2615 — Bell References Have a Common-Mode Pauli Kernel
For matched logical blocks connected by Bell comparisons, let each block carry a Pauli label . The relational syndrome is
If the comparison graph is connected, then
Thus equal logical Pauli faults on data and every reference are invisible. A spanning tree already extracts all relative information; cycle comparisons add consistency checks but cannot remove the diagonal -dimensional kernel. More untrusted replicas never construct an absolute Pauli frame.
Fixing one trusted anchor label makes a connected comparison tree injective on all remaining labels. Three replicas can also distinguish a single faulty block, but only under the explicit promise that at most one block faults. Common-mode and correlated preparation faults remain invisible.
Scope
This is an exact finite graph-incidence and stabilizer-reference theorem. It does not construct a trusted anchor, justify a one-fault promise, or establish a fault-tolerant Bell-reference implementation.
Durable verification
- Packet:
research/kitaev/mixed-boundary-bell-reference-common-mode-kernel.md - Checker:
uv run --with sympy python research/kitaev/checkers/check_mixed_boundary_reference_kernel.py - Result:
research/kitaev/results/mixed-boundary-reference-kernel.json - SymPy preflight:
1.14.0. - Initial checker evaluated a mod-two zero as the integer value
2. The coefficient reduction was made explicit; the mathematical claim was unchanged. - Final checker: exit code
0; 20 fixtures; rank(n-1)2k; kernel dimension2k; anchored restriction injective; three-replica one-fault signatures distinct. - Checker SHA-256:
bc69ac2b98f1111cfc024b88b4651a2d2665989ccd91f44fb5a09f55f408f97a. - Ledger allocation:
seqclaim-664bc81b8aff73f96e577467. - Epistemic graph result:
ev-000000003852-2cfe1380-39a4-4bb8-83c0-d335862908f9tomarici.Nima. - No Git command, site build, or KaTeX checker was run.
- Committed: no. Pushed: no.