CLASS 01
Infrastructure
Historical internal result
Evidence that bounded artifacts can be generated, checked, scored, packaged, and reproduced. The public acute-trajectory note describes this class.
Evidence
What a result can establish.
Evidence class, observed result, and commercial availability are separate. Current public evidence documents infrastructure work; it does not establish a completed model-hardening outcome or broad real-world improvement.
Explore the research informing our questions and evaluation design
Historical internal result · artifact snapshot June 16, 2026
Checked synthetic acute-care trajectories across 96 hourly transitions.
The retained infrastructure work includes 64-hour and 96-hour synthetic trajectory arms. These are accepted synthetic artifacts, not real patients. The recorded validator recheck applies only to its accepted-only snapshot and is not an independent clinical assessment.
This is not a clinical error rate, patient-outcome result, future-patient accuracy estimate, model-safety finding, held-out trajectory evaluation, or model-hardening result. The validator recheck is not an independent clinical assessment.
Read the methodology, provenance, and limitationsSix evidence classes
These describe what evidence establishes, not six buying stages.
An engagement can conclude at a bounded class, with no effect, or with a supported stop. Classes are cumulative in meaning but are not inherited from architecture or a prior class. Each Evidence Record names the level actually established.
CLASS 02
Failure discovery
Not yet demonstrated as a public outcome
A controlled challenge reveals a clinically meaningful weakness ordinary evaluation did not surface. A reproduced known failure is a positive control, not novel discovery.
CLASS 03
Certified synthetic hardening
Closed-loop result not yet externally complete
A frozen intervention improves a declared behavior on a sealed candidate-blind synthetic proof surface without unacceptable regression. This supports a bounded synthetic-world behavioral claim.
CLASS 04
Customer cross-fitted hardening
Not demonstrated in the public evidence shown here
Patient- or episode-grouped cross-fitting can evaluate a frozen procedure when labeled customer outcomes exist. Each evaluated group must be excluded from its fold's adaptive pipeline.
CLASS 05
Real-world and transport validation
No result claimed here
Untouched customer, future temporal, source-distinct, or prospective evidence supports the corresponding real-world, temporal, site-transport, or broader claim.
CLASS 06
Transfer
Not yet demonstrated
Evidence that experience from earlier models improves challenge selection, failure classification, intervention choice, or efficiency on later models.
Proof pathways
Match the evidence design to the claim.
A second dataset is not universally required. Frozen materiality criteria, no-regression checks, and independent clinical or evaluator authority remain part of a defensible proof plan. Ordinary random-split evaluation is not a substitute for the grouped cross-fitting contract.
PATH 1
Sealed candidate-blind synthetic proof
Tests a bounded behavioral claim in the declared synthetic-world scope. Candidate blindness applies to the frozen generation and evaluation policy and its ancestry, not merely to hiding a file. No real-world clinical benefit follows from this proof alone.
PATH 2
Patient- or episode-grouped customer cross-fitting
Can evaluate a customer-scoped procedure from one labeled dataset. For every fold, the held-out patient or episode is excluded from profiling, selection, calibration, thresholding, and other adaptation. A final model fit on all customer data does not inherit direct holdout evidence for that exact candidate.
PATH 3
Untouched, source-distinct, temporal, or prospective evidence
Supports the corresponding stronger real-world claim. The evidence source and design must match the claim; these paths are not universal entry conditions for every bounded engagement.
Full lifecycle
Preserve the question, the result, and what remains unknown.
The sequence may be grouped in a diagram, but it retains initial inspection, failure adjudication, claim-appropriate proof, and negative results.
- 01Evidence Profile
- 02Controlled challenge
- 03Failure adjudication
- 04Targeted intervention
- 05Claim-appropriate proof
- 06Evidence Record
- 07Regression Suite
No result is promised in advance.
A defensible outcome can be improvement, no effect, a regression, a need for more real evidence, a narrower intended use, or a stop under the permitted actions and tested scope. None implies that every possible repair has been ruled out.