G2-N: tiny LSV1M step-current regression on build N, against upstream reference_data, beside the old G2 (logs/g2.txt) Build N: patch 0001 (LGN noise pair order), GSL 2.7.1 with -ffp-contract=off, pinned libomp and libltdl. NEST source and flags as before. Command: python run.py nest 2 param/defaults pytest, run folder runs/tiny-n, RUN differs from runs/tiny/RUN only in sourcing runtime-n and in paths. Environment: runs/tiny-n/env.txt (MPLBACKEND=Agg, PYTHONHASHSEED=0, OMP_NUM_THREADS=2). pynn_seed 995 (model default). Run: exit 0, 44.7 s, peak footprint 0.182 GiB. Tools, unchanged: downloads/compare_raw_output.py, feasibility-010 diag_compare.py, tools/g2_report.py, g2_rates.py, g2_first_diff.py. Script: build/n/g2n-compare.sh. VERDICT: PASS. X_ON 30 of 30 and X_OFF 30 of 30 recorded spike trains equal the reference (old G2: 6 and 9). All sheets: 1656 of 1656 trains equal (old G2: 15). The cortex matched too, which this step did not require. 1. Same tool, server, old Mac and build N (reference vs 1x2) server old Mac build N segment files compared 36 36 36 spike trains compared 1656 1656 1656 spike trains with different times 0 1641 0 largest signal difference (diag_compare) 4.26e-14 25.5 6.34e-09 segments with every signal bit-equal 24 0 0 identical_content (compare_raw_output) 24 0 0 identical_content needs every voltage and conductance bit-equal, so it stays at 0: the spike trains are equal, the signals differ in low bits. compare_raw_output exit 1 is that. Server source: feasibility-010 receipts/decomp/compare-ref-vs-1x2.json and diag-ref-vs-1x2.json. 2. Per sheet against the reference, over the 36 reference segments. old = logs/g2.txt, N = this run sheet trains equal equal spk spk spk Hz Hz Hz max dVm mV max dVm mV rms dVm max dg old N ref old N ref old N old N N N X_ON 30 6 30 150 151 150 30.77 30.97 30.77 12.98 7.11e-14 1.2e-14 0.0e+00 X_OFF 30 9 30 194 196 194 39.79 40.21 39.79 12.98 9.95e-14 1.3e-14 0.0e+00 V1_Exc_L4 642 0 642 29208 28930 29208 279.97 277.31 279.97 23.25 5.80e-10 9.6e-12 1.9e-12 V1_Inh_L4 114 0 114 10455 10347 10455 564.37 558.54 564.37 19.39 1.09e-09 2.5e-11 3.7e-13 V1_Exc_L2/3 696 0 696 32520 32100 32520 287.53 283.82 287.53 25.46 6.14e-09 1.1e-10 2.0e-11 V1_Inh_L2/3 144 0 144 21626 21290 21626 924.19 909.83 924.19 19.75 6.34e-09 1.3e-10 3.5e-12 all sheets: 1656 trains, 15 equal old, 1656 equal N; spikes reference 94153, old 93014, N 94153 annotations equal in 36 of 36 segments Largest signal difference by LGN alone: X_ON 7.11e-14 mV, X_OFF 9.95e-14 mV. The server against the same reference: 4.26e-14 over all sheets. 3. Opening blank (Segment6 to 11), the earliest simulated time: max |dV| at samples 1 to 5, mV, old then N X_ON old 2.6 1.4 1.6 2.8 3.2 N 0 0 0 0 0 X_OFF old 2.1 1.5 2 1.4 2 N 0 0 0 0 0 V1_Exc_L4 old 1.4e-14 1.4e-14 1.4e-14 0.024 0.029 N 0 0 1.4e-14 0 1.4e-14 V1_Inh_L4 old 0 0 7.1e-15 0 0.089 N 0 0 0 0 0 V1_Exc_L2/3 old 3.6e-14 2.6e-13 1.4e-14 2.8e-14 2e-13 N 0 0 0 1.4e-14 1.4e-14 V1_Inh_L2/3 old 0 7.1e-15 1.4e-14 2.1e-14 2.1e-14 N 0 0 1.4e-14 2.1e-14 2.8e-14 4. How far the cortical difference goes X_ON largest |dV| per recorded stimulus, mV: 2.8e-14 2.8e-14 7.1e-14 1.4e-14 5.7e-14 2.8e-14 X_OFF largest |dV| per recorded stimulus, mV: 1.4e-14 2.8e-14 9.9e-14 1.4e-14 5.7e-14 5.7e-14 V1_Exc_L4 largest |dV| per recorded stimulus, mV: 4.7e-10 1.3e-12 1.8e-10 5.8e-10 1.5e-10 1.6e-10 V1_Inh_L4 largest |dV| per recorded stimulus, mV: 7.8e-12 1.1e-12 1.1e-09 9.9e-13 2.4e-10 1.1e-09 V1_Exc_L2/3 largest |dV| per recorded stimulus, mV: 1.2e-09 6.1e-09 2.3e-09 2.5e-09 4.8e-09 1.3e-09 V1_Inh_L2/3 largest |dV| per recorded stimulus, mV: 3e-10 7.3e-10 2.5e-09 5.3e-09 3.9e-09 6.3e-09 Six entries per sheet, in file order, which is not simulated-time order, so this table shows size and not growth. Cortex: about 1e-12 to 6e-9 mV. LGN: at most 1e-13 mV. 5. What this shows The old G2 diagnosed the pair order from source and said it was not confirmed by experiment. This run is that experiment, and it confirms it: with the order restored and the same seed, LGN voltage is exactly equal at samples 1 to 5 of the opening blank (old: up to 3.2 mV), and every spike train in every sheet equals the reference. The run changed two things together, the pair order and GSL (2.8 with fused multiply-add to 2.7.1 without). The size of the old LGN difference says the pair order was the main cause. This run does not separate how much of the old 4.4e-12 mV arithmetic gap came from GSL. 6. Limits Tiny model, 2 threads, 1.275 s, pynn_seed 995. The reference exists at 2 threads only. Identity with the pilot at 4 threads on the 2 mm model is not tested here; G4 tests it. The cortical match over 1.275 s does not promise one over 35 s: cortical signals still differ at about 1e-9 mV, and one threshold crossing that lands in another 0.1 ms step would start a real divergence. LGN cells get no cortical input, so that cannot reach the LGN. Rates here are from a tiny, high-rate run and are not biological evidence.