Act 5 · Trust
Replay and limits
Does the profile predict what happened next, and what can it not show?
Sample values. Synthetic study, not real patients.
Replay: re-fitted on records dated before 2024-12-26, then the rest revealed
The fit saw nothing dated after the cut (latest record in the fit: 2024-12-26). Computed inside the hospital zone; only the results crossed.
| Check | Quantity | Predicted | Actual | |
|---|---|---|---|---|
| Replay | Dropouts in the last third | 18.0 (13.0 to 25.0, 80%) Learned | 13 | inside |
| Replay | Interim visits attended in the last third | 212.0 (202.0 to 222.0, 80%) Learned | 215 | inside |
| Back-test | Months from the cut to full enrolment | 11.01 (9.59 to 12.68, 80%) Learned | 10.15 | inside |
Calibration
Not run for this build. Run python -m calibrate.
What one study can and cannot show
- One study cannot show that local knowledge makes a design better. Simulating twice shows what changes when local inputs replace global ones.
- One study cannot measure how much studies differ. Every prediction for a new study adds a between-study spread that clinicians estimated, and is labelled as such.
- The treatment difference is the brief's assumption. Trial OS does not predict whether a drug works.
- Site activation time has no local value and uses a global default.
- Calibration uses synthetic studies drawn from one truth. It tests the interval method, not real-world transfer.
- Theme agreement and extraction accuracy here are measured on synthetic text written beside the rules. Hand-labelled real documents are needed for a real number.