The control run, round 3
Round 2 (13 Sep, 09:25) put Mo at 11.8 minutes against Go’s 8.2 and Python’s 7.9, with the rest of the gap traced to mo fmt’s line breaking and five stdlib rows. Step 14 fixed all of it. Round 3 reruns the identical experiment on the toolchain after steps 14 to 16: same spec, same model, three fresh sessions in parallel, in worktrees from which the round 2 implementations have been deleted so nothing is copied.
Setup
| worker | worktree | branch | removed before start |
|---|---|---|---|
| Mo | ../mo-lang-control3-mo | control3-mo | examples/programs/logstat/ |
| Go | ../mo-lang-control3-go | control3-go | experiments/control-run/go/ |
| Python | ../mo-lang-control3-python | control3-python | experiments/control-run/python/ |
Briefs are word for word program-2 and control-run; the Mo worker writes to examples/programs/logstat/ as before. Agents mo-r3-mo, mo-r3-go, mo-r3-python. The branches are evidence and are not merged; the numbers and the Mo program’s gaps are what round 3 produces.
Result (13 Sep 2026, 13:58 to 14:21)
| Mo round 2 | Mo round 3 | Go round 3 | Python round 3 | |
|---|---|---|---|---|
| wall-clock | 11.8 min | 23.4 min, void: the machine slept mid-run; includes the toolchain build and the corpus test | 18.3 min, void: the same sleep | 7.3 min, done before the sleep |
| loops to green (code) | 0 | 3 (one real bug, two test mistakes) | 0 | 1 (own test wrong) |
| loops to green (tooling) | 0 | 6 (five syntax or law diagnostics, one stale verified: line) | 2 (staticcheck version, a shell cwd) | 0 |
| functions | 4 modules, median 3, max 12 | 56, median 6 (bodies 4), max 19 | 44, median 9.5, max 36 | 39, median 5, max 24 |
| program + test lines | 791 | 851 | 1,387 | 1,064 |
| checks that caught a real bug | 0 | 2: a test (busiest counts off by one) and a never (a half-built summary) | 0 | 0 |
| stdlib gaps hit | 5 | 1 (Fs.each_line cannot fold), plus 2 toolchain bugs | 0 | 0 |
| verified by Fable | yes, as is | yes, as is: four modules green, five runs against their expected files, mo fmt --check clean, the corpus test green in the worktree | yes: check.sh, go test, go vet | yes: check.sh, 69 tests (mypy not on Fable’s machine; the worker reports it clean) |
Reading. The wall-clock column is unusable for Mo and Go this round: Robert’s machine slept while both were running (the Mo pane shows the API error), and Mo’s time also holds a toolchain build and a full corpus run. The other columns stand. Mo’s loops went from zero to nine, and six of them are the toolchain refusing the model’s first guess at Mo syntax or a law: assert as a case arm, a one-field variant matched by name, a line wrapped inside a string, an assignment on a one-line arm, a _ arm on a closed type, a verified: line edited after --write. Each cost one run and was fixed on sight, and none was a behaviour bug, but the worker singled out MO0101’s wording as unhelpful; that is the cheapest thing to fix. For the first time a control run had a real bug for the checks to catch, and Mo’s caught it twice over: a test found the off-by-one and a never found the invariant break on a half-built value, which the worker then designed away. Go’s and Python’s caught nothing, having nothing to catch. Two toolchain bugs surfaced: any(T) ignores a refinement’s where (a property test generated a Percent of 65,535), which is a soundness bug in the property runner, and Fs.read_lines hands back a String that is not UTF-8. Both go to step 17 before program 4 builds on the same rows. Mo’s functions are still the shortest, and its program is 60 percent of Go’s length.