Portable M03 independent review
==============================

Requires Python with NumPy and SciPy. The saved-field and record audits are
portable. The guarded new-reference runner additionally requires64-bit Windows,
as declared by its source; it has no unguarded OS fallback. This package does not
install dependencies, operate a device, or reproduce a physical experiment.

Extract the exact public author core and this reviewer ZIP into NEW directories.
Never overwrite the supplied original evidence with a reproduction. The author
core contains the current M03 source/results and history/attempt1 with its
PUBLIC-TRANSFORMATIONS.json. Set AUTHOR below to that current source directory,
the directory containing M03-PLAN.json, resource_guard.py and results/.

From the reviewer directory, verify the five analytical helper tests:
  python -m unittest test_audit_helpers -v

Audit the frozen complete attempt using the supplied two reference fields:
  python audit_results.py --study AUTHOR --results AUTHOR/results --independent references --clearance SOURCE-CLEARANCE-REVISION2-FINAL.json --output NEW-COMPLETE-AUDIT.json

Audit the full public interrupted-attempt history:
  python audit_public_history.py --history AUTHOR/history/attempt1 --output NEW-HISTORY-AUDIT.json

That public history audit will report original_private_bytes_locally_reverified
as false. This is expected: it verifies every public member/hash, original/public
mapping, record and retained field without access to the removed private path.
The included PUBLIC-HISTORY-TRANSFORMATION-REVIEW.json separately records the
local comparison against the original bytes. The original244-check partial
receipt is historical; the compact public replay groups some checks differently.

The above audits do not solve another full grid. They independently recover six
current and four historical saved fields and recompute the reported metrics.
The complete audit uses two supplied reference results from the prospectively
fixed midpoint solves; it does not pretend all322 points were separately solved.

Optional: regenerate exactly those two reference solves under their original
120second total,55second per-worker and1GiB Job commit limits. Stop other main
FEM workers first. Use a fresh destination:
  python run_reference.py --guard-source AUTHOR/resource_guard.py --author-summary AUTHOR/results/M03_summary.json --output NEW-REFERENCES
Then repeat audit_results.py with --independent NEW-REFERENCES. Runtime, UTC,
resource telemetry and archive timestamps can vary. Field/metric agreement is
tested at the stated tolerances; source and original evidence hashes do not vary.

To reproduce the author's full322-point main matrix, follow its README and
M03-IO-ADDENDUM.json in a fresh output directory. The provided frozen source
still records its pre-execution hold statement as historical context. Execution
of a new replay is the reader's action; no source file or evidence is overwritten
by these reviewer commands. Existing failed/partial attempts remain evidence.

Read FINDINGS.txt and REVIEW-METHOD.txt before interpreting a pass. Verification
passes coexist with failed discretization-accuracy screens and zero physical
validation. The historical modal denominator and the direct denominator differ.
