Frame
Claude
Structural · factual
Maps clinical frameworks, timelines, diagnoses, procedures, medications, and measurable metrics. Frame is the objective backbone — what happened, when, and in what order.
For clinical trial sponsors and CROs
QIS
Caregiver strain, oxygen non-adherence, social withdrawal, functional decline — the signals that precede a withdrawal are spoken aloud for weeks before anyone acts on them. QIS captures that stretch, reads it through three models at once, and flags what matters to a clinician while the participant is still enrolled.
What a trial cannot see
A participant who withdraws has usually been telling someone for weeks — a caregiver, a family member, a nurse in passing. None of it is a scheduled endpoint, so none of it reaches the sponsor until the patient is already gone.
Figures are published industry benchmarks, cited for scale rather than as QIS results. At a 30% withdrawal rate, a 300-participant study carries roughly $1.7M in replacement cost alone, before any delay to readout. QIS has not yet been measured against live trial data — what it offers today is the mechanism, not a proven effect size.
How it works
Four steps, each one already running in the product.
Caregiver conversations, patient interviews and site notes are ingested as first-class qualitative records — not free-text fields nobody reads. Trials already measure the visits; this is the rest of the month.
Every question runs against three frontier models under distinct analytical lenses. Where they agree is reported. Where they disagree is kept, because a split between models on a risk question is a signal, not noise to average away.
Defined criteria — caregiver strain, oxygen non-adherence, social withdrawal, functional decline — route a record to a clinician instead of into a dashboard. The flag names which criterion fired and why.
Every finding and every suggested action arrives as a proposal. A clinician accepts, edits or dismisses it. This is the property that makes clinical AI deployable, and it is enforced in code rather than in policy.
Why this is deployable
Every clinical AI connector shipping today is read-only, because nobody has a safe way to let a model change a patient record. QIS does not have to choose: a suggestion enters a clinician's review queue, and takes effect only when a person accepts it.
| Product | Surface | Can it change the record? |
|---|---|---|
| ChatGPT × Epic | Health-system EHR | Read-only — states it does not write back to the patient record |
| athenahealth for Claude | Patient records | Read-only |
| Third-party FHIR connectors | Patient records | Restricted to read operations so the model cannot modify data |
| QIS | Clinical digital twin | Can propose. A clinician approves, or it never takes effect. |
Each row above reflects that vendor's own published documentation as of November 2026. The difference is not that QIS trusts models more — it is that a care action has always been born as a proposal here, waiting for a clinician. Connecting an assistant changed nothing about that boundary; it only gave the proposal a new way in.
See it in practice
The same question, run two ways on synthetic trial data: once against the quantitative endpoints alone — the view a reviewer has now — and once across the full qualitative record. The difference between the columns is the thing a visit schedule cannot capture.
Clinician asks
Is she trending toward diabetic complications, and what am I missing?
Retrieving twin context · three lenses running in parallel
A1C 7.8% and rising albuminuria across the last two panels. Trend is the signal, not the single value.
She stopped mentioning her walks in the March transcript. Adherence may be slipping before the labs show it.
ADA guidance puts this profile in intensified management. Comorbid hypertension raises the CVD tier.
Convergence · 2 of 3 agree, 1 adds context
Risk is trending up on labs and on language. Intensify monitoring and probe adherence at the next visit.
Review the evidence and flagged risks before making care decisions.
What is already built
Ingestion, meeting capture with consent gating, the review queue, the care calendar and the connector surface are all running today. This is what a pilot would actually be run on.
Dashboard
Library
Chat & QIS Review
Meetings
Care actions
Calendar
Who uses it day to day
A sponsor buys the retention signal; a care team is who produces and acts on it. QIS is built for the people on the ward and at the kitchen table, because that is where the qualitative record actually comes from.
Keep your documents together and ask questions about your history. Return to your dashboard to see suggested actions and upcoming care.
Create your accountLearn QIS
The same guided tours the product ships with, readable here before you sign in. Every step is written out; open any one to be walked through it.
FROM INFORMATION TO UNDERSTANDING
Take a guided look at a familiar care journey.
Choose a step to see what happens next.
01 / THE RECORD
Start by choosing a patient and adding a document to the Library. QIS processes the text so you can read it, ask about it, and find suggested care actions.
You can add PDFs, Word documents, text files, spreadsheets in CSV format, or images of records. Wait until a file is ready before asking about it.
Start your own journeyYour files stay connected to the selected patient.
The QIS perspective
For questions that need a closer look, QIS Review considers the clinical facts, the human context, and potential risks together.
Frame
Claude
Structural · factual
Maps clinical frameworks, timelines, diagnoses, procedures, medications, and measurable metrics. Frame is the objective backbone — what happened, when, and in what order.
Soul
GPT
Contextual · human
Reads interview transcripts, caregiver notes, and patient language for goals, values, burden, and emotional signals — the narrative behind the numbers.
Steel
Gemini
Risk · evidence
Stress-tests findings against contradictions, red flags, medication interactions, and care-coordination gaps — the early-warning lens for what does not fit.
You stay in control. Review the source evidence and flagged risks. Care actions require confirmation, and meeting transcripts are reviewed before approval.
A little guidance
The essentials, without the technical detour.
Trials capture what happens at scheduled visits. Withdrawal is usually decided in between them — caregiver strain, medication or oxygen non-adherence, social withdrawal, functional decline. Those signals are spoken aloud to family and care staff for weeks before they reach a sponsor. QIS ingests that qualitative record, reads every question through three frontier models at once, and routes anything meeting defined clinical criteria to a clinician for review while the participant is still enrolled.
No. QIS runs on synthetic and de-identified data today and has not been deployed against live trial data, so no retention effect has been measured. What exists is the mechanism and a working product: ingestion, triangulation across three models, convergence and divergence mapping, eight defined human-gate criteria, and the review and scheduling workflow. Any figures we cite about the cost of attrition are published industry benchmarks, not QIS results.
No, and this is enforced in code rather than policy. Every finding and every care action is created in a 'proposed' state and waits for a clinician to accept, edit or dismiss it. That applies equally to a connected AI assistant: an assistant can propose a care action through our connector, and there is deliberately no tool by which it can approve one. Every clinical AI connector shipping today is read-only for this reason; the human gate is what lets QIS accept a proposal safely.
The disagreement is preserved and surfaced, not averaged away. Convergence is scored, findings reported by only one model are kept as divergent, and low convergence on a result containing risk signals is itself one of the criteria that routes a record to a human. A split between models on a risk question is treated as information about uncertainty rather than as noise.
Not in the demo environment. Production use with protected health information requires a signed business associate agreement and a dedicated deployment, and we will walk you through what that involves before you commit to anything. Everything you can see today runs on synthetic records.
Create an account or sign in to your dashboard. If a patient record is already linked, open the Library and add a document. Otherwise, caregivers and clinicians can create a patient from Chat; patients without a linked record should contact their care team. Once a file finishes processing, you can ask questions about it in Chat.
It is a connected collection of a patient’s records, approved meeting transcripts, and saved findings. QIS uses this shared context when answering questions, so you do not have to explain the history from scratch each time.
Start with Chat for everyday questions and explanations based on your records. Choose QIS Review when you want a deeper comparison: three AI perspectives examine the facts, personal context, and risks, then show where they agree or differ.
Upload PDF, Word (DOCX), CSV, TXT, or Markdown files. Supported images and scans can also be processed when image recognition is available. Wait until the file is ready before asking about it. If processing fails, the Library shows the error so you can check the file and try again.
Open Care actions to review suggestions extracted from records or meetings, together with their source evidence. Confirm, edit, or dismiss them as appropriate. Confirmed actions can then be scheduled in Calendar.
Bring us a trial.
Tell us the indication, the phase, and where you are losing participants. We will walk you through what QIS would have surfaced, on synthetic data, and be straight with you about what it has and has not been proven to do.
Research stage
QIS runs on synthetic and de-identified data today. It has not been deployed against live trial data, and no retention outcome has been measured.
PHI requires a BAA
Protected health information needs a signed business associate agreement and a dedicated deployment. We will tell you what that involves before you ask.
Patent pending
The triangulation and human-gate architecture is the subject of US patent application 63/935,100.