 | 3 | 2 | 2 | 3 | 2 | 3 | 21 | FOUNDATIONAL |
| Aaqib-bashir1/unified-health-record | 2 | 1 | 0 | 2 | 1 | 2 | 1 | 3 | 12 | FUNCTIONAL |
| YouBase/white-paper | 2 | 1 | 0 | 0 | 0 | 1 | 0 | 1 | 5 | HISTORICAL* |
*Bottom tier scored as PROBLEMATIC by rubric; HISTORICAL is the accurate characterization. See individual audit for context.
Columns: PAI = Patient Agency Index, AI = Architecture Integrity, TM = Technical Maturity, SDP = Safety & Disclaimer Posture, IS = Interoperability Stack, CA = CLAIM Alignment, SM = Sustainability Model, SH = Scope Honesty
| Dimension | Top Scorer | Score | Notes |
|---|---|---|---|
| Patient Agency Index | Tula / OwnChart / OpenKP | 3/3 (tied) | Three-way tie at maximum |
| Architecture Integrity | Tula / OwnChart / OpenKP | 3/3 (tied) | Three-way tie at maximum |
| Technical Maturity | Tula / OpenKP | 3/3 (tied) | OwnChart penalized for alpha; YouBase/UHR no code |
| Safety & Disclaimer Posture | Tula / OwnChart | 3/3 (tied) | OpenKP missing medical device disclaimer |
| Interoperability Stack | Tula / OwnChart | 3/3 (tied) | OpenKP KP-specific; UHR/YouBase no standards |
| CLAIM Alignment | OwnChart / OpenKP | 3/3 (tied) | OwnChart evidence contract; OpenKP interrogative stance |
| Sustainability Model | Tula | 3/3 | Only project with commercial path + multi-contributor |
| Scope Honesty | Tula / OwnChart / OpenKP / UHR | 3/3 (four-way) | Only YouBase fails scope clarity |
| Dimension | Batch Mean | Notes |
|---|---|---|
| Technical Maturity | 1.6 | UHR and YouBase with 0/3 drag the mean; three FOUNDATIONAL repos at 3/3 |
| Interoperability Stack | 1.6 | YouBase and UHR with 0/3 and 1/3; even strong repos limited by platform scope |
| Sustainability Model | 1.6 | Only Tula has a commercial path; three projects with single-maintainer risk |
| CLAIM Alignment | 2.2 | Tula’s 2/3 is the only gap among FOUNDATIONAL repos |
These are capability absences across ALL five repos in this batch:
1. Non-technical patient access. Every FOUNDATIONAL project requires Docker Compose, Python venv, terminal commands, or Claude Desktop. The patient AI tool accessible to a 65-year-old with a Windows PC and no developer background does not exist in this batch.
2. Caregiver-mode support. OwnChart lists it on the roadmap. No project implements it. Caregivers represent one of the highest-leverage patient AI use cases — managing a parent or spouse’s fragmented records — and none of these tools serve them in a live, accessible way.
3. Non-Kaiser portal coverage. OpenKP covers KP Northern California. Tula covers generic SMART on FHIR, which means Epic/Cerner/athena in theory but not with the depth of OpenKP’s per-tool implementation. The equivalent of OpenKP for CommonSpirit, Ascension, HCA, or Intermountain does not exist in this batch.
4. Mental health record integration. Absent from all five repos. Mental health records occupy a distinct legal category (42 CFR Part 2 for substance use, state-level MH privacy laws) and are disproportionately fragmented. No project addresses this.
5. Rare disease and multi-institutional patients. Patients with rare diseases typically have records across five or more institutions, often including academic medical centers and international specialists. No project specifically addresses cross-institutional longitudinal aggregation for this population.
6. Patient-outcome evaluation. Tula uses Waza evals for technical compliance. No project measures whether patient use of these tools actually changes patient outcomes — comprehension, appointment preparation quality, medication adherence, clinical communication effectiveness. The field is producing governance evals, not patient outcome evals.
7. DICOM / imaging integration. OwnChart lists it as roadmap. No project has live imaging integration. For patients with cancer, orthopedic conditions, or cardiac disease, imaging reports and DICOM files are often the most important records.
8. AI literacy scaffolding for patients. OwnChart has the philosophy (PHILOSOPHY.md §19) but not yet the UX. No project teaches patients how to interrogate AI outputs, recognize uncertainty, or build reasoning capacity over time. The tools are built; the pedagogy is absent.
Specific underserved populations and use cases not covered by any repo in this batch:
Immediate whitespace:
Medium-term whitespace:
Structural whitespace:
Three convergence patterns are visible across the FOUNDATIONAL repos:
1. Self-hosting + local-first as the privacy standard. Tula, OwnChart, and OpenKP all chose self-hosted architectures. There is no FOUNDATIONAL cloud-hosted patient AI project in this batch. The patient AI field appears to be converging on “your server, your data” as the baseline rather than the exception.
2. PolyForm Noncommercial as the emerging license standard. OwnChart and OpenKP both chose PolyForm NC. Tula chose Apache 2.0. UHR chose AGPL. But the PolyForm NC pattern — “free for personal patient use, commercial use requires permission” — appears to be emerging as the community norm for patient-directed tools. This has significant implications for the open-source health AI ecosystem: it creates a class of tools that are explicitly not privatizable while remaining open for patients.
3. Hugo Campos / Critical AI Health Literacy as the philosophical anchor. OwnChart (§19 of PHILOSOPHY.md), OpenKP (README explicitly), and implicitly Tula (patient agency framing) all draw on the critical AI health literacy tradition. Hugo Campos is credited by name in two of the FOUNDATIONAL projects. This intellectual convergence around a specific community (aipatients.org, Hugo Campos, the ePatient movement) suggests the patient AI space has a coherent philosophical core, not just scattered technical experiments.
Risk: OwnChart and Tula are solving overlapping problems (self-hosted patient health AI) with different architectural emphases. If both attract developer communities and then stall at single-maintainer scale, the patient AI space gets two incomplete platforms instead of one complete one. The overlap is currently productive (different audiences, different emphases) but fragmentation risk increases as both mature.
Opportunity: The three FOUNDATIONAL repos have complementary capability profiles:
A patient using all three together would have capabilities that none offers alone. An interoperability layer between them — or a shared standard for patient AI output format — would compound the ecosystem’s value significantly.
Based on this batch, proposed calibration anchors for scoring the 886-repo PAVE corpus: