Research-facing overview
Technical scope without exposing the implementation
Detailed enough to evaluate the idea. Not detailed enough to reconstruct proprietary logic.
Study design
A single-subject (N-of-1) proof of concept. One annotator, continuously self-annotating across 68 consecutive days — 12 October to 18 December 2025 — with no gap days. Each treatment action was captured with its pre-decision state, context, uncertainty, timing and judgment, then aligned to device data on the same timescale.
Annotation is synchronised to the Moment-of-Action: the point of insulin dosing, recorded before the outcome is known. That ordering is the methodological claim. A record written after the fact is a reconstruction, and reconstruction is what the architecture exists to avoid.
Record structure
The dataset holds 536 annotated records, which decompose as 468 dated decision events plus 68 day-close records — one per day, timestamped at midnight, carrying cumulative day-level fields rather than a discrete decision. 468 + 68 = 536, and the day-close count equals the day count because there is exactly one per day.
Each record spans 341 columns across 21 annotation layers. Paired device data comprises 21,226 CGM readings and 498 insulin dose log entries.
Variable classification
Not all variables carry the same role, and the type determines how each is validated and interpreted. Of the 341 columns: 200 are derived, computed from source variables through defined formula logic; 69 are source, directly observed or entered and never formula-derived; 69 are mixed; and 3 are reserved and currently inactive.
Source variables are ground truth — the glucose reading, the time of injection, the affective snapshot taken at the moment of action. Derived variables encode relationships between them. Interpretive variables encode the system's reading of a state rather than the state itself, and are treated as probabilistic throughout.
The public architecture
ESP360° preserves state capture, temporal synchronization, context load, decision representation, trajectory comparison, deviation classification, human review and provenance locking as separable functions. Divergence between expected and observed trajectories is classified — missed, delayed, resistant, over-corrective — before any root-cause attribution is attempted. Classifying before explaining is deliberate: it keeps the deviation record independent of whatever theory is later applied to it.
Research-use boundary
The architecture is suitable for framework-validation research, inter-rater reliability design, qualitative analysis of decision states and—under controlled research conditions—exploratory AI/ML work. It is not currently validated for autonomous clinical recommendation or patient-facing decision support. Any future clinical deployment would require independent replication, intended-use definition, formal risk analysis and the appropriate regulatory pathway.
What remains to be established
- Independent design verification and a formal risk analysis.
- Inter-rater reliability. Every annotation in the current dataset comes from one annotator.
- Cross-user replication, including whether a reduced variable set is viable for non-expert annotators.
- Prospective clinical usefulness and safety under intended use.
- The appropriate regulatory pathway for any future intended use.