Batch 09 — Prediction Science & Hardening
The prediction engine measured rather than read: a learning-curve evaluation (40 seeded simulations per cycle-profile, mirroring the app's own error-feedback loop) established that the engine reaches its accuracy floor within 2-3 cycles for stable users and is honestly uncertain for genuinely variable ones — and surfaced defects in the learning loop that silently cost or falsified error samples, three measurable calibration gaps, and a set of scientific constants carrying no citation.
This batch fixes the loop, calibrates the engine against its own measurements, commits the simulation as a permanent guard, and writes down the evidence for every constant that encodes a claim about menstrual physiology.
Landed when: the learning-curve guard is committed and green at its final thresholds, every defect below is closed with a named regression test, and docs/architecture/SCIENCE.md exists with a verdict for each biology constant.
