The buildbot-sheathing programme asks whether a robot can hang a 4x8 sheet of wall sheathing and land all 62 nails on the studs. Its paper for Automation in Construction quotes 194 numbers. The previous step added reproduce.py --check, which re-runs an experiment in a scratch copy and compares the fresh results file with the committed one, character for character. Twelve experiments went through it, and two had drifted. This step ran the other ten: the MuJoCo and Drake contact simulations and the wall pushover and cyclic-loading runs, about 2 h 20 min on four CPUs.
Eight matched exactly apart from wall-clock lines: the ledger seat (m3), the humanoids (m3b), the bowed sheet edge (m6b), the frame and soft-nail limit analyses (m15, m16), the wall pushover (m19), the CUREE cyclic runs (m20) and the strength-degrading nails (m21).
The current Drake can't run m6. The Drake cross-check compares two contact solvers, SAP and TAMSI. Drake 1.58.0, the version pip installs today, rejects discrete_contact_approximation="tamsi". The experiment was committed under 1.51.1, and its results.json records that version. With pip install drake==1.51.1 in a separate environment it runs. Nothing else in the project uses TAMSI, and the Drake tests and m6b pass on 1.58. The pin is now written into reproduce.py, the module docstring and the paper's data-availability statement. On 1.58 the run still stops with Drake's own error, but reproduce.py --list now names the version it needs.
m6 also drifted, for a cause we had already traced. One m6 table re-runs the 24 ledger pipelines with each engine's seated error. Its "no residual" column now equals the refreshed whole-cycle table (m5) line for line. So this is the same cause the last step found: a later milestone drew two random tilts in the opposite order. The paper's one number from that table, the worst gap between Drake and MuJoCo in clean-sheet rate, moves from 0.06 to 0.05 points. The pre-registered bar was 10, so the verdict is unchanged.
Two bookkeeping finds. In m9, one missed-nail mean printed as 0.48 and now prints as 0.47. The JSON shows why: the old formula, (1 − mean hit) × 62, gave 0.47500000000000164, and the per-sheet sum gives exactly 0.475. That's floating-point error at a rounding edge, and the paper doesn't quote the value. In m21, a diagnostics file ended in the word DONE, a wait marker an earlier session's shell had echoed into it. The script never prints it.
The check needed two fixes too. m6 reports seconds per seat as a bare table column, and three experiments end with wall time N s. Neither matched the timing patterns the checker masks, so a faster machine looked like a failed reproduction. Both patterns are now masked, with tests.
What this means for the paper. Every results file it quotes has been re-run from the committed code, and 194 of 194 quoted numbers match. Still open before submission: authorship, two standards' editions, and a tested end-nail strength. Code, paper and log: research/buildbot-sheathing.