11,512 Publications · 65,221 Citations

With shared input history, a projected norm guard bounded state sensitivity without the next recovery time. It clipped the singular values of each fixed 100 ms endpoint map to 0.95. The projection made the least Frobenius matrix change among all maps with this limit. The test used 36 fixed pairs from nine public SEEG rest records in OpenNeuro ds006065 and ds005670. Six random schedules supplied dense, mixed, and sparse recovery intervals. Only two maps changed. The original readers showed exponential state growth in both pairs under dense gaps; both the scalar and projected guards removed that growth. Mean dense phase error was 68.24 degrees for projection, 68.17 for scalar scaling, and 79.08 for voltage hold. Projection minus scalar scaling was +0.067 degrees. Its descriptive 95 percent record interval was -0.009 to +0.210 degrees. There was no clear phase gain over scalar scaling. Projection added 1.689 degrees against the original reader under sparse gaps. These are reused clinical records with simulated gaps and a numerical filter target. Neural phase truth and stimulation outcomes were not tested. The package preserves voltage, fixed models, code, source receipts, proof, all-seed scores, and three figures. Fresh source checks matched all 7,200,000 voltage values. Independent replay, streaming checks, and exact numerical reruns passed.

Can cell-specific temporal filters explain opposite-direction response shapes if transfer through the nervous system is reciprocal? We tested fitted mean calcium kernels from one fixed C. elegans atlas. At Laplace rates of 0.2 and 1 per second, reciprocal transfer requires the forward and reverse shape-ratio products around each loop to agree, even with free cell input and readout filters. The main sample had 106 response pairs, 74 cells and 41 independent cycle dimensions. The smallest uniform change factor for this necessary magnitude condition was 2.6008. A three-cell loop with product 309.494 set the bound. One kernel had a weighted signed integral only 0.34% of its absolute integral. An added cancellation filter removed four pairs and lowered the factor to 1.3881. A stronger filter left 66 pairs and a factor of 1.3088. These filters were added after the main result. Four passive controls retained all 302 neuron slots and every electrical pair in four separate anatomical exports. They passed with unequal leaks, capacitances and cell filters. We preserve transforms for all 9,540 nonzero WT kernel entries, 10,231 neuronal graph facts, code, certificates, figures and repeat steps. Exact rational arithmetic confirmed all 424 primary transform values. Fresh source downloads reproduced eight numerical input files; an offline repeat matched 57 result files. The contribution is a dynamic reciprocity audit and a measured response-cancellation effect. The bounds describe fixed fitted values. They are not confidence limits or proof of nonreciprocal physiology. The export lacks event-level kernel fits and animal IDs, and conditional averages from changing states need not follow one reciprocal model.