CelerisLab/tests/validation/multigrid/cpuref_0_3_evidence/README_grok.md

Grok independent tests for CPUREF-0.3 (2026-10-13)

Runner: Grok's isolated sandbox (not aws-la, not Horsea). Python 3.13.5, NumPy 2.5.3, Linux x86_64. No installs, no Horsea/remote writes. Command: OPENBLAS_NUM_THREADS=1 prlimit --as=268435456 python test_cpu_reference.py then test_tron_additions.py and test_open_corner_negative_control.py (reference copied unchanged to cpu_reference.py; CPU_REFERENCE_PATH accepted). Budget: <=8 coarse steps per case, address space capped at 256 MiB, peak RSS ~43 MiB, stdout <2 MiB.

SHA256

• cpu_reference_v0.3.py (Drive 1G_a5i4ywwAgz20Binl9jI_d_SyXvfCJK): bba2a06a9f6517b32079ffff15e4060d8c3b7ed35f16ca134e1877226c9e4e1e
• test_cpu_reference.py: ade5d4827f971aed7174b59d83769e7aa542338e625a9fc8db1e8b1805cb74ca
• test_tron_additions.py: 8e7a6a90b85e64c7326719e7b8bca2f6d34fb4c80690f0685f7e291ab93bd2a4
• test_open_corner_negative_control.py: 59a6f8ca80cdd63cb7827fbebf9b4b555c3793f19dd6bca801544ec10b5cbe1b
• stdout_final.txt (20 unittest OK): 73e4ef01138842981daaa534390f1463839e63833c11f4e8872f94f8eaadc717
• stdout_open_corner.txt: 927f8233957fe08f671cf3716e810431c40331caa4cf406da2b09793337c7ee8
• contract still ee8010e763d38937b95c34dadaf97415cdb8b6c23e3c79a9f1f32bd5d4e112f6

Result: 20/20 pass (15 core adapted from v0.2 suite + 5 additions: NaN masks, 8 unique corner keys/packets, coalesce_paths reconstructed test-side, hop2_valid, v0.2↔v0.3 field delta) plus open-corner negative control OK. Reference unchanged.

Adaptations vs v0.2 tests: drop explosion_count; fine_valid_after→hop2_valid; closure_error→bookkeeping_check; corner_packets keyed lower_left_q8 etc. (8 unique); coalesce_paths rebuilt in tests; invalid tick (odd/float/bool/str) rejected; even tick advances by 2.