How to read this
H6 — humans vs models
Literature / GlucoBench context
Data verification
1. Cohort
Who is in the cohort
Each slice is a unique person in one diabetes × CGM bucket. This is head-count, not accuracy.
How accurate people were
Each tick is one person's MAE (mg/dL). Lower is better. Colour is the diabetes × CGM bucket.
Accuracy on each task
A is generic, B is own data, C is mixed. One point per person who played that format. Lower MAE is better.
3. Primary hypotheses
4. Secondary hypotheses
Challenge the unknown / opposite trait
People in the current filter
| id | category | trait | runs | formats | challenge | MAE | generic | own | same | opposite |
|---|
Human vs AI on the same task
Age and diabetes type do not move the model, so this tab does not repeat H1–H4. Each tick is one person, and only if that person has both a human score and an AI score on that task. Order is Generic (A), Own (B), Mixed (C). Mixed is format C, not a blend of A and B.
The slider snaps to the middle (and to each end). Left is only the human, right is only the model, middle is both at 50%. The four category buttons act on the cluster and the deviation charts below. Use Expand on a chart to open it full screen.
Per-task MAE
Same person, same task. Blue is the human, purple is the model. We expect the AI group to sit tighter; a person may be better on generic or on own.
Same user: generic vs own
Each person who has a generic score and an own score is one point. Colour is the diabetes × CGM category. A line joins that person's human point to their AI point when both exist. Hollow marks are people with only one of those tasks, sat on the equal-MAE line so they stay visible. Below the diagonal = better on own data.
All-round deviations
Same 0% line as the sugar-sugar ending page. Rounds are folded into at most eight areas (four diabetes × CGM groups, human and AI). Human bands keep the category colours; AI bands use a second palette so they are not the same hue with a different dash. The edge is min–max; the thicker inner line is the median. Persistence, linear, and GluMind each get their own three panels.
Per-person traces
One person at a time, every reconstructed round on one screen (up to 12). Slate is the original CGM, blue is the human forecast, orange / green / purple are the AI models. Under the traces, this person's hidden-hour error is the same kind of overlapping area as the AI tab: one band per series (human and each model), min–max edge and a median inside.
This person's deviations
Horizontal 0% line, then percent error across the hidden hour. Generic (A), Own (B), Mixed (C). Each colour is one predictor — blue human, then a different colour per model — as a transparent area, not one line per round.