Scenario fields

Generated from chromantis.model.scenario.Scenario (CODES, MEANING) by chromantis.evaluation.report.write_scenario_key, rewritten by every report; do not edit by hand. A scenario string names only the fields that differ from these defaults, as key=value tokens joined by ,; booleans are written 0/1. See the architecture and Workflow.

Field	Type	Default	Allowed values	Meaning
history_source	str	'finalized'	artificial, finalized, reported	Choice A1, where the network's training histories come from: final reference histories (finalized), actual archived Wednesday reports (reported), or final histories made preliminary with artificial reporting errors (artificial; choices B and C). Was part of pilot_method.
history_correction	bool	0	0, 1	Choice A2: correct the artificial training histories with the cross-fitted correction model (choice D) before the network learns from them (was pilot_method=two_stage/corrected).
reconstruction_labels	bool	0	0, 1	Choice A3: also learn the final values of the last reconstruction_weeks context weeks (was pilot_method=joint). Labels, not inputs.
error_scope	str	'all'	all, early_actual	Apply artificial reporting errors in all training seasons, or early two with actual/proxy recent-season reports.
error_signals	str	'admissions'	admissions, all, targets	Choice C: admissions perturbs counts only; targets also perturbs ED but keeps covariates unchanged; all perturbs both targets and covariates.
corrector_examples	str	'synthetic'	real, synthetic, synthetic_then_real	Choice D, what the correction trees (the nowcaster) learn from: artificial errors paired with final values (synthetic), real archived reports paired with their mature values (real), or synthetic pretraining then real (synthetic_then_real, MLP only). Was pilot_nowcaster.
corrector_model	str	'tree'	mlp, tree	Correction model: gradient-boosted trees (tree) or a residual MLP (mlp). Was pilot_nowcaster.
input_set	str	'all'	named by the dataset specification	Named target-history input set declared by the dataset specification.
ili_units	str	'own'	flu_scaled, own	Own-source ILI proportions or auxiliary flu-head pseudo-tasks rescaled by fitting-only modern Q95 / historical ILI Q95.
ili_steps	int	200	any int (checked in Scenario.__post_init__)	Historical forecasting pretraining updates; each uses 64 auxiliary examples (series models) or batch_size historical panel episodes (flu MLP).
correction_weeks	int	8	any int (checked in Scenario.__post_init__)	Recent admissions ages corrected by the nowcaster, clipped to lookback.
correction_ed	bool	0	0, 1	Also fit/apply report-to-mature ED correction on the recent correction_weeks, restricted to allowed input channels; future labels unchanged.
ili_training	str	'none'	joint, none, pretrain	No historical transfer, forecasting pretraining, or joint forecasting on the pinned pre-2022 ILI archive.
ili_path	str	'data/processed/historical_ili.npz'	any str (checked in Scenario.__post_init__)	Historical ILI auxiliary dataset; its hash is pinned by the experiment manager.
ili_weight	float	0.25	any float (checked in Scenario.__post_init__)	Historical ILI loss weight for joint training after within-history normalization.
growth_anchor	bool	0	0, 1	Add damped observed two-week transformed growth to the level anchor.
reporting_probability	float	1.0	any float (checked in Scenario.__post_init__)	Probability a training episode receives artificial reporting errors.
reporting_random_strength	bool	0	0, 1	Scale each drawn error window by a uniform(0, 1) multiplier.
reporting_recent	int	0	any int (checked in Scenario.__post_init__)	Perturb only this many most recent context weeks; 0 = all.
joint_weight	float	1.0	any float (checked in Scenario.__post_init__)	joint only: weight of the reconstruction loss relative to the forecast loss.
correction_penalty	float	10.0	any float (checked in Scenario.__post_init__)	Ridge penalty of the correction trees' linear leaves.
correction_strength	float	1.0	any float (checked in Scenario.__post_init__)	Multiplier of the fitted correction, 1 = full.
reporting_missingness	bool	1	0, 1	Transfer donor reporting masks with the errors; false transports numerical errors only and retains native input availability.
reporting_method	str	'local_log'	calendar_log, local_additive, local_log, phase_log, synchronous_log, synchronous_phase_log	Donor matching of artificial reporting errors; see dataset/reporting_error.py.
reporting_strength	float	1.0	any float (checked in Scenario.__post_init__)	Multiplier of drawn log reporting errors.
reconstruction_weeks	int	2	any int (checked in Scenario.__post_init__)	joint only: recent context weeks whose final values are reconstruction labels.
lookback	int	12	any int (checked in Scenario.__post_init__)	Context weeks per episode (the history the network sees).
count_transform	str	'fourth_root'	fourth_root, log1p, rate, raw, sqrt	Admissions in model space: raw counts, rate per 100,000, or its sqrt / fourth root / log1p (model/network.py transform_counts).
ed_transform	str	'linear'	fourth_root, linear, logit	ED proportions in model space (model/network.py); scores stay in native units.
geography	bool	1	0, 1	Adds log population and a native-US flag per location as features.
coordinates	bool	0	0, 1	Census state internal-point latitude/longitude and non-US indicator.
dynamics	bool	1	0, 1	Recent-dynamics feature block (30 slope/acceleration/age/validity features); see architecture.md.
loss_weights	str	'objective'	admissions_only, ed_only, objective	Training target weights: the problem objective, count targets only, or proportion targets only.
encoder	str	'mlp'	conv, mlp, multiscale_conv, series_mixer, series_mlp	Temporal context encoder; see architecture.md.
spatial	str	'none'	attention, distance, gated_pool, gravity, joint_location_target, national_broadcast, neighbors, none, pathogen_spatial, pooled, target_spatial	Cross-location information exchange (none, shared attention, pathogen/target/joint scopes); see architecture.md.
heads	str	'shared'	shared, state_us	State and US output heads shared or separate (state_us).
decoder	str	'legacy'	legacy, quantile, quantile_small, residual2	Horizon decoder: existing modulated residual (legacy), residual2, or direct ordered quantiles; see architecture.md.
decoder_blocks	int	0	any int (checked in Scenario.__post_init__)	Residual decoder depth (0 = original head).
sum_wis_weight	float	0.0	any float (checked in Scenario.__post_init__)	Auxiliary four-week flu admission sum WIS weight; sampled heads only; normalized by four times training-only admission Q95.
noise	str	'global'	global, local	Global latent noise, or global plus a per-location latent (local).
us_error	str	'none'	none, shared_factor	Extra common noise factor (shared_factor); see architecture.md.
latent	int	16	any int (checked in Scenario.__post_init__)	Global latent (noise) dimension.
width	int	64	any int (checked in Scenario.__post_init__)	Hidden width of the network.
epochs	int	50	any int (checked in Scenario.__post_init__)	Epoch cap (the fixed number of epochs when patience = 0).
patience	int	0	any int (checked in Scenario.__post_init__)	Early-stopping patience in epochs; 0 = fixed epochs, no validation weeks; > 0 selects the epoch on the validation weeks, then refits on the full training seasons (design §4).
batch_size	int	8	any int (checked in Scenario.__post_init__)	Episodes per optimizer step.
members	int	128	any int (checked in Scenario.__post_init__)	Sampled members per training episode (fair CRPS loss).
lr	float	0.001	any float (checked in Scenario.__post_init__)	Adam learning rate.
head_sharing	str	'shared'	pathogen, shared, target	Output sharing: shared, three pathogen heads, or six target heads; see architecture.md.
annual_calendar	bool	1	0, 1	Annual sine/cosine and Christmas-timing features.
location_embedding	int	0	any int (checked in Scenario.__post_init__)	Dimension of a learned location-ID embedding (0 = none).
fit_partition	str	'all'	all, pathogen, target	One model for all targets, or separately fitted models per problem group / target; see architecture.md.
validation_members	int	256	any int (checked in Scenario.__post_init__)	Members drawn for the early-stopping validation loss.
weight_decay	float	0.0	any float (checked in Scenario.__post_init__)	Adam weight decay.
covariate_encoder	str	'raw'	growth, raw, shared, smooth, summary	Raw standardized history, signed-log trailing-three-week smoothing, six summaries, a shared 4-dimensional encoder, or recent growth (level, 1/2-week log growth, acceleration) plus coverage/age.
signal_features	str	'none'	multiscale, none, smooth_multiscale	Optional causal 3/6/12-week level, slope and curvature features for targets and covariates, with optional three-week smoothing.
covariate_set	str	'none'	named by the dataset specification	+-joined named covariate sets declared by the dataset; none = none; see workflow.md and experiments/b-2-t0/index.md#protocol.
input_normalization	str	'none'	b0, none	none, or B0 per-location transformed target scales fitted on training contexts only.
validation_calendar	str	'season'	b0, season	season-relative blocks, or B0 blocks counted from the first stored week of each season.
validation_weeks	int	3	any int (checked in Scenario.__post_init__)	patience > 0 only: consecutive early-stopping weeks hidden per block (design §4).
validation_spacing	int	16	any int (checked in Scenario.__post_init__)	patience > 0 only: one validation block every this many weeks of a training season.
validation_offset	int	4	any int (checked in Scenario.__post_init__)	patience > 0 only: week of each training season where the first block starts (0-based, counted from the season's first epiweek, CDC week 31).
error_seasons	str	'latest'	all, latest	Choice B: archived seasons supplying artificial reporting errors and real correction pairs: latest permitted training season, or every archived training season with recency weights 1, 1/2, 1/4.
actual_share	float	0.0	any float (checked in Scenario.__post_init__)	Probability a training episode uses the actual archived report (finalized fill where none) instead of an artificial-error draw.
correction_realizations	int	1	any int (checked in Scenario.__post_init__)	Independent artificial-report-then-correction realizations per training episode; one is drawn per minibatch.
uncorrected_share	float	0.0	any float (checked in Scenario.__post_init__)	Probability a corrected-history training episode is shown uncorrected instead (robustness to correction failure).
correction_noise	float	0.0	any float (checked in Scenario.__post_init__)	Scale of sampled out-of-fold correction residuals added to corrected admissions; 0 = point correction. Evaluation uses 16 sampled histories per issuance.
correction_noise_train	bool	0	0, 1	Also add sampled correction residuals to corrected training histories each minibatch.
log_loss_weight	float	0.0	any float (checked in Scenario.__post_init__)	Weight of an extra weekly admission loss (the trained admission targets) on the log(1 + count) scale, added to the native-scale loss.
covariate_dropout	float	0.0	any float (checked in Scenario.__post_init__)	Probability a training episode has all covariates (e.g. Kinsa) hidden.
training_window	str	'all'	all, last2, recent2	all = every non-held-out season; recent2 = two preceding seasons; last2 = latest two permitted non-held-out seasons (nested removal of oldest training season).
stress_views	bool	0	0, 1	Also evaluate delayed-admission and covariate-missing input views of the same fitted model.
error_reference	str	''	any str (checked in Scenario.__post_init__)	Choice B: prescribed reporting-error season for training errors and prescribed evaluation inputs, independent of the held-out epidemic season; empty = the fold's own seasons.
forecast_view	str	'corrected'	corrected, half, raw	Choice F, how the fitted pair is used at forecast time, part of the recipe: inputs as given (raw), newest weeks corrected by the fold's correction model (corrected), or the 50/50 mixture of those two forecast distributions (half). Ranking and production use this view; the others are diagnostics.