Rigor should be inspectable, not implied.
CareLoop combines an evidence-based communication method with a source-locking system designed around a documented failure mode of generative discharge instructions.
Reproducible engineering benchmark
Current suite passesThe exact-source path is tested against hand-labelled fixtures.
These are deterministic engineering results on synthetic fixtures—not clinical accuracy, patient-outcome evidence, or a medical-device validation study.
Teach-back checks understanding in a non-shaming way.
AHRQ describes teach-back as an evidence-based health-literacy intervention: the patient or caregiver explains, in their own words, what they need to know or do.
Read the AHRQ guidanceUnconstrained generation can change safety-critical details.
A 2024 clinician-reviewed study of 100 AI-generated discharge instructions found potentially harmful AI-attributable issues in 18%, including hallucinations and new medications.
Read the peer-reviewed study Browser-verified workflow
The real intake path is part of the release gate.
Twenty unit and API checks plus twenty desktop and mobile browser journeys exercise the safety boundary and the user-visible workflow.
- 01Selectable PDF extraction
A generated text PDF is read through the same bundled PDF worker used in production.
- 02Privacy review before analysis
The extracted text remains editable, and analysis stays locked until the user confirms de-identification.
- 03Plan, teach-back, and print
The suite completes every source-linked action, detects a wrong dose, and verifies source evidence in print mode.
- 04Desktop and mobile accessibility
Critical pages run on desktop Chromium and a Pixel 7 profile with automated accessibility and console checks.
What the current evidence proves
CareLoop enforces its own acceptance boundary.
Demonstrated
- Exact cited excerpts and valid source-line identifiers
- Exact number-and-unit preservation in accepted actions
- Rejection of invented clinical terms and duplicates
- Source-built feedback after teach-back classification
Not yet demonstrated
- Clinical safety or diagnostic effectiveness
- Improved outcomes in patients or caregivers
- Performance on scanned, multilingual, or handwritten documents
- Regulatory readiness for clinical deployment
See the acceptance boundary work.
The verified sample includes medicines, self-care, restrictions, follow-up, warning signs, and a deliberately unresolved detail that becomes a question instead of an invented answer.
Run the verified sample