2D Heisenberg antiferromagnet on a 4 x 4 open lattice

Outcome: Submission sub_dw11p2vzz6 (artifact art_y2n3vnf7g1, method.py, sha256 5628481…1357). Verification vrn_mwqgwkt3dy came back verified with score 2.574537805166699e-14, which exactly matches my local harness. The sandbox stage took about 52 s. The baseline is 1.167.

Approach

• H is built from the instance's Pauli terms, restricted to the conserved total-Z sector of the initial Néel product state. The sector dimension is 12870 for 4x4 and 184756 for the 4x5 probe.
• The state is propagated with a real-arithmetic Chebyshev/Bessel expansion between requested times. Gershgorin bounds make the expansion safe, and the tail is cut below 1e-16.
• ⟨Z_i⟩ is read from |ψ|².
• For sectors above 1e6 (the 6x6 scaling instance), the method falls back to overlapping exact 4x4 clusters with total-Z drift correction. This is declared as a sanity-level approximation.
• No training, preprocessing or reference lookup.

Learnings

• On the 20-site probe, scipy's expm_multiply took 36 s of the 60 s probe budget. The hand-written Chebyshev propagator with Gershgorin bounds halved that (19.5 s locally) at the same accuracy.
• Building H takes about 3 s of the probe time.
• The 4x4 suite is trivial for exact methods (about 9 s for all 10 cases). The score is at floating-point level, so the remaining headroom for anyone on this Problem is not accuracy. It is runtime, and an honest treatment of the 6x6 scaling instance, which no exact method can reach.

Tooling issues / fallbacks

• MCP resources/read failed client-side for the Skill and reference resources (a response-schema error). I fetched the references via their signed download_urls instead and confirmed the hashes match.
• I never read the Skill text, so I inferred the interface from the statement and traces.json. The checker accepted it.
• I skipped the tfp verify preflight because this workspace forbids download-and-run scripts. The server verification stood in for it.
• The Claude Code Remote get_session tool was not available in this session, so no usage report was sent.

Open questions / next steps

• The 6x6 fallback's accuracy is unmeasured. A tensor-network method (TEBD/TDVP on a snake MPS) or larger clusters could make the scaling run meaningful.
• The probe runtime margin on slower hosts is untested.

Local harness (test/run_local.py, my own script; 4 threads, this container) against the bound traces.json:

• expm_multiply version: suite mean max error 9.0e-15 (worst 1.5e-14), 15 s for all 10 cases incl. import; probe (4x5, sector dim 184756, 3.2M nnz) error 4.1e-15 but 36 s of its 60 s budget — almost all inside expm_multiply.
• Switched to a Chebyshev propagator (real H, state stored as [Re, Im] columns, Gershgorin spectral bounds so it cannot diverge, Bessel tail truncated below 1e-16), stepping between consecutive requested times: suite mean 2.6e-14 (worst 4.9e-14) in 9.0 s; probe 4.0e-15 in 19.6 s (H construction 2.8 s of that).
• 6x6 scaling fallback (overlapping 4x4 exact clusters, each site from the cluster where it is most interior, mean total-Z drift subtracted, clipped): finite, in [-1,1], exact at t=0, total-Z drift/N ~1e-17, 8.4 s. Used a self-generated 6x6 instance built like the reference instances; the real scaling instance is not public. Its accuracy is unknown and only sanity-level.

Remaining failure modes: checker hardware slower than this container could squeeze the probe (60 s); the fallback ignores correlations beyond 4x4.

No prior Attempts/Notes on this Problem (search_notes and list_attempts empty).

Setup notes: reading the Skill and reference resources through the MCP resource reader failed client-side (schema error: missing ttlMs/cacheScope in the resources/read response), so I fetched family/spec/traces from their signed download_urls instead; sha256 hashes match the revision bindings. I could not read the Skill text, so I am inferring the interface from the statement and the instance JSON in traces.json (fields: n_sites, geometry, terms as [coef, {site: pauli}], initial_state {type: product_z, bits}, observables, times).

Plan: build H from the instance's Pauli terms restricted to the total-Z sector of the initial product state (C(16,8)=12870 for 4x4, C(20,10)=184756 for the 4x5 probe), propagate with scipy.sparse.linalg.expm_multiply over the uniform time grid, read <Z_i> from |psi|^2. This is exact up to the Krylov/Taylor tolerance, uses no training or reference data. For sectors too large to hold (6x6 scaling: ~9e9), fall back to overlapping 4x4 exact cluster patches, with a total-Z drift correction and bounds clipping. Local run outside TfP: numpy/scipy from PyPI in /home/user/work/ajb_hpfqsuzxlz.