Additional controls for empirical model structure and experimental units

Source: Ivanovich et al., Frontiers in Microbiology 2025,16,1540664.
DOI:10.3389/fmicb.2025.1540664. Source measurements are reused with attribution;
no new experiments or covariates are represented as measured.

Run
python run_fairness_controls.py
python verify_fairness_controls.py
Both scripts write only to their own directory. They read ../empirical_mic
source extraction and previously verified XGBoost predictions. Use the scientific
environment documented in the main analysis; no tuning or data selection.

Suggested Methods paragraph
To separate reconstruction from model structure, we additionally fit a direct
Ridge model with all seven factorial contrasts of the two pressure levels,
flow category and inoculation status, including the three-way interaction.
The factorized Ridge baseline is correspondingly extended to pressure, flow
and their interaction; its microbial ratios remain estimated from training
controls only. Contrasts are effect-coded as -1/+1. Ridge retains alpha=1,
an unpenalized intercept and training-only feature and target standardization,
without hyperparameter search. Complete direct and factorized ordinary least
squares designs supply an unpenalized saturation check. These two designs span
the same eight treatment means, whereas Ridge penalties act in different
coordinates after target transformation. We report descriptive errors for both
32 coupons and16 treatment-biological-replicate means, keeping the two coupons
within each experimental unit together in every train/test split.

Suggested Results paragraph
Adding all direct treatment interactions changes Ridge coupon RMSE from
0.014280 to0.003820 mm/y; the corresponding factorized interaction model gives
0.003630 mm/y. Thus, the much larger improvement relative to direct main-effects
Ridge is associated with the interaction information carried by the calibrated
context factors, not a universal accuracy benefit of factorization. Unpenalized
saturated direct and factorized fits both reproduce the condition-mean
reference (RMSE0.003515 mm/y; maximum prediction difference2.78e-17 mm/y).
On the16 treatment-replicate means, these reference errors are0.003194 mm/y;
full-interaction Ridge gives0.003527 and0.003320 mm/y for direct and factorized
prediction, respectively. The complete control table and fold-level records
are supplied. Results describe biological-repeat prediction at the four
observed pressure-flow contexts.

Suggested abstract/conclusion refinement
The measured-data example demonstrates training-control calibration and
reconstruction under matched experimental conditions. Saturated direct and
factorized models are equivalent; differences for regularized or tree models
depend on representation and approximation. It does not externally validate
the frozen15-feature pipeline regressor or calibrate its unmeasuredHL gate.

Interpretation for submission
1. Keep cell means as a visible reference, not hidden in supplementary material.
2. Do not attribute the main-effects Ridge contrast solely to correction
   placement: the calibrated four-context factors add treatment interactions.
3. The full direct and factorized designs can represent all8 treatment means.
   With identical nominal alpha, penalties are still not invariant to target
   transformation; report both without claiming matched regularization.
4. No test target enters calibration or fitting. Training ratios depend on
   the training responses, which is legitimate calibration but additional
   learned information compared with a direct main-effects model.
5. Unit-mean errors address the experimental-unit reporting question; they
   do not enlarge the sample, establish uncertainty or demonstrate unobserved
   pressure/flow transfer. No inference based on32 independent cultures is made.
6. Artificial double-multiplication strategies are diagnostic failure cases.
   State their computational role and retain ordinary direct prediction as
   the primary practical comparator.

Files
coupon_predictions.csv: all9 retained model-strategy combinations on32coupons.
biological_unit_mean_predictions.csv: two coupons averaged within each of16units.
metrics.csv: full-precision RMSE,MAE,R2 on both reporting levels.
fold_metrics.csv: same metrics per held-out biological replicate fold.
training_factors.csv: all8 training-only context ratios across the two folds.
protocol.json andsummary.json: design, limitations, hashes, complete results.
publication_table_compact.csv: suggested4-row table,6-decimal rate unitsmm/y.
independent_audit.json: closed-form verification and nestedMSE decomposition.
