10 / Educational simulation · 2026
Adaptive dosing as an educational RL simulation
A synthetic control benchmark with a strict non-clinical boundary.
The work is deliberately framed as teaching software. Its synthetic parameters, reward and dynamics cannot support patient dosing or clinical claims.
Evidence register
20×20
state grid
Explicit tabular representation
4/4
tests
Environment and API contracts
held-out
seeds
Separate from training episodes
01 / Problem
An interesting controller can become harmful if a synthetic experiment is presented as a medical recommendation.
02 / Approach
Define a small MDP, compare learned and fixed synthetic policies on unseen seeds, and enforce the educational boundary in API and documentation.
03 / Outcome
A coherent end-to-end RL demonstration now includes evaluation, tests, CI, Docker and a strict simulation interface.
System trace
How the evidence is produced.
Repositioned and rebuilt the concept as bounded educational software with reproducible evaluation.
- 01Synthetic ODE → discrete state
- 02Q-learning → policy
- 03Held-out seeds → comparison
- 04Strict API → simulation only
Validation scope
Four tests pass; evaluation compares policies on held-out random seeds under the same synthetic environment.
Known limitation
No patient data, pharmacological calibration, clinical utility or safety validation. It must never recommend a real dose.
What is inspectable
- Non-clinical language is enforced.
- Training and evaluation seeds are separated.
- A fixed policy remains visible as baseline.
Next proof to add
- 01Add sensitivity analysis over dynamics.
- 02Report uncertainty across many seeds.
- 03Keep all outputs explicitly synthetic.
Main stack