# G4 gate, written before the run starts (7 Nov 2026)

G4 reruns protocol 022's intact pilot on this Mac: the pilot's exact command, seed and sequence, on build NM with `run_controller_mem.py`. Pass conditions are fixed here, before any result exists.

## Must hold for the Mac route to continue
1. The run exits 0 and memcap did not stop it.
2. Peak physical footprint is at most 3.515 GiB, which is 5 % under the 3.7 GiB cap.
3. All 3,600 LGN spike trains (1,800 X_ON, 1,800 X_OFF) equal the pilot's export over the whole 35,133 ms. Pilot export SHA-256: 952b469b9a8e74a482f704854a0b5462f6c68798ba3466e1a15aa0e893eae281.
4. The frozen `analyze_plaids.py` accepts the export, with the lesion marked FAILED, and each primary measure lies within 3 pilot standard errors of the pilot's value: XOS 0.047, Dc 0.100, ds 0.074. Standard errors come from the pilot's 95 % intervals (width / 3.92). The tolerance is descriptive: it sizes and sets nothing in protocol 023.

## Reported, not a pass condition
- Cortical trains identical to the pilot's, per sheet, and the first differing time if any. The first 2,289 ms were identical in all 18,600 neurons, but rounding differences in a chaotic network may still appear later.
- Run time, memory curve, swap, free disk, machine load.

## If a condition fails
- Condition 2 fails below the cap: add R5, then R6, each with a byte-identity proof, and repeat the measurement.
- Condition 3 fails: find the first differing spike. The exact lead-in check needs only the first 504 ms, so the protocol can still run if that part holds, but the Mac-to-server link is then statistical from that time on, and the spec must say so.
- Condition 4 fails: stop and report. The Mac run would then not reproduce the pilot.
