# Reproduction record ## The run of record - **What was run:** a byte copy of this supplement as it stood when the run started (10 November 2026, 13:59 UTC), run into a separate output directory. Since then `README.md`, `SHA256SUMS.txt` and `experiments/` have changed, and `thmB/` and three files of `full/` (`cert2v2_v2main_M1800.json`, `v2_v2main_M1800.log`, `v3_v2main_M216000_NB50_NQ50.log`) have been added. On 13 November 2026 the two separately written implementations were replaced by corrected versions (`sep_p4/p4rig3.py` with `sep_p4/vconv.py`, and `sep_p5/p5rig.py`; see README.md), together with the four outputs they produced (`expected_small/sep_p4/cert3_M144000.out.txt`, `expected_small/sep_p5/rig_M36000.out.txt`, `full/cert3_M288000.out.txt`, `full/rig_M432000.out.txt`); steps 8 and 9 were then rerun (see "Steps 8 and 9 with the corrected programs" below). Apart from these, the code, inputs, `expected_small/` and `compare.py` that the run uses are unchanged. Of the added files, the run reads only `full/cert2v2_v2main_M1800.json` (compare.py's input-hash check), and the input hash it records is one that `full/cert4v3_v2main_M216000.json` already records, so the checks below are unaffected. - **Command:** `PY= bash run_all.sh.txt `, at nice 10, single-threaded, with a CPU-time limit per process. - **Software:** Python 3.14.4, numpy 2.5.2 and python-flint 0.9.0, on Linux. - **Result:** ALL REPRODUCED. All 13 steps exited 0 (the certificates at small resolution, the caps pass, and the four recomputations of the audits); all 13 stored outputs were reproduced; and all 22 checks passed: the stored full-resolution outputs against the margins stated in the paper (and the caps output's sup F < 4), the decoded input pickles against the hashes recorded in those outputs, and the plain-text exports of the inputs against the pickles. - **Header line:** "run_all.sh.txt started 2026-11-10 13:59:27 UTC; python 3.14.4 numpy 2.5.2 python-flint 0.9.0". ## How the comparison treats elapsed times compare.py ignores elapsed times: the leading "[ 12.3s]" stamps of text outputs, the times printed inside lines ("done in 79.1s", "(9.4s)"; six such fragments occur in the stored outputs, all of them elapsed times), and the "seconds" field of JSON outputs. Controls on a copy of the stored (1/18, 1/6) separate output: changing a computed value on a line that also carries a time ("WK max width", and an enclosure endpoint on the "(9.4s)" line) is reported as a difference; changing only the time is not. ## Elapsed time per step (monotonic clock, /proc/uptime; the machine was shared, so the times depend on its load) | Step | Output (under OUTDIR) | Exit | Elapsed | |---|---|---|---| | 1 | exact1/runb_e11_M13200.log | 0 | 5.6 s | | 2 | exact1/runb_e14_19c_M53760.log | 0 | 61.1 s | | 3 | exact1/run4v3_M7200.log | 0 | 14.0 s | | 4 | engine2/v2_p5frozen_M3600.log | 0 | 128.2 s | | 5 | engine2/caps_p5frozen_M3600.log | 0 | 129.9 s | | 6 | engine2/dev_g93adjQ32_M960.log | 0 | 1649.9 s | | 7 | engine2/dev_p7r_M960.log | 0 | 2198.0 s | | 8 | sep_p4/cert3_M144000.out | 0 | 168.7 s | | 9 | sep_p5/rig_M36000.out | 0 | 264.9 s | | 10 | audit/p3a/spot_least_cell.log | 0 | 0.4 s | | 11 | audit/p5/p5_independent.log | 0 | 0.1 s | | 12 | audit/p6/p6_independent.log | 0 | 0.1 s | | 13 | audit/p7/p7_independent.log | 0 | 0.2 s | | | total | | 4621 s | ## The shipped scripts against the scripts as run The shipped scripts differ from the scripts that produced the stored outputs only in comments, docstrings and the declared path lines (README.md). Their behaviour was also compared with earlier runs of the scripts as run, at small resolution, before the edits: - the exact certificates for (1/11, 1/10) at M = 13,200 and (1/14, 1/8) at M = 53,760 reproduce the earlier outputs in every field except the elapsed time; - the exact certificate for (1/18, 1/6), run with the earlier run's block counts (NB = NQ = 32) at M = 7,200, reproduces that run's output in every field except the elapsed time; - the second-order engine for (1/20, 1/5) at M = 3,600 reproduces the earlier output in every field except the two that record the SHA-256 of the engine and of dpl.py, which differ because those files were edited; - the engine for (1/30, 1/4) at M = 1,920, with the full run's settings, reproduces every field of an independent audit's replay of the engine as run, again except the two hash fields; - the separately written implementations reproduce their earlier outputs at M = 144,000 and M = 36,000 line by line, apart from elapsed times; - the caps pass for (1/20, 1/5) at M = 3,600 reproduces every field of an independent audit's replay of the pass as run, except the two hash fields; - the four scripts in audit/ reproduce the audits' own outputs exactly, except one field of spot_least_cell.json that records the path of its input. ## Steps 8 and 9 with the corrected programs (13 November 2026) - **What was run:** a byte copy of this supplement as it stood after the corrected programs and their outputs were put in, with the two commands of steps 8 and 9 of `run_all.sh.txt` run in copies of `sep_p4/` and `sep_p5/`, at nice 10, single-threaded, with the same settings (`PY=`, one thread for the numerical libraries, no bytecode files). The other eleven steps were not rerun: their code, inputs and expected outputs are unchanged. - **Software:** Python 3.14.4, numpy 2.5.2 and python-flint 0.9.0, on Linux. - **Result:** both commands exited 0. `compare.py`, run on that output directory with `inputs/` copied and decoded as `run_all.sh.txt` does, reports SAME for `sep_p4/cert3_M144000.out.txt` and `sep_p5/rig_M36000.out.txt`, MISSING for the outputs of the eleven steps not rerun (its eleven failures), and PASS for all 22 checks of its other sections, among them the two replaced full-resolution outputs against the margins stated in the paper (printed outward as 0.04587 and 0.01609). - **Elapsed time:** step 8 took 165.8 s (monotonic clock). Step 9 was paused by the machine for part of its run, so its elapsed time (1455.8 s) is not a fair run time; it used 254.4 s of CPU time (GNU `time`; peak memory 550 MiB). A first attempt at step 9 was stopped by the same pause and ended with the shell that ran it; it is not counted.