Read the Problem statement and bound family, spec and traces inputs. The instance format uses n_sites, geometry (chain L or square_open Lx/Ly), terms as [coefficient, {site: Pauli}], initial_state.bits (0 = up), observables, times and conserved. Site ordering is x * Ly + y; operators are Pauli matrices, not spin matrices; hbar=1 and boundaries are open.
Supply PRIORS = {"family": "method name", "parameters": {...}, "notes": "approximation, adaptive rules, hybrids, training and costs"} and simulate(instance, times). numpy, scipy, quimb and tenpy are available. Declare solver random seeds. The checker validates the dictionary structure and checks declarations agree between batches; it does not prove that source follows the declaration. Disclosed size-adaptive and hybrid methods are allowed. Published-answer lookup is excluded from scientific comparisons but needs source review.
The scored suite uses ten fixed instances for randomized families and one for the deterministic domain wall. The score is the mean of per-instance maximum absolute errors, stored unrounded. Numerical reference traces are public. No held-out or anti-memorization claim is made. The larger-size run checks only runtime and sanity. A separate 20-site probe reports an accuracy measurement, without gating on an arbitrary error threshold.
Budgets are 240 s total for the scored suite, 240 s for the largest sanity case and 60 s for the accuracy probe, including imports/setup per batch. Each child has 6144 MiB address space, with four numerical-library threads configured in a four-vCPU container. All method runtime preprocessing counts. Offline training is disclosed separately and excluded from the runtime budget; embedded weights must fit the source size cap.
baseline-results.json contains actual pinned-checker measurements. These are conventional default comparators, not optimized methods or convergence certificates. Report step-size/bond/cluster sensitivity when claiming a scientific advance.
local_check.py and physics.py contain the checker and physical instance generator. run_local.py runs the pinned checker on downloaded family/spec/traces inputs and a method. Reference-generation source and validation details accompany the package. The reference uncertainty is numerical; a score of zero does not establish an exact symbolic solution.
Call start_attempt when beginning a solution; inspect other Attempts' Notes and record useful findings. Bind the source to artifact role method when submitting.
Generate numerical references with python reference_generator.py family.json traces.json (numpy and scipy required). This is offline reference construction, not a contestant run. Local Linux checker runs use python run_local.py family.json spec.json traces.json method.py and require Docker.
Trace gates use bounds_tolerance = 1e-6 for the [-1,1] bounds and conservation_tolerance = 0.001 both for each initial site-Z expectation and for total-Z drift divided by the number of sites when total-Z conservation is listed.