Evidence & limits
What has been demonstrated — and what has not.
Credibility depends on keeping proof of concept, verification and validation separate.
Demonstrated
That structured capture is possible. A real-world longitudinal record — 536 annotated records over 68 consecutive days, aligned to 21,226 CGM readings and 498 dose entries — can pair physiological telemetry with structured decision-state context and hold that context as an auditable record. Internal dataset and codebook verification confirms that the counts reconcile, the day structure is complete with no gaps, and every variable is classified by type in the codebook.
That the layer carries signal the device does not. Annotation surfaced recurrent within-subject patterns absent from the glucose trace alone — corrections delayed after recent hypoglycemia, under-correction under elevated cognitive load, environment-dependent dosing adjustment.
Forensic case-study series
The proof-of-concept record has also been examined through structured forensic case studies. These are not clinical trials and are not presented as population validation. Their purpose is narrower: test whether the architecture can distinguish decision states that look similar in telemetry but differ materially in provenance.
Tests whether the sequence of observations, insulin-on-board constraints, treatment actions and later outcomes changes the interpretation of an otherwise familiar glucose event.
Tests whether a system can possess relevant information yet still fail because it has classified the human participant's role incorrectly — patient, expert operator, reviewer or override authority.
Examines the state after an irreversible action, when the relevant cognitive task changes from choosing an action to monitoring consequences and preserving response capacity.
Examines why identical or alarming physiological endpoints cannot, by themselves, establish whether the preceding decision was error, delay, constraint-driven action or expert judgment. Detailed treatment logic is deliberately not reproduced publicly.
Independent professional review
Separate professional-review portfolios contain assessments from physicians, allied-health professionals, commercial/regulatory specialists and professional reviewers. Those records are evidence that qualified readers found aspects of the problem definition, documentation, clinical plausibility or cross-domain relevance credible. They are not treated here as proof of efficacy, regulatory clearance, population validity or institutional endorsement.
This distinction matters: external review can validate that a structural problem is recognizable and worth testing while leaving the architecture's predictive performance and clinical usefulness unresolved.
Not yet established
General clinical benefit. A single-subject implementation does not establish cross-patient generalizability, clinical efficacy, safety benefit or causal attribution, and nothing here should be read as claiming otherwise.
Reliability. Every annotation comes from one annotator, who is also the subject and the architect. Inter-rater reliability is unmeasured. Until it is, the possibility that the manifold encodes one person's introspective idiom rather than a generalizable structure remains open.
Feasibility at lower expertise. The annotation density here was produced by someone with six decades of lived expertise and a strong motive. Whether a usable subset survives contact with ordinary clinical conditions is unknown.
What would falsify it
Stating this in advance is part of the method.
- Independent annotators applying the framework to the same events produce materially divergent records. That would indicate the layer captures idiom, not structure.
- The decision layer adds no predictive or explanatory information beyond what is already recoverable from physiological data. Then the layer is redundant, whatever it feels like from the inside.
- Patterns observed here fail to appear in any other subject under comparable capture. Then this is autobiography with a schema attached.
- Capture at this density proves unsustainable outside expert self-annotation, with no viable reduced set. Then the architecture is correct and unusable, which is the same as unusable.
Path to validation
- Inter-rater reliability study — multiple annotators, shared events, measured agreement.
- Cross-user evaluation, including a reduced variable set for non-expert annotation.
- Independent design verification and formal risk analysis.
- Prospective study design against defined intended use.
- Regulatory pathway determination for any future intended use.
Ethics and data
The dataset is single-subject and self-annotated. The subject, the annotator and the architect are the same person, who consents to its use and publication. No third-party patient data is involved, and no institutional cohort was recruited. That removes the usual consent complications and introduces a different one, stated plainly above: the annotator cannot be blind to the hypothesis.