Completed main sweep: October 6

All 600 configurations × seeds 42/43 completed. Training allocations ended around 08:21 EDT; common-manager ranking completed at 08:25:27 EDT in ranking-98cf1e1c8cb0. The CSV tables in this folder contain the full 1,200-run results. Both native and log leaders from the provisional analysis remain unchanged. Their models, training inputs, labels, training/evaluation seasons and evaluation histories are described below. These remain development-season results rather than an untouched test.

The native leader is the joint flu MLP with Kinsa trained on real-tree-corrected synthetic histories, predicting finalized flu admissions and ED, evaluated with the newest two flu-admission weeks corrected and ED left reported. Two-seed mean composite 0.779856, admissions native 0.772371, admissions log 0.847984, forward ED 0.787072. The log leader is the joint flu MLP trained with artificial reporting errors plus recent finalized reconstruction, evaluated with two-week corrected admission histories: log 0.795342, native 0.791751, forward ED 0.894088. The forward fold trains 2022–25 and evaluates 2025–26; the retrospective fold trains 2022–24 plus 2025–26 and evaluates 2024–25. Finalized next-four-week labels are used in both. Scores below 1 beat the frozen ensemble; lower is better. Admissions seasons are weighted equally. Frozen retrospective support excludes ED; both-season raw ED is in pilot-raw-wis.csv.

All 150 matched recipes and both seeds are complete. Holding each recipe's architecture, training treatment, finalized labels, seasonal folds and correction rule fixed, independently retraining with COVID/RSV/both increases median corrected-view native composite by 4.38%/0.73%/0.65%; admissions native by 4.96%/1.08%/1.19%; admissions log by 1.83%/1.12%/2.42%; and forward ED by 5.06%/1.21%/1.94%. Thus other-pathogen inputs are not consistently beneficial. Detailed matched effects for all three evaluation-input views are in matched-input-effects.csv.

Both-season raw ED confirms a tradeoff between the selected models. The native leader's mean raw ED WIS in 2024–25 is 0.006035 states/DC and 0.005162 US, versus the log leader's 0.005722 and 0.004964. In 2025–26 the native leader scores 0.005070 and 0.004069, versus the log leader's 0.005713 and 0.005056. Lower is better. These raw scores use a common full-window support, unlike the ensemble-relative ED table.

Historical ILI results: within the 40 matched shared-architecture/input-set recipes, replacing no historical auxiliary task with flu-scaled ILI pretraining lowers the median corrected-view composite by 3.00%, native admissions by 3.35%, and log admissions by 2.14%. Own-source pretraining changes those by -1.06%, -0.90%, and +4.99%; flu-scaled joint learning by +0.41%, +0.32%, and +1.04%. Each comparison independently retrains the same shared forecaster on the same modern seasons and finalized flu labels, changing only the historical pre-August-2022 auxiliary mode/units. Scaling uses fit-only target quantiles and changes source-head identity as documented in the main protocol; it is not observed historical hospitalization data. All use the recipe's same modern input treatment and corrected Wednesday/holiday evaluation histories. This suggests scaled pretraining can help these architectures, but does not make them competitive with the leading existing MLP: the best corrected-view shared mixer composite is 0.915490 and shared MLP 1.043145, versus 0.779856. Architecture minima are comparisons of searched configurations, not controlled architecture effects.

One focused follow-up was submitted and started at 08:58:29 EDT: 24 configurations × two seeds, comparing contemporary ILINet/Kinsa and 2/4/8-week correction windows around both leaders. No additional refinement batch may launch. Main scientific ranking is complete; refinement scoring and analysis remain pending.