Artificial reporting vintages: overnight research, 5 October 2026

This is an iterative development comparison using the strongest original B2 C1, C2 and C3 architectures. Each arm retrains a forecasting model. These seasons have already been used for development; none is an untouched prospective test.

Data and labels

When evaluating Wednesday reports in 2024–25, forecast models train on finalized trajectories and future labels from 2022–23, 2023–24 and 2025–26. Only archived/final reporting pairs from 2025–26 train the artificial reporting generator. This is a retrospective split.

When evaluating Wednesday reports in 2025–26, forecast models train on finalized trajectories and future labels from 2022–23, 2023–24 and 2024–25. Only archived/final reporting pairs from 2024–25 train the artificial reporting generator.

Every forecasting model learns the same finalized future values at t+1 through t+4. Artificial reporting errors alter training inputs only. Reference normalization and target scales are fitted on the permitted unperturbed historical episodes; stochastic errors change minibatch inputs without refitting those reference scales. Generator seasonal peaks use permitted training dates from the whole season, not only dates before each historical origin; these offline generator features are not separately supplied to the forecasting network. Evaluation always uses Wednesday archived numerical values under the original B2 source-availability schedule. Targets are available through the latest completed week; covariates keep their source-specific reporting delays and the Vermont exception. If an archive value is absent while the assumed schedule says it is available, the frozen final value is an explicit numerical proxy. This assumption does not establish operational availability. The generator's empirical training pairs never use such proxies. The arm using actual reports for the recent training season does use the same explicit final-value proxy policy as evaluation.

The full-availability assumption applies only where the panel has historical final truth. It does not invent absent historical series: RSV admission inputs are absent in 2022–23, so artificial RSV-admission errors can affect 2023–24 within the first-two-season treatment but cannot create 2022–23 observations. Input availability and finalized-label support remain unchanged.

Forecast architectures

C1 fits one multilayer perceptron per pathogen, shares spatial information with distance weights and has no covariates. C2 fits one multilayer perceptron per target without spatial sharing and uses inpatient, wastewater WVAL-like, Kinsa, ILINet, clinical laboratory and FluSurv covariates. C3 fits one multilayer perceptron per target with neighbor spatial sharing and Kinsa covariates. All use 12-week histories, fourth-root admission transforms, logit ED transforms and training-fitted B0 normalization. Epoch selection has a cap of 300 and patience 30. The first comparison removes the original B2 finality channel and 20% episode missingness regularizer.

First comparison

The original plan assigned seeds 42, 43 and 44 to each arm on each architecture. The final completed plan contains 78 runs: arms 3 and 5 below retain only seed 42 on each architecture; the other eight arms have all three seeds. The narrowed arms are explicitly single-seed pilots and are excluded from confirmed three-seed comparisons. Each run evaluates both seasonal folds.

1. Finalized inputs in all three training seasons, evaluated on archived reports.
2. Half-strength empirical reporting errors in all three training seasons.
3. Full-strength empirical reporting errors in all three training seasons.
4. Half-strength empirical errors in 2022–23 and 2023–24; finalized recent season.
5. Full-strength empirical errors in 2022–23 and 2023–24; finalized recent season.
6. Half-strength empirical errors in 2022–23 and 2023–24; actual recent-season reports.
7. Full-strength empirical errors in 2022–23 and 2023–24; actual recent-season reports.
8. Learned conditional reporting means plus centered empirical residual windows in 2022–23 and 2023–24; actual recent-season reports.
9. Learned conditional reporting means only in 2022–23 and 2023–24; actual recent-season reports.
10. Learned conditional reporting means plus centered empirical residual windows in all three training seasons.

Treatment scope follows the episode origin season. A context crossing a season boundary receives its origin season's treatment. The fitted empirical reporting-error library still excludes every reference week outside the named donor season.

The empirical method transports a signed log ratio on the original floor of 5% of the location/signal seasonal peak, at least one admission or 0.0001 for ED proportions. Half strength multiplies this log ratio by one half. Donor windows match location, final epidemic phase and calendar position.

The learned generator fits a ridge regression separately for each report age and signal. It predicts the signed log ratio from finalized normalized level, preceding-week growth, origin growth, annual sine/cosine, proximity to Christmas and a location indicator shrunk toward the shared regression. The ridge penalty is 20; the intercept is effectively unpenalized. Fewer than 30 observed pairs leave that age/signal's conditional mean at zero. Finalized epidemic features are legitimate inputs to this training-data generator, but would not be available to an operational nowcaster. The forecasting model never receives these generator features separately.

For residual draws, the generator subtracts its fitted conditional mean from observed reporting errors and transports a matched complete residual window. Those residuals are centered in-sample: shrinkage limits overfit, but the residual variance may be conservative. Predicted log means are capped at a factor of four; ED inputs remain between zero and one. National covariates remain shared across locations. Missing donor archive cells have zero empirical residual, while a learned conditional mean may be predicted from other supported pairs. This is an explicit modeling assumption.

Early stopping excludes hidden validation weeks from the generator's fits and peak scales. No evaluation-season reporting error or label enters the fitting library. Recent-season actual training inputs preserve the episode's available cells. Finality indicators and artificial missingness regularization are omitted for the first comparison to match the weekend architectures. This is a choice of training regularization, distinct from the assumed evaluation availability.

Score

Lower is better. Each location's summed forecast weighted interval score (WIS) is divided by the hub ensemble's WIS on identical frozen tasks. States/DC share 80% of the weight equally and US receives 20%. Admission targets have twice the weight of ED proportions. Seasons count equally; report means and variation across three seeds. A value below one beats the hub ensemble under this objective. Report the two evaluation seasons separately as well as their equal average. Frozen benchmark support differs: 2024–25 scores flu and COVID hospital admissions only; 2025–26 scores all six targets. Models still learn finalized future labels for all six targets in both folds. A secondary diagnostic restricts both seasons to flu and COVID admissions, equally weighted, using the same fitted models and predictions. It does not replace or modify the official objective.

Code and manager commands

Private local source: /tmp/chromantis-vintage-overnight-20261005. Private remote project: /proj/jlessler/projects/tapestry-all/tapestry-vintage-overnight-20261005. The shared source tree is not edited. Inputs and the environment link to the existing reporting-augmentation project; experiment names are unique.

The plan wrapper calls the common manager's plan command and saves the exact expanded invocation in each experiment's manager-plan.txt.

cd /proj/jlessler/projects/tapestry-all/tapestry-vintage-overnight-20261005
PYTHONPATH=src .venv/bin/python scripts/plan_vintage_overnight.py -e vintage-overnight-smoke-20261005 --smoke
LANES=3 GPUS=1 sbatch --job-name=vintage-overnight-smoke-20261005 --array=0 --time=00:20:00 scripts/jlessler.sbatch vintage-overnight-smoke-20261005
PYTHONPATH=src .venv/bin/python -m chromantis.experiment.planner status -e vintage-overnight-smoke-20261005
PYTHONPATH=src .venv/bin/python -m chromantis.experiment.planner rank -e vintage-overnight-smoke-20261005 --no-plots

The two-epoch execution pilot (job 3800115) checks the learned generator's path, not forecasting efficacy. Research results will be added as the full comparisons finish.

Donor-model diagnostic before forecast results

The conditional ridge reporting generator was evaluated by withholding each of four contiguous blocks of donor origins in turn. This predicts archived/final log reporting errors within one donor season, not future disease forecasts. Lower root mean squared log error is better. For current-week flu admissions from 2024–25, the conditional model scored 0.1227, compared with 0.1373 for a shared donor-training mean; using 2025–26 pairs, the scores were 0.1255 and 0.1379. This is evidence for structured reporting bias, not evidence that the forecast model improves. The diagnostic ridge is retrained excluding each block.

Current-week ED has only about 17% archive support among the 2024–25 donor windows, concentrated in one of the four blocks. The remaining blocks cannot train its current-week conditional model. A ratio of one for this case reflects that lack of training support, not demonstrated equivalence of the two models. Age-three COVID ED log-error RMS is 0.0555 in the 2024–25 donor season but 0.1823 in 2025–26. This large shift limits confidence in transporting ED errors between seasons. The reporting-generator-blocked-validation.csv table includes exact support fractions, errors and season definitions.

First research launch: job 3800158, four L40 allocations, four processes each, six-hour limit, 30 configurations and 90 seed runs (180 outer folds).

PYTHONPATH=src .venv/bin/python scripts/plan_vintage_overnight.py -e vintage-overnight-round1-20261005
LANES=4 GPUS=4 sbatch --job-name=vintage-overnight-round1-20261005 --array=0-3 --time=06:00:00 scripts/jlessler.sbatch vintage-overnight-round1-20261005
PYTHONPATH=src .venv/bin/python -m chromantis.experiment.planner status -e vintage-overnight-round1-20261005
PYTHONPATH=src .venv/bin/python -m chromantis.experiment.planner rank -e vintage-overnight-round1-20261005 --allow-incomplete --no-plots

The outer empirical library contains 34 origins from 23 November 2024 to 19 July 2025 for the 2024–25 donor season, and 48 origins from 2 August 2025 to 25 July 2026 for the 2025–26 donor season. Calendar matching therefore cannot supply observed early-autumn 2024 reporting patterns. The preexisting metadata field target_evidence_fraction describes finalized-truth support, not actual archived/final pair coverage; use available_pairs_fraction from the diagnostic CSV for archival evidence.

Admissions-only follow-up

The donor archive diagnostic and a separate H100 development run both implicated ED revision transport. The H100 evidence is a design clue, not pooled with L40 results. The second L40 comparison contains four arms on each C1/C2/C3 model, again with seeds 42, 43 and 44:

• Keep 2022–23 and 2023–24 inputs finalized; use actual Wednesday reports for the recent permitted training season. This isolates the effect of adding artificial vintages to the first two seasons.
• Add half-strength empirical reporting errors to admissions in 2022–23 and 2023–24; retain actual reports for the recent permitted training season.
• Add half-strength learned conditional reporting means plus empirical residuals to admissions in 2022–23 and 2023–24; retain actual recent-season reports.
• Add half-strength empirical errors to admissions in all three training seasons; keep all ED targets and covariates finalized in training. This arm tests whether admission gains survive without introducing ED/covariate errors.

Here admissions-only means the three NHSN targets. Early-season ED targets and all covariates remain finalized. For the first three arms, actual recent-season inputs still include every signal on the documented source schedule. The fourth arm instead uses synthetic admission inputs and finalized ED/covariate inputs in all three seasons. The recent season is 2025–26 when forecasting the held-out 2024–25 season, and 2024–25 when forecasting 2025–26. Prediction labels, evaluation reports, reporting lags and final-value archive proxies remain exactly as specified above.

The second code snapshot also makes a missing-adjacent-week growth feature zero rather than interpreting missingness as a large epidemic change, and caches unchanging conditional means for speed. The first snapshot is retained inside its experiment and is not modified. These are separately identified research runs; any comparison between their learned generators includes the stated growth-feature change.

Array 3804709 requests four L40 GPUs, four processes per GPU, with an allocation limit of 135 minutes and an 08:25 EDT scheduling deadline. It is queued behind the first comparison. Slurm's status, not this protocol, determines actual starts.

PYTHONPATH=src .venv/bin/python scripts/plan_vintage_overnight.py -e vintage-overnight-admissions-focused-20261005 --admissions
LANES=4 GPUS=4 sbatch --job-name=vintage-overnight-admissions-focused-20261005 --array=0-3 --time=02:15:00 --deadline=2026-10-05T08:25:00 scripts/jlessler.sbatch vintage-overnight-admissions-focused-20261005
PYTHONPATH=src .venv/bin/python -m chromantis.experiment.planner status -e vintage-overnight-admissions-focused-20261005
PYTHONPATH=src .venv/bin/python -m chromantis.experiment.planner rank -e vintage-overnight-admissions-focused-20261005 --allow-incomplete --no-plots

At 02:00 EDT the first completed three-seed arm was the C1 pathogen MLP with learned reporting means plus empirical residuals in all three training seasons. Trained on 2022–23, 2023–24 and 2025–26, with reporting errors learned only from 2025–26 and finalized future labels, it scored 0.88178 on 2024–25 Wednesday reports. Trained on 2022–23, 2023–24 and 2024–25, with errors learned only from 2024–25, it scored 0.99093 on 2025–26 Wednesday reports. Lower relative WIS is better. Its matched all-finalized-input training control was not yet complete, so these early numbers do not establish a benefit from artificial vintaging.

Budget adjustment at 02:20 EDT

Observed target-partition seed runs take roughly 60–70 minutes on four concurrent processes per L40, while C1 takes roughly 30–40 minutes. To leave time for the admissions follow-up before 08:30 EDT, twelve unstarted replications were removed from the live dispatcher queue: seeds 43 and 44 for full-strength empirical errors in all three seasons, and for full-strength early-two errors with recent-season finalized inputs, on each of C1/C2/C3. Their seed-42 pilots remain. All matched controls, half-strength arms, recent-actual-report arms and learned-generator arms retain three seeds. This change uses the earlier weekend finding that full-strength blanket perturbations hurt the forward season; it is not hidden selection of favorable overnight seed results.

The active first-batch budget is now 78 seed runs. The initial manager plan recorded all 90. deferred-seeds.json records the initial temporary deferral. The subsequent vintage_overnight_narrow.py helper removed only these twelve unstarted entries from both jobs.csv and the dispatch state under its lock. narrowed-plan.json records the exact change and design.csv lists each arm's planned seeds. Ordinary manager resubmission now preserves the 78-run plan. No running fit was cancelled. Partial rankings must not present these one-seed pilots as equivalent to completed three-seed comparisons.

The originally queued follow-up array 3801336 was cancelled before any fit started. Its half-clean learned arm was replaced by the all-three-season admissions-only empirical arm, keeping the 36-run budget unchanged. The first matched L40 C2 seed-42 results motivated this change: with training on 2022–23, 2023–24 and 2024–25, errors learned only from 2024–25 and finalized future labels, the model evaluated on 2025–26 Wednesday reports scored 0.96577 after all-finalized input training, 0.97333 with half-strength errors on all signals in all three seasons, and 1.02152 with half-strength early-two errors plus actual recent-season reports. Lower is better. For the all-three half-strength arm, flu admissions improved from 0.93842 to 0.87188, while COVID ED worsened from 0.90394 to 0.98138 and RSV ED worsened from 0.82830 to 1.05088. These are single-seed development clues, not three-seed confirmations.

To reproduce the narrowed scheduling after a new first-round plan, apply the following helper before launching. It changes seed lists only; no configuration, task ID, cost estimate, fitted result or pinned source is changed.

PYTHONPATH=src .venv/bin/python scripts/vintage_overnight_narrow.py -e vintage-overnight-round1-20261005

Existing-experiment resumes already use the narrowed plan automatically. An exact new reproduction of the original generator must set PYTHONPATH=reproduction-source-v1/src while planning under a fresh name. That independent, checksum-verified source copy includes the notifier required by the planner. Using an experiment's own snapshot as a re-planning source is unsafe because planning replaces the destination snapshot; the bare snapshot also lacks the expected notifier path. No such re-plan was executed. The private working source now contains the documented second-version growth-feature change. The normal Slurm launcher always uses each existing experiment's pinned source. See manager-commands.md for the complete safe reproduction commands.

At 03:10 EDT, completed seed runtimes averaged 31.4 minutes for C1 (four runs), 58.1 minutes for C2 (ten runs) and 59.0 minutes for C3 (six runs). The pending follow-up limit was increased from 120 to 135 minutes to allow for the last few fits. With its unchanged 08:25 EDT scheduling deadline, it must start by 06:10 EDT. Actual scheduling and completeness will be reported, not inferred from this estimate.

At approximately 03:20 EDT, the dispatcher briefly held unstarted noncontrol entries until the remaining five matched-control seeds were claimed. It then automatically restored all 35 held fits. No planned seed, scenario or active fit was removed. control-priority.json records this scheduling change. All three C1 controls and the remaining C2/C3 controls were running when the hold ended.

First matched three-seed result

The C1 pathogen MLP, with distance-based spatial sharing and no covariates, trained on finalized 2022–23, 2023–24 and 2024–25 inputs and finalized future labels, scored 0.93984 ± 0.01230 on the six-target 2025–26 Wednesday-report benchmark. Retraining that same architecture with learned conditional reporting means plus empirical residuals from 2024–25 applied to all three training seasons scored 0.99093 ± 0.08785, a 5.44% worsening in the mean. Lower is better; the spread is standard deviation across seeds 42, 43 and 44. Two of the three paired seeds worsened.

For the retrospective 2024–25 benchmark, which scores flu and COVID admissions only, C1 trained on 2022–23, 2023–24 and 2025–26 improved from 0.93467 to 0.88178 when the learned errors came only from 2025–26. All future labels stayed finalized, and both models received the same Wednesday reports, source lags and explicit absent-archive final proxies at evaluation. The secondary forward flu/COVID-admission score also worsened, from 0.95531 to 0.97765. Thus C1's forward deterioration is not solely an artifact of adding ED/RSV targets to the forward benchmark. Learned blanket vintaging has not solved the forward problem.

Per-seed and per-season results and the graph are in vintage-overnight-round1-20261005/. Only comparisons with three completed seeds in both the treatment and matching control receive a plotted percentage change.

The location-specific archive audit confirms that the national series itself has only seven current-week ED reporting origins in 2024–25, versus 42–43 in 2025–26. National admissions have 33–34 versus 46–47 origins. Counts refer to observed report/final pairs, not final-truth support. The full-season audit contains 52 and 53 origins respectively; the fitted donor windows use the subset documented above. Missouri has no current-week ED pairs in 2024–25, and Wyoming has none for flu ED. These gaps are not zero-error evidence. The conditional model can extrapolate a mean to missing archive cells, while empirical residuals are zero there; this is a material modeling assumption, especially for ED. See archive-support-by-location.csv.

Effective perturbation audit

transport-error-magnitudes.csv and its graph measure the actual changes made to available training inputs, after nonnegative and proportion clipping. The audit uses the first-round pinned generator, seed 20261005, and five draws per origin. It compares current-week and three-week-old inputs in each training season. The reporting model is fitted only to the permitted donor season; this diagnostic does not fit a forecasting model or score predictions. Future labels are never changed. The plotted subset concerns 2022–23 inputs and errors learned from 2024–25, as used for the forward forecast fold. RSV admissions are absent in 2022–23 and omitted rather than filled with synthetic observations.

At full strength, empirical donor windows leave 98% of available current-week 2022–23 RSV ED cells unchanged (effective log-error RMS 0.019), whereas the learned conditional mean plus residual generator leaves 5% unchanged (RMS 0.064). For flu ED, the unchanged fraction is 91% versus 1% (RMS 0.012 versus 0.022); for COVID ED, 86% versus approximately zero (RMS 0.057 versus 0.073). Admissions perturbation RMS is similar: flu 0.122 versus 0.120 and COVID 0.203 versus 0.207. Thus the learned generator is also an extrapolation into absent archive cells, with a much denser ED perturbation pattern. This is a plausible mechanism to investigate, not proof that it caused any forecast score change. The admissions-only follow-up avoids synthetic ED perturbations.

Reproduce the input audit with:

PYTHONPATH=data/experiments/vintage-overnight-round1-20261005/code/src .venv/bin/python scripts/vintage_overnight_transport_diagnostic.py

All three seeds of each matched finalized-input training comparison are now complete on L40. The following models use finalized training inputs and unchanged finalized t+1 through t+4 prediction labels. For evaluation on 2025–26 they train on 2022–23, 2023–24 and 2024–25. For retrospective evaluation on 2024–25 they train on 2022–23, 2023–24 and 2025–26. Every evaluation uses Wednesday archived values, B2 source lags and the documented final-value proxies. Scores are mean relative WIS ± standard deviation over seeds 42, 43 and 44; lower is better. The two columns have different target support and should not be compared as a common six-target metric.

Model trained on finalized inputs	Retrospective 2024–25: flu/COVID admissions	Forward 2025–26: six targets
C1: pathogen MLP, distance sharing, no covariates	0.93467 ± 0.03536	0.93984 ± 0.01230
C2: target MLP, no spatial sharing, six covariate groups defined above	0.93397 ± 0.04085	0.98733 ± 0.01868
C3: target MLP, neighbor sharing, Kinsa	0.97920 ± 0.06795	0.97501 ± 0.05111

The focused scientific checks in scripts/vintage_overnight_scientific_checks.py passed for both held-out seasons and empirical/learned generators. They check that donor origins belong only to the permitted reporting-error season, finalized future labels are identical before and after augmentation, input episodes are not mutated, and admissions-only draws leave ED values and covariates exactly unchanged in 2022–23/2023–24. The recent-season actual-report treatment is intentionally allowed to change all available input signals. Run these checks with PYTHONPATH=src .venv/bin/python scripts/vintage_overnight_scientific_checks.py.

C2 matched three-seed empirical comparisons

C2 is the target MLP without spatial sharing, using inpatient, wastewater, Kinsa, ILINet, clinical-lab and FluSurv covariates. For the forward benchmark it was retrained on 2022–23, 2023–24 and 2024–25 with the following training-input treatments, using only 2024–25 to construct artificial reporting errors. All future prediction labels remained finalized. Evaluation inputs were identical 2025–26 Wednesday reports with B2 source lags and documented absent-archive final-value proxies. Mean six-target relative WIS is lower when better.

Training-input treatment	Mean forward score	Change versus finalized training	Seeds improving out of three
Finalized inputs in all three seasons	0.98733	—	—
Half-strength empirical errors in all three seasons	0.98778	+0.05%	1
Half-strength errors in 2022–23/2023–24; 2024–25 finalized	0.97395	−1.36%	1
Half-strength errors in 2022–23/2023–24; actual 2024–25 reports	0.97874	−0.87%	2

These small forward gains are variable across seeds. The secondary forward flu/COVID-admission score worsens from 0.94281 after finalized-input training to 0.94925, 0.95130 and 0.97991 for the three respective treatments. In the all-three-season half-strength arm, mean COVID admissions improve from 0.9712 to 0.9343 and RSV admissions from 1.1863 to 1.0822, but flu admissions worsen from 0.9144 to 0.9642 and all three ED targets worsen. Thus the admissions-only follow-up remains a hypothesis, not a confirmed fix derived from the seed-42 result.

For retrospective 2024–25 evaluation, C2 instead trains on 2022–23, 2023–24 and 2025–26, constructs errors only from 2025–26 and learns the same finalized future labels. The flu/COVID-admission score is 0.93397 for finalized-input training, 1.00346 for all-three-season half errors, 0.99514 for early-two half errors with recent finalized inputs, and 0.84579 for early-two half errors with actual recent reports. The last improves all three seeds by an aggregate 9.44%. The recent-actual-only control in the follow-up is needed to determine how much of this comes from real recent-season inputs rather than artificial early-season vintaging.

The admissions-focused report explicitly imports the matched finalized-input training controls from the first round on the same L40 hardware:

PYTHONPATH=src .venv/bin/python scripts/vintage_overnight_report.py -e vintage-overnight-admissions-focused-20261005 --reference-experiment vintage-overnight-round1-20261005

Its primary percentage changes compare every arm to those finalized-input controls. A separate column compares the early-two synthetic arms to the follow-up's finalized-early-two/actual-recent-input control; this isolates the additional effect of artificial early-season vintaging. That second contrast is intentionally absent for the all-three-season synthetic arm. Seed-paired absolute differences and the number of improving seeds accompany means; they are only reported after three matched seeds complete. Three seeds quantify training variability, not independent prospective evaluation seasons.

Nonlinear reporting-generator screen

A second learned reporting model was screened before forecast retraining: shallow histogram gradient boosting with 80 iterations, eight leaves, minimum 50 cells per leaf, learning rate 0.05 and L2 penalty 20. It predicts signed log report/final error from normalized final level, two growth features, annual sine/cosine and holiday proximity. It was compared with ridge and with ridge plus boosted residuals, which retains ridge's location effects. Each model was fitted to three contiguous-origin blocks within the permitted donor season and predicted the fourth block; neither forecast evaluation season supplied fitting pairs. These are reporting-error prediction results, not forecast scores.

For current-week admissions from donor 2024–25, ridge RMSE is 0.1227/0.1660/0.1762 for flu/COVID/RSV; standalone boosting gives 0.1325/0.1760/0.1884, and ridge plus boosting gives 0.1256/0.1669/0.1780. For donor 2025–26, ridge gives 0.1256/0.1550/0.1781, standalone boosting 0.1321/0.1647/0.1822, and ridge plus boosting 0.1351/0.1576/0.1783. Lower RMSE is better. Neither nonlinear candidate improves any of the six current-week admissions comparisons, so no forecasting GPU batch was assigned to this candidate. This does not rule out other nonlinear models, nor does it contradict a report-to-final nowcaster: the direction and prediction task here are final-to-report error generation.

The four-block comparison cannot evaluate temporal generalization of current-week 2024–25 ED errors because their archive pairs are concentrated in one block; identical ED entries are not evidence of equivalence. Full results are in nonlinear-reporting-blocked-validation.csv. Reproduce with:

OMP_NUM_THREADS=2 OPENBLAS_NUM_THREADS=2 PYTHONPATH=src .venv/bin/python scripts/vintage_overnight_nonlinear_diagnostics.py

At 05:20 EDT, the first admissions-focused allocation started after idle first-round owner 3800158_2 was released. A second idle first-round owner, 3800158_3, was released shortly afterward. No active fit was cancelled. An attempted extension of the first follow-up owner's running time limit to 08:24 EDT was denied by Slurm (Access/permission denied); the original 135-minute limits remain. The contemplated balanced-pilot scheduling hold was therefore abandoned before any seed state changed, to retain the common manager's existing cost and load balancing. The 36-run plan is unchanged.

Completed first-round outcome

All 78 planned seed runs finished by approximately 05:29 EDT with both folds scored and no failures. Six deliberately narrowed full-strength comparisons retain only their seed-42 pilots; all other configurations have seeds 42/43/44. C1, the pathogen MLP with distance sharing and no covariates, remains the best forward mean among these completed runs when trained on finalized inputs: 0.93984. Every three-seed C1 artificial-vintaging treatment worsens its mean; the closest is half-strength empirical errors in 2022–23/2023–24 while retaining finalized 2024–25 inputs, at 0.94604 (+0.66%). Training labels stay finalized; evaluation is the same six-target 2025–26 Wednesday-report benchmark with source lags and absent-archive final-value proxies.

The clearest within-architecture early-season candidate is C3, a target MLP with neighbor sharing and Kinsa. Trained on 2022–23/2023–24/2024–25, with learned conditional mean-only reporting errors from 2024–25 applied to the first two seasons and actual reports in 2024–25, it scores 0.94960 versus 0.97501 after all-finalized input training (−2.61%). Both use unchanged finalized future labels and identical 2025–26 evaluation inputs. Two of three seeds improve; the paired absolute score difference is −0.02541 with SD 0.03055 across seeds. This does not beat the stronger C1 finalized-input model. Adding empirical residuals to that C3 learned mean gives 0.95873 (−1.67%, two of three seeds improve).

For the retrospective fold, the corresponding C3 models train on 2022–23/2023–24/2025–26, learn reporting errors only from 2025–26, and receive 2024–25 Wednesday inputs. The flu/COVID-admission score improves from 0.97920 with finalized training inputs to 0.83555 with early-two learned means and actual recent reports, or 0.82555 with learned means plus residuals; all three seeds improve in both cases. Final labels are unchanged. The follow-up's actual-recent-only control is still needed to attribute these benefits to artificial early-season vintaging rather than real recent-season reports.

All four admissions-focused allocations started between 05:20 and 05:29 EDT. Owners 0/1 retain 135-minute limits; pending owners 2/3 were extended successfully to 150 minutes before starting. Thus the last allocations end around 07:58 EDT, within the 08:30 research deadline. The complete common-manager ranking and matched-comparison figures are in vintage-overnight-round1-20261005/.

First admissions-focused three-seed results

For C1, trained on 2022–23/2023–24/2024–25 with finalized future labels and reporting errors learned only from 2024–25, half-strength empirical errors restricted to admissions in all three seasons score 0.94444 ± 0.03289 on the six-target 2025–26 Wednesday-report benchmark. ED inputs remain finalized during training. This is better than the matched all-signal half-strength treatment (0.95561), but still 0.49% worse than finalized-input training (0.93984); only one of three seeds improves against that finalized-input control. Evaluation inputs, source lags and absent-archive final proxies are identical for every arm.

C1 with learned half-strength admission errors in 2022–23/2023–24 and actual 2024–25 reports scores 0.95857 ± 0.03448 (+1.99%; one of three seeds improves). That treatment changes both strength and affected signals relative to the first-round full-strength learned generator, so their difference cannot be attributed solely to restricting perturbations to admissions. The actual-recent-only control was not complete at this interim report.

For retrospective 2024–25 flu/COVID-admission scoring, training instead uses 2022–23/2023–24/2025–26 and errors only from 2025–26, with unchanged finalized future labels. The two respective C1 treatments score 0.87064 and 0.86100 versus 0.93467 after finalized-input training. The latter improves all three seeds. The forward and retrospective columns have different benchmark target support.

Final small dilution comparison

After the first two C1 admissions-only treatments failed to beat finalized-input training on the forward fold, one additional three-seed C1 comparison was queued: half of training episodes remain clean and half receive half-strength empirical reporting errors across all three training seasons. The same strong C1 architecture learns unchanged finalized future labels. For 2025–26 evaluation it trains on 2022–23/2023–24/2024–25, with errors only from 2024–25; for retrospective 2024–25 evaluation it trains on 2022–23/2023–24/2025–26, with errors only from 2025–26. Both use unchanged Wednesday-report evaluation inputs, B2 source lags and the explicit final-value proxies. This tests the clean/noisy mixture that was most consistent on other fully replicated weekend backbones, for which C1 had no completed comparison. It is not an architecture or missingness-mask sweep.

Job 3818337 uses one L40 with three concurrent fits and a 45-minute limit. Its 08:25 EDT deadline requires a start by 07:40. Admissions-focused completion has priority; this queued job will be cancelled if that work needs the capacity. Plan and manager commands, in the private remote project:

PYTHONPATH=src .venv/bin/python scripts/plan_vintage_overnight.py -e vintage-overnight-c1-diluted-20261005 --diluted
LANES=3 GPUS=1 sbatch --job-name=vintage-overnight-c1-diluted-20261005 --array=0 --time=00:45:00 --deadline=2026-10-05T08:25:00 scripts/jlessler.sbatch vintage-overnight-c1-diluted-20261005
PYTHONPATH=src .venv/bin/python -m chromantis.experiment.planner status -e vintage-overnight-c1-diluted-20261005
PYTHONPATH=src .venv/bin/python -m chromantis.experiment.planner rank -e vintage-overnight-c1-diluted-20261005 --no-plots
PYTHONPATH=src .venv/bin/python scripts/vintage_overnight_report.py -e vintage-overnight-c1-diluted-20261005 --reference-experiment vintage-overnight-round1-20261005

reporting-metadata-clarification.json distinguishes final-truth support from observed archived/final-pair support for every target and the C2 covariates in each permitted donor season. The older inherited target_evidence_fraction field means final-truth support only. Its inherited unsupported_cells=unchanged description is correct for empirical errors but not for learned conditional means, which can extrapolate into cells without archived pairs. This sidecar clarifies the completed pinned runs without altering their fitted inputs, checkpoints or scores. The integration patch now labels those quantities and the learned extrapolation explicitly in newly generated metadata. The metadata addition changes no numerical computation. Reproduce the audit with PYTHONPATH=src .venv/bin/python scripts/vintage_overnight_metadata.py.

The nonlinear reporting-model diagnostic uses scikit-learn 1.7.2, already available in the shared Longleaf virtual environment. The ridge generator uses NumPy. The paired-forward plot can be regenerated after the first-round report with PYTHONPATH=src .venv/bin/python scripts/vintage_overnight_paired_plot.py.

Allocation handoff at 06:52 EDT

One C3 learned half-strength admissions fit, seed 44, started at 06:37 on an allocation ending at 07:35; its matched seed had taken 63.5 minutes. After eight minutes the fit subprocess was deliberately stopped and its failed attempt preserved. Two earlier-ending queue-manager processes were paused for about seven minutes while their independent training children continued. The same seed and unchanged numerical configuration restarted on owner 3804709_2 at about 06:52, with an allocation ending at 07:58. Both managers resumed and pending work was restored. The guarded controller had an eight-minute fail-safe and a finally block to resume managers. Its record is data/experiments/vintage-overnight-admissions-focused-20261005/reroute-record.json. This was an allocation-timing intervention, not a selection based on fit quality. The new attempt is included once, through the ordinary common scorer.

Forward C3 score components before attribution control completes

The C3 target MLP with neighbor sharing and Kinsa, trained on 2022–23/2023–24/2024–25 with learned conditional mean errors in the first two seasons and actual recent-season reports, has forward six-target mean WIS ratio 0.94960 versus 0.97501 after finalized-input training. Both learn unchanged finalized future labels and evaluate the same 2025–26 Wednesday reports with the stated schedules and proxies. Its flu-admission target ratio changes from 0.9430 to 0.8747 and RSV admissions from 1.0906 to 0.9758. Their normalized underprediction penalties fall from 0.4144 to 0.3327 and 0.7442 to 0.5970. Nominal 90% interval coverage rises from 68.4% to 78.5% for flu admissions and 67.7% to 75.3% for RSV admissions; neither reaches nominal coverage. COVID admissions worsen from 0.9425 to 1.0139 as overprediction rises. These are three-seed target diagnostics, not independent confirmation or proof that artificial errors caused the gain. The actual-recent-only reference is needed to isolate that contribution.

The generator threshold of 30 supported cells does not mean 30 independent reporting dates: cells include locations. Donor-origin counts and archive-pair support, including their seasonal concentration, are reported separately. Three seed replicates quantify optimization variability; they do not account for all epidemiological or reporting-process uncertainty.

For current-week ED errors in the 2024–25 donor season, all supported dates fall into one of the four contiguous validation blocks. When that block is held out, the fitting blocks have no observed current-week ED error pairs. The shared-mean and ridge diagnostics therefore both fall back to zero and their equal RMSE is not evidence that the reporting processes are equivalent or that the learned model was meaningfully validated for those ED cells. The admission diagnostics have evidence across blocks and are more informative.

Numerical proxy fractions in scheduled inputs

scheduled-input-proxy-audit.csv counts the cells actually used by the scheduled input protocol, before model fitting, by season, signal and report age. Unlike the donor-error library audit, it includes all eligible origins: 51 in 2024–25 and 52 in 2025–26. At age zero, 2024–25 scheduled admission inputs use frozen-final proxies in 33.1%–36.7% of available cells, and ED inputs in 88.5%–88.7%. For 2025–26 the respective fractions are 11.5%–13.5% and 20.5%–22.6%. Denominators exclude absent historical final truth and cells hidden by source-specific lags. The same source-season mixture applies to actual-recent training inputs and evaluation inputs. Consequently, “actual recent reports” always means archived numerical values where present plus the explicitly assumed frozen-final proxies; it does not imply a fully observed archive. Labels are unchanged finalized future values. Reproduce with PYTHONPATH=src .venv/bin/python scripts/vintage_overnight_proxy_audit.py; this audits inputs and fits no model.

C2 attribution with three reference seeds completed

The C2 target MLP without spatial sharing, using inpatient, wastewater, Kinsa, ILINet, clinical labs and FluSurv, scores 0.96640 ± 0.02106 on forward 2025–26 when trained on finalized 2022–23/2023–24 inputs and actual-or-proxy 2024–25 reports. All labels remain finalized t+1 through t+4 and evaluation inputs are identical Wednesday archived/proxy values. This improves the all-finalized-input reference, 0.98733, by 2.12%; all three paired seeds improve.

Adding artificial errors to the first two seasons does not improve that three-seed forward mean in the original comparison: half-strength empirical errors worsen it by 1.28%, full-strength errors by 7.62%, learned conditional means by 2.75%, and learned means plus residuals by 11.46%. The half-strength arm improves two of three seeds but its worse seed reverses the mean gain. For retrospective 2024–25 scoring, the actual-recent-only reference instead uses 2025–26 reports and scores 0.90910 versus 0.93397 after finalized-input training; adding half-strength empirical errors improves that retrospective mean by a further 6.96%. The forward six-target and retrospective two-admission objectives differ. These results support distinguishing use of actual recent reports from the contribution of artificial first-two-season errors.

actual-recent-attribution.csv, .md, .png and .pdf explicitly compare those matched inputs; three-seed references are required before an architecture appears. Regenerate after the principal export with PYTHONPATH=src .venv/bin/python scripts/vintage_overnight_attribution.py.

The admissions-only arms restrict which inputs are perturbed; they retain the same joint donor-window eligibility rule as the original generator. That rule requires some archived evidence within the latest four context weeks for every target signal, excluding pre-archive/onboarding origins. Thus admission-only perturbations still use the 34 eligible 2024–25 donor origins or 48 eligible 2025–26 origins, rather than all archived admission pairs from those seasons. The restriction does not recover missing early-season donor dates. Calendar and phase extrapolation beyond those supported dates remains a model assumption.

C1 attribution with three reference seeds completed

For C1, the pathogen MLP with distance sharing and no covariates, finalized 2022–23/2023–24 inputs plus actual-or-proxy 2024–25 reports yield forward 2025–26 mean 0.98968 ± 0.07812, compared with 0.93984 ± 0.01230 after finalized-input training in all three seasons. Actual recent inputs worsen this mean by 5.30%. Adding learned half-strength admission errors in the first two seasons improves the actual-recent reference by 3.14%, to 0.95857; two of three paired seeds improve. Nevertheless, this remains 1.99% worse than all-finalized-input training. Half-strength empirical errors in all early-season signals likewise improve the actual-recent reference by 1.81%, but do not beat all-finalized-input training. These are partial rescues of a worse C1 training setup, not a better overall C1 result. Labels remain finalized t+1 through t+4 and every arm evaluates identical Wednesday archived/proxy inputs.

Retrospective C1 actual-recent-only training uses 2025–26 reporting inputs and evaluates 2024–25 flu/COVID admissions. It scores 0.86245; adding learned half-strength early-season admission errors scores 0.86100, a nearly unchanged mean. The large retrospective improvement against all-finalized-input training therefore comes mainly from actual recent reporting inputs in this comparison.

At 07:06 EDT, admissions owner 3804709_0 had zero active fits and no pending work. It was released after a queue-lock check, allowing the final diluted C1 job 3818337_0 to start on its L40. No active fit was stopped in this release. The dilution job has a 45-minute limit, ending about 07:51; all remaining admissions fits continue on their existing allocations.

Completed C2/C3 admissions-only all-season comparison

For C2, the target MLP without spatial sharing using six covariate groups, training on 2022–23/2023–24/2024–25 with half-strength empirical admission errors in every training season, while ED and covariate values stay finalized, scores 0.96053 ± 0.03444 on 2025–26. Errors come only from 2024–25; finalized future labels and Wednesday archived/proxy evaluation inputs are unchanged. This improves all-finalized-input C2 training, 0.98733, by 2.72%; all three paired seeds improve. It is an alternative to actual-recent training, not a pure first-two-season attribution contrast: the recent season and affected signals differ. Its retrospective score is 0.90603 when training instead on 2022–23/2023–24/2025–26 with errors only from 2025–26, versus 0.93397 after all-finalized-input training.

The C2 learned half-strength admission generator in the first two seasons, retaining actual recent reports, scores 0.97044 ± 0.01683 forward. Although it improves all-finalized-input training by 1.71% in all three seeds, it is 0.42% worse than actual-recent-only training, 0.96640; only one of three seeds improves that attribution reference. Its retrospective score is 0.84738.

C3 with neighbor sharing and Kinsa, using empirical half-strength admission errors in all three training seasons and finalized ED/covariate inputs, scores 0.96745 ± 0.03656 forward versus 0.97501 after all-finalized-input training. Only one of three paired seeds improves despite the 0.77% lower mean. Its retrospective mean is 0.92468. These results use the same labels and evaluation protocol above. The strongest completed C1 mean remains 0.93984 after all-finalized-input training; no cross-architecture superiority claim is based solely on one favorable seed.

Transporting reporting errors to 2022–23/2023–24 assumes that the permitted donor season's reporting process is informative about those earlier seasons after matching level, epidemic phase, calendar position and location. The experiment does not establish that historical reporting systems were stable across those years. The artificial-input goal is robustness to the specified evaluation regime, not reconstruction of unobserved historical archives.

C3 attribution with three reference seeds completed

The C3 target MLP with neighbor sharing and Kinsa, trained on finalized 2022–23/2023–24 inputs plus actual-or-proxy 2024–25 reports, scores 0.94373 ± 0.02378 on 2025–26. All future labels remain finalized and evaluation inputs are identical archived/proxy Wednesday values. This is 3.21% better than all-finalized-input C3 training, 0.97501; two of three paired seeds improve. Its retrospective mean, using actual 2025–26 training inputs and evaluating 2024–25 flu/COVID admissions, is 0.86213.

Against that matched actual-recent-only reference, the original artificial first-two-season treatments have worse forward means: learned mean only +0.62%, learned mean plus residuals +1.59%, full-strength empirical +1.39%, and half-strength empirical +5.32%. Hence their apparent improvements against all-finalized-input training do not demonstrate added value from synthetic first-two-season errors.

The admissions-only half-strength empirical treatment in the first two seasons does score slightly better: 0.94058 ± 0.03156, 0.33% below the actual-recent reference, with two of three seeds improving. It improves all three seeds relative to all-finalized-input C3 training, by 3.53% in the mean. Its retrospective score is 0.83249. The tiny forward incremental benefit over actual recent reports alone should not be described as a robust breakthrough; it is also essentially tied with the simpler all-finalized-input C1 model's 0.93984. The learned half-strength admissions-only C3 fit remains pending at this interim comparison.

C2's empirical half-strength admission errors in the first two seasons, with actual recent reports, finish at 1.00168 ± 0.06650 forward: 3.65% worse than its actual-recent-only reference and 1.45% worse than all-finalized-input training. Its retrospective mean is 0.86422.

Idle admissions owners 1 and 3 were released after locked checks confirmed zero active fits and no pending work. At about 07:20 EDT, only the rerouted C3 learned-admissions seed 44 remained on owner 2; the final C1 dilution comparison continued on its separate allocation.

Shared-target seasonal diagnostic

Without retraining, the secondary score restricts both evaluation seasons to flu and COVID admissions with equal target weights. On forward 2025–26, C2 trained on finalized 2022–23/2023–24 inputs and actual-or-proxy 2024–25 inputs scores 0.96370 on these shared targets versus 0.94281 after all-finalized input training: 2.22% worse despite its better official six-target score. C3 with the analogous inputs scores 0.94502 versus 0.94275: 0.24% worse despite its official six-target improvement. Both learn unchanged finalized future labels and evaluate identical archived/proxy Wednesday inputs. Their retrospective shared-target gains are 2.66% and 11.96%, respectively, when training instead uses 2025–26 actual inputs and evaluation uses 2024–25.

The C2 model with half-strength empirical admission errors in all three training seasons improves both forward objectives: 2.72% on the official six-target score and 1.20% on shared flu/COVID admissions, to 0.93146. In contrast, C1 with empirical half-strength admission errors in the first two seasons plus actual recent reports improves forward shared admissions by 3.11% but worsens the six-target score by 3.66%. Target support therefore explains part, but not all, of the seasonal difference. The six-target forward score remains primary. Complete values and unchanged-fit definitions are in each experiment's common-admission-season-results.csv.

Admissions follow-up completed

All 36 admissions-focused seed runs completed both seasonal folds by about 07:27 EDT. The last C3 learned half-strength admissions treatment, applied only to 2022–23/2023–24 while retaining actual-or-proxy recent reports, scores 0.94701 ± 0.01996 forward and 0.84071 ± 0.02983 retrospectively. Forward training uses 2022–23/2023–24/2024–25 with errors learned only from 2024–25; retrospective training instead uses 2025–26 as the recent donor season. Both learn unchanged finalized future labels and evaluate the same archived/proxy Wednesday inputs as their references.

The learned treatment is 2.87% better than all-finalized-input C3 training, but 0.35% worse than C3 trained with actual recent reports and no artificial early-season errors. Only one of three seeds improves the latter reference; the paired score difference is +0.00328 with SD 0.01701. Consequently, none of the tested learned first-two-season generators improves the forward mean beyond actual recent reports alone on C2 or C3. C1's learned admission generator partly repairs its worse actual-recent setup but still does not beat all-finalized-input C1 training.

The final common ranking for the completed follow-up uses the ordinary rank -e vintage-overnight-admissions-focused-20261005 --no-plots, without --allow-incomplete. Its owner 2 exited normally. The only intentional active fit interruption was the documented timing-driven restart; that seed completed on its second attempt. The final diluted C1 comparison is the only remaining training work at this point.

Target tradeoffs for the strongest C2 artificial-input treatment

C2's empirical half-strength admission errors in all three training seasons reduce forward RSV-admission relative WIS from 1.1863 to 1.0153 and COVID admissions from 0.9712 to 0.9476. Flu admissions are almost unchanged, 0.9144 to 0.9153. All three ED targets worsen: COVID 0.9635 to 0.9877, flu 0.9736 to 1.0094, and RSV 0.8052 to 0.8910. This comparison retrains the same C2 target MLP with six covariate groups on 2022–23/2023–24/2024–25, uses only 2024–25 reporting errors, preserves finalized future labels and evaluates identical 2025–26 Wednesday archived/proxy inputs. The aggregate improvement is driven mainly by RSV admissions under the specified admission-twice-ED weighting; it is not an improvement for every target.

“Admissions-only” identifies which training input signals receive artificial errors. Every target-specific MLP still uses the other target signals as predictors, so perturbed admission inputs can change ED forecasts. Comparisons with all-signal perturbations also restore finalized covariates, not only ED values; C2/C3 cannot isolate an ED-only effect from that contrast. C1 has no covariates.

shared-target-season-comparison.csv, .png and .pdf place the retrospective flu/COVID-admission objective beside the forward shared-admission diagnostic and the primary forward six-target objective. They use the same fitted forecasts and three seeds in every cell. The completed C3 learned half-strength admission treatment improves forward shared admissions by 4.02% relative to all-finalized-input C3 training, despite failing to improve the primary six-target score beyond actual recent reports alone. Target tradeoffs remain material even when a learned reporting model helps selected admissions. Regenerate with PYTHONPATH=src .venv/bin/python scripts/vintage_overnight_shared_targets.py.

Final diluted result and completion

The final C1 pathogen MLP with distance sharing and no covariates, trained with half clean episodes and half half-strength empirical-error episodes across all three permitted training seasons, scores 0.94765 ± 0.02118 on forward 2025–26. Training uses 2022–23/2023–24/2024–25, with errors only from 2024–25; finalized future labels and archived/proxy evaluation inputs are unchanged. This is 0.83% worse than C1 trained on finalized inputs, 0.93984; only one of three paired seeds improves. Retrospectively, training instead uses 2025–26 as the recent error-donor season and evaluates 2024–25; its mean is 0.86518 ± 0.05055 versus 0.93467 after all-finalized-input training, a 7.43% improvement. The two score columns have the previously documented different target support.

All 117 successful research seed runs completed both folds by 07:35 EDT: 78 first-round runs, 36 admissions-focused runs and three final dilution runs. There are 37 fully replicated three-seed conditions and six single-seed pilots. The primary Slurm arrays have exited; at 07:36 no research GPU allocation remained. final-job-state.json records the scheduler and dispatcher audit. Idle-owner cancellations are distinct from the single intentional timing-driven fit restart, which completed successfully on attempt 002.

The compact final report is results.md. Full reproducible planning, launching, status and ranking commands, including the narrowed first-round dispatch and source-version distinction, are in manager-commands.md. All scores, paired comparisons, graphs and scientific assumptions are retained in this directory.

Final source and metadata integrity checks

Independent reproduction-source-v1 and reproduction-source-v2 directories were prepared on Longleaf by copying the original pinned sources and their notifiers. All 74 source/configuration files in each copy match the originals; checksums are saved in reproduction-source-checksums.json. These copies avoid self-replacement during a new plan and preserve the old generator versions. No new experiment was planned or launched during this preparation.

The latest integration source also limits learned-model metadata fields to generator=ridge. The older pinned empirical runs inherited a conditional-model and centered-residual description even though they used raw empirical error windows. reporting-metadata-clarification.json explicitly corrects this description, alongside the truth-support/archive-pair distinction. These descriptive changes do not alter fitted numerical inputs, checkpoints or scores. The final integration patch applies cleanly to the saved initial private source.