Machine: the shared swarm box (32 vCPU, 115 GiB), 1-minute load average 58 to 216 during these runs (other workflows), MemAvailable above 77 GiB throughout. Every run is one single-threaded process, under 1 GB, checkpointed per window c into units/<variant>_<kind>.json (a rerun skips finished windows). At most seven of these processes ran at once, within the cap of eight. All times are wall-clock on that loaded box: measured, not pinned by any test, and only good for the next estimate.
| run | unit timed first | estimate written before launch | actual (per window, then total) |
|---|---|---|---|
ghost.py arb dh (primary, 7 c) | c = 29 in the scratchpad probe: 6.5 s ground state, under 3 s counting | about 30 s per c, 4 min | 22 to 37 s, 206 s (after one void launch, below) |
ghost.py arb zeta (7 c) | c = 30: 5 s at 900 bits, plus the 3200-bit LDL steps | about 1.5 min per c, 10 min | 52 to 90 s, 497 s |
ghost.py windows, first version (Z_epstein scan) | one Z_epstein evaluation: about 10 s on the loaded box | 450 evaluations: over an hour | stopped after about 12 min; replaced (RESULTS s0) |
ghost.py windows, PARI version | lfunzeros: 0.01 s; confirmations 10 s per Z_epstein evaluation | about 10 min | 117 s (Epstein), 544 s in all |
ghost.py arb dh --variant=q1, sham, scan, scanfine | as primary | 30 s per c, 12 min for all four | 16 to 29 s, 162 s; 19 to 28 s, 163 s; 15 to 23 s, 116 s; 17 to 24 s, 61 s |
ghost.py float dh (dps 60) | its first window, c = 29 | written after that unit: 3 min per c, 21 min | 138 to 180 s, 1088 s |
ghost.py float zeta (dps 320) | scaled from float dh by the cost of dps 320 arithmetic (about 3x) | 9 to 15 min per c, 1 to 1.75 hours for 7 c | 301 to 464 s per c |
ghost.py arb epstein, dedekind (primary and prec2) | scaled from the stored Epstein scan (4 s per cell at N = 128) plus the counter | 1 min per c, 30 min for the four runs | 19 to 28 s, 12 to 22 s, 27 to 42 s, 18 to 26 s; 698 s in all |
ghost.py float epstein, dedekind (dps 60) | as float dh | 3 min per c | 127 to 212 s; 79 to 145 s per c |
ghost.py arb * --variant=prec2 (dh, zeta) | as primary at 1.6x the precision | 1.5x primary | 25 to 37 s, 203 s; 66 to 75 s, 482 s |
ghost.py arb dh 29 31 47 --variant=N64, N192, N256 | N = 64 first | N^3 scaling from N = 128: 5 s, 1 min, 3 min per c | 7 to 15 s; 39 to 56 s; 76 to 180 s |
ghost.py gap2, s2, resolution | reuse of the primary ground states | under 1 min per c | 7 to 17 s per c (gap2); resolution 1 to 3 min per diagnostic |
Void launch, kept on record. The first ghost.py arb dh launch produced two void units (c = 31 and 63/2): the residual of the shifted inverse iteration reached rounding level, its ball contained 0, sqrt of the residual's dot product returned NaN, the eigenvector radius eta became NaN, and every sign and every segment of those units was undecided (the unit printed "ghost = False" with no sign change at all). Fixed by ghost.residual_upper (square root of the sum of squared upper bounds) plus a finiteness guard; the two units were discarded and the whole DH run repeated; test_ghost.py pins the fix.
Stop and resume, kept on record. The campaign's workflows were stopped at the usage limit at 18:10Z on 2026-10-21, after the last float unit (float_epstein, c = 47) had been written but before ghost.py summary had merged the float units. The lane was resumed at about 21:20Z by a second agent in the same worktree, after checking that no process was still writing it (no file had changed since 18:03Z). No unit was recomputed; summary was rerun, which completed S7, and the prose was revised against the merged table. The resume ran only the summary and the lane's tests (single process, under a minute of compute besides the tests).
Failed launch, kept on record. The first ghost.py float epstein and float dedekind launches stopped at their first window with a TypeError: zeta.epstein.log_derivative_coefficients takes a callable, and was handed a list. Fixed in _float_epstein; both runs repeated from their first window.