M02 portable independent review Requires Python, NumPy and SciPy. Obtain the author's frozen M02 source and results as sibling directory ../m02, preserving the exact inherited inputs, 36 direct vectors, CSVs, reports, summary and manifests. This reviewer package contains one independent reference field under reference/, not duplicate author arrays. SHA256 values bind exact inputs and results. Do not overwrite outputs. From this reviewer directory: python -m unittest -v test_audit_helpers.py python audit_public_results.py --study-root ../m02 --results ../m02/results --independent-direct reference --omission-record ../m02/publication/RETAINED-CHECKPOINT.json --output new-audit.json python final_checks.py --study-root ../m02 --results ../m02/results --output new-final-checks.json The original local INDEPENDENT-M02-AUDIT.json verifies the full saved manifest. The public thin core intentionally omits the optional n12 modal-basis NPZ. The public replay requires its explicit metadata record, checks that declaration against the original manifest, and reports that one member as unavailable and unverified. It does not silently excuse other missing files. The scalar/field checks do not require this basis because modal participation/projection data and every direct displacement field are supplied. PUBLIC-CORE-REPLAY.json is the separate actual replay against the exact extracted public study ZIP. The first command checks rigid translation, hydrostatic strain and harmonic engineering shear against analytical identities. The next commands read saved fields; they do not run new eigenproblems or direct solves. New timestamps differ. Optional reproduction of the one prespecified independent direct solution: python run_reference.py --guard-source ../m02/resource_guard.py --output new-reference-run python audit_public_results.py --study-root ../m02 --results ../m02/results --independent-direct new-reference-run/reference --omission-record ../m02/publication/RETAINED-CHECKPOINT.json --output new-audit-with-new-reference.json This optional run uses the reviewed Windows-only guard and a 1 GiB worker-job commit limit, 115 s worker/120 s total stop budget and 32 MiB output guard. It solves only n10 structural steel at 100000 Hz using fem_reference.py, which is the existing separately reviewed tensor implementation and does not import the author FEM engine. It may produce small platform-dependent floating differences. Optional resource-only diagnostic, on Windows: python probe_job_memory.py --guard-source ../m02/resource_guard.py --output new-memory-probe This creates a 64 MiB job and tests a small allocation plus an over-limit request. It performs no FEM solve. Native peak-counter semantics remain unresolved; this diagnostic must not be described as an instantaneous resident-memory proof. The supplied prospective/source review records are historical checks of their exact frozen snapshots. They are not new tests automatically rerun by these commands. The initial failed author preflight remains in the author's preserved history. Resource-observation/REPORTING-NOTE.txt documents a harmless reviewer console-summary failure after the complete observation JSON was already saved; its exact script is preserved for provenance, not used as the default command. Only declared numerical consistency and evidence fidelity were reviewed. Failed accuracy criteria remain failed. None of this is a physical test or a device control program. Matplotlib, browser rendering and hardware APIs are not used.