The original stages ran on the authoring host on 2026-08-07; their wall-clock times are the seconds fields of results.json. This file starts with the 2026-10-19 correction (NOTES.md, "Correction").
2026-10-19: re-run with the relative cutoff (with hunts/jensen_clock)
Unit, timed on this host before spending. One DHFlow.H_and_dH evaluation: 0.035 s at dps 54 (pair 1, 1300 nodes) and 0.67 s at dps 165 (pair 9, 12096 nodes). A contour evaluation is 96 of them (3.4 s and 64 s); a landing time takes 8 to 12 contour evaluations, plus polishing the zero.
Multiply. The recorded stage times sum to 118 minutes on the authoring host: survey 78.2 (pair 9 alone 18.8), precision 6.9, pair1 4.3, lesions 1.9, validate 0.02, and the three jensen_clock stages 26.9 (additivity 9.5, shift 13.1, additivity2 4.3). On a GitHub runner at about half that speed, the longest chain (pair1 and precision, then jensen_clock shift + additivity2) is about 50 minutes wall, over 13 parallel jobs.
Where it ran. .github/workflows/dh-floor-rerun.yml, on push to the correction branch. mpmath pinned to 1.3.0 (the authoring host's), Python 3.12. Not run: ode and figure, which evaluate no DH weight and read only the polished zeros and the PDE t\* values.
Outcome. Run 39355357518 (commit 36d6b19f), all 15 jobs green, 25 minutes wall (01:47 to 02:12 UTC, 2026-10-20). The runner ran at about the authoring host's speed (pair1 245 s against 256 s recorded), so the estimate's factor of two was pessimistic. Its validate output equals this host's re-run bit for bit. Comparison: flow_repair 12 stages, 370 fields, 364 identical, 6 changed (validate's evenness defect, NOTES.md "Correction"); jensen_clock 4 stage blocks, 71 fields, all identical. Reports: rerun_2026-10-19.json here and in hunts/jensen_clock/.
2026-10-30: the v6 census sweep (NOTES.md section 7)
Unit. One pair, census_sweep.land: 0.7 s at height 240 (65 window zeros) to about 4 s near height 20,000 (113 zeros, 4 neighbouring pairs), one core. The pure-Python ode_null_control takes 7 s and 27 s on the same inputs.
Run. census_sweep.py --procs 16 on a 32-core host: 1,268 pairs, 203 s wall, 2,138 CPU-s. A first version re-found each window's zeros by sign scan and argument principle, as stage_ode does; near height 20,000 that cost 7 to 12 CPU hours a pair (16 pairs in 12 hours on 16 cores) and was stopped. Its 16 landings agree with the receipt-built ones to 2.6e-6 outside the three edge pairs. Those 16 landings are not committed, so this agreement cannot be rechecked from the tree (noted at landing, 2026-10-31).