Read the contract and proof. The construction replaces the growing all-to-all refresh with an explicit bounded-degree expander circuit. A fixed positive fraction of arbitrary faulty locations preserves the live old majority and leaves a correctable new word. The quantum export decoder is staged, so its error bound is uniform in concatenation depth. Together they give a source realization with noisy quantum computation, runtime control and actual public central records at fixed nonzero subthreshold elementary strength. No ideal runtime controller is added.
The proof uses credited expander and computation-code theorems. Its graph degree 8^12 and other fixed constants are deliberately conservative and very large. This is a constructive positive-threshold result, not a measured native-machine threshold or a practical hardware proposal. Source/CCG, static scheduling, control resources and logical accounting retain their declared status. The branch builds on merged main; it does not need an unmerged PR's files to verify its evidence.
- Six exact powered graphs, through 256 vertices, are generated by forward walks and checked by independent symmetric edge insertion and matrix exponentiation. Exact integer row-norm certificates verify their finite mixing bounds. Loops and repeated edges are retained.
- The indexed compiler can address operations in the full 8^12-degree schedule without expanding it. Fully executed unpowered m=2,3 circuits have 1,372 and 3,087 locations. Bit-sliced and Boolean-array engines agree on all 91,740 single-fault cases with zero or one old error. The weaker unpowered circuits have failing cases; those are retained, not filtered.
- Integer powered-voter controls exercise adversarial shared-source and private-voter damage. The universal live invariant is the analytic proof, not a claim to execute the huge powered native circuit.
- The actual punctured RM(5,11) decoder runs on all 2,047 input bits. Tensor parity reduction and an independent packed-bit finite-difference decoder agree on 25 cases through radius 31, a 32-error wrong-label witness, and a retained failure/fallback case. The proof covers every radius-31 error set.
- Complete full-M6 flagged reduction is checked for d=2,4, including every input coherence and the private complement Kraus indices. The finite export experiments check full Choi matrices, spectator coherence, noisy archive labels and retained application aborts. Their deliberately visible errors are channel controls, not samples below the enormous machine's symbolic threshold.
- Exact tail coefficients, integer refinement ceilings and lifetime exponents are independently reconstructed. Coherent staged-channel controls contrast the uniform decoder bound with an unprotected serial error floor. Relative checks retain even the smallest positive stage terms; replacing them by zero, negative values or incorrect multiples is rejected. Full-degree schedule tests independently reconstruct random walks and bubble-sort row boundaries, rather than only checking agreement between operand lookups.
- A complete full-M6 erasure channel and a classical bit-flip channel give a concrete fixed nonzero noise family inside the derived interval. Every matrix unit and the quantum dilation-distance bound are checked.
No full universal concatenated apparatus is claimed executed. Hashes bind
the sources and scoped claims; graph, Boolean, decoder and operator replay
supply the semantic checks. Hostile tests reject substituted graphs, missing
idles, shared-voter corruption, forged decoder results, dropped outcomes,
invalid growth laws and malformed receipts. Harmless private Kraus-basis
changes pass. Verification works with the producers disabled and under
python -O.
From the repository root, set PYTHONPATH=code and run:
python -m m1_expander_archive.build
python -m m1_expander_archive.verify
python -m pytest -q code/m1_expander_archive
The dedicated workflow performs the same replay and hostile controls on Linux and Windows. The receipt also recursively verifies the merged noisy records, fixed-noise and source evidence. The compact receipt stores no raw fault tapes; every experiment is regenerated from these scripts.