Three bugs and a framework swap later

Six weeks between the scaffold and a working pipeline. The architecture was right. The code had three bugs and a dependency that turned out to be wrong for this environment.

All five strategies now run. The numbers are in.

Bug one: the data schema

The NON_IID_SPEC in schema.py was missing ADL entries for two activity types (adl_eating and adl_toileting) for IL and SNF care types. CTGAN filled in the gaps by learning mid-range values — approximately 3.7 on a scale of 1–5 — which are clinically wrong for Independent Living facilities. IL patients are largely self-sufficient. A value of 3.7 implies significant dependency.

The downstream effect: the calibration threshold function was producing demand/supply ratios well above the expected 12% ceiling. The actual mismatch rate was 41.7%. That kind of skew in synthetic data would corrupt any result it touched.

The fix was adding clinically appropriate values — mean ≈ 1.0 for IL (low-acuity), mean ≈ 2.5 for SNF (moderate). The calibration function also had a hard upper bound of 5.0 on the binary search; facilities with high demand ratios hit the ceiling. Raised to 50.0.

Bug two: XGBoost 3.x

XGBoost 3.x removed BytesIO support from Booster.save_model() and Booster.load_model(). The Flower client was using that path for model serialisation. Replaced with save_raw("ubj") for serialisation and a NamedTemporaryFile + XGBClassifier.load_model() round-trip for deserialisation — which also correctly sets n_classes_ and other sklearn wrapper attributes that predict_proba needs.

The bigger problem: Flower needs Ray

Flower 1.29.0 requires Ray to run fl.simulation.start_simulation(). Ray installs at approximately 2 GB. For a single-machine research environment running 10 simulated clients on local data, that is roughly 1.98 GB more than the problem warrants.

The replacement is a manual FL round loop — _run_manual_simulation — that directly instantiates FedAcuityClient objects and steps through fit → aggregate → evaluate without any Ray or Flower server machinery. The public CLI interface is unchanged. The ResultsLogger output is unchanged. The tests that validate the client interface still pass.

The Flower dependency stays in the codebase for its client base class and protocol types. But the simulation runner is now self-contained.

The numbers

10 rounds, held-out evaluation:

Strategy	AUC-ROC
Local (no federation)	0.9749
Centralised oracle	0.9777
FedAvg	0.9643
FedProx (μ=0.1)	0.9643
Clustered FL (C1)	0.9828

CFL outperforms FedAvg by +1.85 percentage points overall. The IL cluster specifically improves from 0.912 under FedAvg to 0.983 under CFL — a 7.1pp improvement on the cluster that had the most clinically distinct patient profile. That is the dissertation’s core claim, demonstrated.

72/72 unit tests pass. The midsem evaluation is June 13.

---

Next: full 50-round simulation, data fidelity evaluation, DP epsilon sweep.