The control run, round 5

The first pre-registered round, as empirical-validation-plan asked: the thresholds below were written before any session started (14 Sep 2026, 01:12; the sessions started 01:14). Same spec, same model, three fresh sessions in parallel in worktrees from which the earlier implementations are removed, briefs word for word program-2 and control-run, so the round compares with round 4, the timing round. The toolchain since round 4: step 21 (green threads, three diagnostics for round 4’s loops, mo fix for the one-line if) and program 1’s evidence. Nothing in the briefs changed; the bolted-on-checks arm for the baselines (Go with a contracts library, Python with pydantic) waits for a round with a second task, since changing the baselines’ briefs would break the comparison this round exists for.

Setup

workerworktreebranchremoved before start
Mo../mo-lang-control5-mocontrol5-moexamples/programs/logstat/
Go../mo-lang-control5-gocontrol5-goexperiments/control-run/go/
Python../mo-lang-control5-pythoncontrol5-pythonexperiments/control-run/python/

Agents mo-r5-mo, mo-r5-go, mo-r5-python. The branches are evidence and are not merged.

Pre-registered

Three predictions, decided before the run; all three must hold for the round to count as “held”, any one failing makes it “mixed”, and the reading says which.

predictionthresholdround 4
P1, wall-clockMo at most 1.5 times Go1.73
P2, loops to greenMo at most Go’s plus 25 to 1
P3, the lawsno loop from a shape law, and none from the three forms step 21 gave diagnostics for0 law loops, 3 form loops

Two new columns, recorded for every worker: loops by cause (a law, a grammar form, a diagnostic, a test mistake, a real bug) and whether the worker wrote its language directly or wrote a generator for it (from the pane and the commits). Tokens: the pane’s final status line per session (Claude Code’s own count), amortized preamble not separated this round.

Result (14 Sep 2026, 01:14 to 01:31; no sleep, timing valid)

Mo round 4Mo round 5Go round 5Python round 5
wall-clock to the last commit16.1 min14.8 (17.1 to done, with the corpus test)11.614.9
loops to green55: 3 test mistakes, 1 grammar form (is inside == needs parentheses, MO0101), 1 law (assert and a fixture in a shared helper, MO0214 and MO0403)2 (an unused import and gofmt with a broken staticcheck; a layout bug and a test mistake)3 (1 code: a test mistake beside a real bug; 2 tooling, ruff’s flags)
functions61, median 3, max 1767, median 3, max 1447, median 10, max 3751, median 4, max 15
program + test lines8498311,5521,334
checks that caught a real bug001 (a test: a layout rule)1 (a test: a leaked descriptor)
stdlib gaps, toolchain bugs1, 13, 0 (fold_lines fails a file at its first non-UTF-8 line; Fs.list does not tell files from folders; no thousands separator)0, 1 (the installed staticcheck cannot read Go 1.27)0, 1 (mypy not installed, so no type check in any round)
wrote the language directly—yesyesyes
verified by Fableyesyes: zig build test green in the worktree, mo fmt --check clean, five run linesyes: check.sh, go test, go vetyes: check.sh, unittest

After the fact (14 Sep, morning, once mypy was installed as a uv tool): the round 5 Python passes mypy --strict with no issues in its 9 files, so its unchecked type hints were sound; the row above stands as the round ran. Tokens: not captured; the panes’ status lines had scrolled off by the end (Mo showed 54k output tokens at 13 minutes, Python 72k). The next round reads them before the report.

Reading, against the predictions

predictionthresholdround 5held
P1, wall-clockMo at most 1.5 times Go1.28 to the last commit, 1.47 to doneyes
P2, loops to greenMo at most Go’s plus 25 to 2no, by one
P3, the lawsno loop from a shape law or the three step 21 formsnone from eitheryes

Mixed. Two of three held. Mo’s wall-clock ratio fell from 1.73 to 1.28, the best of any round, on the timing round’s own terms; Go’s time rose to 11.6 (a broken staticcheck cost it a rebuild). Mo’s five loops are the same count as round 4 but a different shape: three were the worker’s own test mistakes (a star count, an argument order, a padding width), which Go and Python had one and one of; one was a grammar form the three step 21 diagnostics did not cover (is inside a comparison), and one was the honesty laws doing their job (assert outside a test, a fixture outside a test), which the worker fixed in one pass and named as a diagnostic that helped. No shape law tripped for the fourth round running. The three forms step 21 gave diagnostics for did not recur. No check caught a real bug in Mo this round; Go’s and Python’s tests each caught one, and in both cases the bug was in code Mo’s shape would have written the same way (a padding rule, a descriptor leak the platform owns). Mo’s program is half Go’s size and its functions a third the length, as before.

What changes. The laws stay; four rounds without a shape-law loop is the evidence Robert asked for on 13 Sep. The next ergonomics item is the is form inside a comparison, one diagnostic, to step 22. P2 was set on round 4’s ratio and missed by one loop on a count of five; the next round pre-registers loops by cause rather than a total, since test mistakes are the worker’s, not the language’s. Tokens are read before the report. The baselines’ bolted-on arm still waits for a second task.