Rejected prospective semantic campaign, September 18, 2026.
128 full-cohort updates plus one bootstrap update issued 129 trial NEURO.
Quality failed: single 13/16, combined 11/16, zero retained-correct losses.
Three fresh quality auditors agreed. No new graph was promoted.
Use the matching execution-source archive and its declared Python dependencies.
Set PYTHONPATH to its src directory, ATEN_CPU_CAPABILITY=default,
MKL_ENABLE_INSTRUCTIONS=SSE4_2, OMP_NUM_THREADS=1, OPENBLAS_NUM_THREADS=1,
MKL_NUM_THREADS=1; run python evidence/replay_public.py.
Expected: 446 accepted transactions, 9398 headers, 129000000 issued atoms.
Ledger replay does not redo GPU work or establish operator independence.
Every numerical boundary has a hash-verified public publication receipt in jobs/.
The complete repair archive is development evidence only; it does not change
this failed result or count as an independently successful learning cohort.
