Audited results

Synthetic smoke evidence

The table below comes from the successful MATLAB R2024b cross-language workflow at commit d475fbf. It uses one synthetic category, one seed, two GA generations, and five Pattern Search iterations. These deliberately small settings test the system; they are not thesis performance claims and must not be used to rank the model families.

Model	Variant	RMSE	MAE	MASE	sMAPE (%)
Seasonal naive	sales only	605.05	487.54	0.824	10.76
ARIMA	sales only	556.67	460.69	0.779	10.17
LSTM	sales only	2102.26	1993.53	3.369	54.51
EFS	sales only	634.29	511.09	0.864	11.37
EFS	target-date weather	710.56	577.22	0.975	12.98

All rows use the same 68 target dates and original sales scale. In this smoke run, target-date weather increased EFS RMSE by 12.03% relative to sales-only EFS (reported as a -12.03% improvement). This is a valid same-scale ablation result for this artifact, not evidence that weather is generally harmful.

Rule evidence

The sales-only smoke model learned 7 rules; the weather model learned 60. Both training summaries record zero change in upper membership parameters during the final uncertainty stage, confirming that only the type-2 lower scale/lag parameters were tuned then.

For example, the learned association “sales_lag_3 is Low → Medium” had support on 66 of 72 prepared test samples, mean activation 0.371, and maximum activation 0.796. This is an empirical association, not a causal statement or business prescription.

Sales onlyTarget-date weather

The complete lightweight snapshot, including CSV predictions, rules, and activations, is stored under data/precomputed/synthetic-smoke. Its provenance file identifies the exact workflow run, commit, and data SHA-256.

Publication status

No private-data accuracy result is published yet. The full eight-category, three-seed run must be performed locally with configs/benchmark.private.example.yaml; only its resulting manifest, hash, per-category table, macro summary, and weather ablation can support thesis conclusions.