8. The interpreter milestone, and every open bet
The milestone
Lex, parse, typecheck, and run Payments.Refund from chapter 4 with its tests. Contracts run at tier 2. rejects tests trip their requires. The verified: line is computed. No solver, no C backend, no real platform beyond Mo.Sim. Processes and supervisors can come in a second step, but the syntax must parse from day one.
This is where every syntax pick meets reality. What breaks gets fixed on the pick’s page as a Session note, never by rewriting history.
Session 5: the milestone is met
12 Sep 2026, four worker steps, about two hours of Opus time. mo test examples/payments/refund.mo lexes, parses, checks, runs six tests (four rejects tripping their requires; seven tests today), a property under 200 seeds, and the refund queue as a process under Mo.Sim, then prints the verified: line. All 52 corpus files behave as declared. What broke and got fixed on the way is in chapter 4’s Session 5 notes and decisions/decision-log.md. No solver, no C backend, sim (not run) until fault injection exists.
First tested by
Every provisional decision from session 3, and what tests it.
| Bet | First tested by |
|---|---|
| Anonymous functions are call-arguments only (d31) | the example corpus: how often an agent wants a stored closure |
| Loops without invariant syntax | tier 3 small-model checking on the corpus |
| Unbounded integers in contracts, sized in bodies | the tier-2 contract runtime, then the tier-3 proof rate |
Tier 3 v0 is testing, proving is mo prove (d32) | how far property tests and simulation get before anyone misses a prover |
| Bounded mailboxes, sender crashes (d33) | Mo.Sim with a slow consumer |
| Zig, and its incremental build claim | the toolchain’s own rebuild time |
| Packages as recipes (d34) | three recipes (the rate limiter, the store, the model client, under examples/recipes/) implemented by agents: token cost, defects against the recipe’s tests, attempts to exceed needs |
| The stdlib ceiling | stdlib misses on programs 2 and 4 of the menu |
| Law numbers (70, 6, 3, 12; the 500-line file law dropped 14 Sep) | lines-per-function distribution on the corpus |
| 50 ms and 100 ms targets | the benchmark suite, from day one |
| Spec altitude is readable | a timed reading study on program 1 |
The control run
On the program menu, agents writing Mo against agents writing Go, Python, and from round 10 Elixir, with the same checks bolted on (contracts via a verifier, tests, linters). Measured: time to green, defects found later, human review minutes, total token cost including spec and diagnostics. And a Quasar-style control: Mo’s checks over a Go or Python subset. If the subset wins, that changes the project.
Session 6 (Robert, 14 Sep 2026, after round 6): the measures above were the wrong ones. Agent time to green is not what Mo is for: no model has seen Mo in training, so an agent will take longer in it than in Go or Python for a long time, and rounds 1 to 6 measured mostly that. From round 7 the control run predicts, in this order: how many defects a hidden adversarial suite finds (reliability, the first column, chapter 1); how fast the program runs and how much memory it holds, native, against Go and Python; how fast the feedback loop is (check and test time on the program, mo build time); and the dependency count (third-party packages at run time, third-party tools at build time; Mo is 0 by construction, round 6’s Go was 1, Python 3). Agent time, loops, lines, and whether a check caught a bug stay recorded columns, not predictions. The null hypothesis of round 6 (that bolted-on checks do as well) stands as recorded and is not re-run; what remains to earn is whether Mo’s checks catch what tests do not, and that waits for a program built around invariants, the ledger.
Session 6, evening (Robert, 15 Sep 2026, on Fable’s direction 43): two diagnostic columns join every round from round 8 on, recorded, never part of the reading rule: tokens read per correct change (what an agent consumed from the codebase before its change passed the suite; the honest measure of readability for an agent, and the number spec altitude exists to lower) and first-fix rate per diagnostic (for each Mo diagnostic, how often the first fix after seeing it is right; a diagnostic under 0.7 with ten or more sightings is rewritten to say what to write). The reading rule stays the conjunction of chapter 1: reliability at zero dependencies.
Open questions, not yet asked
Held for after the milestone, one at a time: supervisor strategies and escalation; flow narrowing into refinement types; a linear resource kind; copyable capabilities; a spec-only contract kind for never clauses too costly to run; restart from init or from the last committed state; hot reload in development; a compatibility promise at v1; two versions of one recipe in a program; forbid reaches(A, B) as the general form of flows.