The control run, round 2

Round 1 (13 Sep, 00:18) put Mo at 25.5 minutes against Go’s 12.9 and Python’s 8, with every minute of Mo’s loss traced to the toolchain: no multi-module programs, no stdlib, a slow interpreter. Steps 7 through 12b fixed all of it. Round 2 reruns the identical experiment on today’s toolchain: same spec, same model, three fresh sessions in parallel, in worktrees from which the round 1 implementations have been deleted so nothing is copied.

Setup

workerworktreebranchremoved before start
Mo../mo-lang-control2-mocontrol2-moexamples/programs/logstat/
Go../mo-lang-control2-gocontrol2-goexperiments/control-run/go/
Python../mo-lang-control2-pythoncontrol2-pythonexperiments/control-run/python/

Briefs are word for word program-2 and control-run; the Mo worker writes to examples/programs/logstat/ as before. The branches are evidence and are not merged; the numbers and the Mo program’s gaps are what round 2 produces.

Result (13 Sep 2026, 09:25 to 09:37)

Mo round 1Mo round 2Go round 2Python round 2
wall-clock25.5 min11.8 min8.2 min7.9 min
loops to green (code)2001
loops to green (tooling)602 (staticcheck, gofmt)1 (mypy hints)
functions784 modules, median 3 lines, max 1246, median 10, max 31median 5, max 19
program + test lines1,1747911,386983
checks that caught a real bug20 (nothing fired)00 (own test wrong)
stdlib gaps hit9550
verified by Fablewith the file law liftedyes, as isyesyes

Verification: Mo’s 30 tests pass from the four source files, mo run over the fixture matches the expected output, and the worker derived its expected files from an independent Python calculation rather than from mo run. Go and Python pass their checks and tests.

Reading. Mo’s time halved and every failure mode of round 1 is gone: zero failed runs, no toolchain workaround, no lifted law. It is still about 45 percent slower than Go and Python, and the worker’s report says where the time went: mo fmt produced awkward shapes it reworked three times, and five stdlib rows were missing (sort_by descending, two-value min/max, a streaming read). Mo’s functions are a third the length of Go’s. No language’s checks caught a bug this round, because there were none to catch: the spec is now well understood by the model after three readings across rounds, which is itself a finding about spec-driven work. The remaining gap is toolchain ergonomics, and it is measured in single-digit minutes.