Benchmark Cards

These cards summarize the public small-scale validation metrics stored in experiments/public/outputs. They are not leaderboard claims. They are compact reproducibility cards for package behavior, tensor contracts, solver residuals, and example coverage.

The measured public validation summaries are also collected in Results.

How To Reproduce

silva-experiment --config graph_silva_smoke
silva-experiment --config solver_sweep
silva-experiment --config iris_tabular_silva

Run every checked artifact:

python scripts/release_audit.py
pytest tests/test_public_experiments.py

Core Cards

Card	Device	Task	Metric	Residual evidence
`graph_silva_smoke`	CPU	graph classification	accuracy `1.0000`	`1.0055`, `0.6067`
`iris_tabular_silva`	CPU	UCI Iris node classification	accuracy `0.8333`	`1.8182`, `0.1265`
`molecular_smoke`	CPU	molecular regression validation	MAE `0.0644`	`4.9945`, `7.1137`
`custom_operator_experiment`	CPU	custom local operator	accuracy `0.7500`	`3.3509`, `3.6641`, `2.8922`
`fully_configurable_graph`	CPU	graph classification	accuracy `0.6667`	`2.8020`, `2.8972`, `4.6459`
`vision_channels_smoke`	CPU	vision validation	accuracy `0.5000`	output shape check only

Dataset Cards

The public dataset cards exercise real download, preprocessing, tensor packing, solver execution, and gradient flow. They use small budgets so they can be run locally before release.

Card	Dataset family	Route	Metric
`tabular_dataset_suite`	UCI tabular	stratified kNN graph to `SILVAGraphNetwork`	accuracy and residuals per dataset
`cifar10_vector_smoke`	TorchVision CIFAR10	flattened image to vector SILVA classifier	accuracy `0.1667`, residual `3.8260`
`cifar10_cortex_smoke`	TorchVision CIFAR10	convolutional retina to cortex SILVA hierarchy	accuracy `0.1250`, residuals `8.8553`, `5.9625`
`torchvision_dataset_suite`	TorchVision images	vector preset across MNIST-family, CIFAR, and SVHN	opt-in multi-dataset validation

Solver Sweep

The solver sweep uses a small equilibrium system and records convergence, residual, stability residual, spectral radius, and a Jacobian norm estimate.

Solver	Iterations	Converged	Residual	Stability residual	Spectral radius
Picard	15	yes	`7.6803e-06`	`3.7923e-06`	`0.2602`
Anderson	16	yes	`9.1485e-06`	`2.5385e-06`	`0.2752`
Broyden	13	yes	`6.6698e-06`	`6.6698e-06`	`0.2744`

The card is a package diagnostic, not a proof that one solver dominates. On a new problem, compare residual curves, wall time, gradients, and memory.

Operator Ablations

Vision-vector ablation:

Case	Attention	Graph mode	Accuracy	Residual
full	simple	GAT	`0.9375`	`5.1449`
no_global	none	GAT	`0.8125`	`3.4175`
no_local	simple	none	`0.9375`	`4.2750`
none	none	none	`0.9375`	`3.7282`
static_global	static	GNN	`0.8750`	`4.5311`

Graph-operator options:

Case	Attention	Graph mode	Local depth	Accuracy
full	simple	GAT	1	`0.7778`
no_global	none	GAT	1	`0.8333`
no_local	simple	none	1	`0.6667`
none	none	none	1	`0.6667`
static_global	static	GAT	1	`0.6667`
topk_global	topk	GAT	1	`0.5000`
local_depth_2	simple	GAT	2	`0.7778`

Reading The Cards

For each card, ask:

Does the tensor contract match the intended case?
Did the solver residual move into a reasonable range for the validation budget?
Does the example run on CPU without optional production dependencies?
Is the reported metric identified as a small-scale validation metric unless a full benchmark protocol is documented?

Citation Rule

Reports that use these cards should cite the package and the SILVA article. If a card uses a specific external lineage, cite that lineage too: DEQ for equilibrium claims, GAT/GCN/MPNN for graph branches, RAFT/DEQ-Flow for optical flow, and OptNet/CVXPYlayers for optimization layers.

Where to Go Next

Question	Page
Where are the complete measured outputs explained?	Results
What evidence is required beyond compact validation?	Reconstructing Paper Experiments
How should metrics and claims be cited?	Citation-Aware Reporting