Outside review, 13 Sep 2026: evidence
Every probe below was run against toolchain/zig-out/bin/mo built from main on 13 Sep 2026 (zig build test green, 3m52s). The conclusions are in outside-review-2026-09-13.
P1. Recursion is unbounded and uncaught
fn forever(n: UInt64) : UInt64
forever(n)
endmo check exits 0. mo run prints a Zig stack trace ending in start.zig:190 and exits 134. spec/design-v0/02-laws.md says “bounded everything”; spec/design-v0/03-semantics.md says recursion is “structural only, proved terminating; to be confirmed.” Nothing enforces it, and the failure is not a Mo crash report but a runtime abort.
P2. The for-only law produces fictional bounds
examples/programs/kv/server.mo L105–128: talk runs for _ in 0..10_000 around talk_awhile, which runs for _ in 0..10_000 around exchange. Comment: “a connection gets at most 100 million lines.” The accept loop in the same file uses the same shape. The law asked for a bound; the program gave it a number that bounds nothing.
P3. invariant polarity
invariant "count never exceeds 3"
state.count <= 3
endOne Bump then Read:
FAIL test "bump once": invariant "count never exceeds 3" tripped in Counter; state = Counter(count: 1)
The block must be true when the invariant is broken (spec/grammar.md). The natural reading fails on the first valid state.
P4. Fault injection and vacuous assertions
examples/programs/kv/log.mo L309–321:
assert first is Ok(Ok(108)) or first is Ok(Error(_)) or first is Error(_)server.mo L197–202, heard_so_far?: Error(_): true, and Ok(lines) passes for an empty list. Neither is a strict tautology (a wrong byte count still fails), but both pass with zero successful operations, so the test titles overclaim.
P5. Exhaustiveness without grouping
log.mo L171–181, applied: six arms with the identical body Error(BadLine(number: number)). The no-catch-all law (MO0309) is right; the missing piece is A | B | C: body.
P6. Handles are not capabilities
toolchain/src/caps.zig L140–159: capIn recognises .cap inside list/option/set/result/map/tuple, nothing else. Handle(T) is not .cap. server.mo answer(store: Handle(Store), ...) performs store.ask(...) with no capability parameter, so the premise’s “no capability param ⇒ pure” is false for any function holding a handle.
P7. Closures may capture and use a capability
fn shout(out: Out, xs: List(String)) : UInt64
xs.map(fn(x) out.write_line(x) end)
xs.size
endChecks and runs (prints a, b, 2). “Captures are read-only” is true; “captures are effect-free” is not. The discarded map result is not flagged either.
P8. update is not a transaction across effects
toolchain/src/sim.zig L363–383: ask delivers to the target and runs its update before returning to the caller; a later crash in the caller cannot undo that. Comment on L366: “A Timeout’s message still arrives.” toolchain/src/server.zig L276–301: fs.write fsyncs, then checks lateness; the doc comment says “a write that took longer is Timeout, and is on disk all the same.” Restart empties the mailbox (sim.zig L378 comment) and reruns init; examples/programs/kv/store.mo L95–99 sets restart: :never because a restarted store “would forget every change since.”
P9. The durability never checks self-reported witnesses
store.mo L9–13 checks Step.all for s.after != s.before and !s.logged. Step is built by the implementation at L208–209 with literal logged: true after a successful Append. An implementation that skips the journal and still writes logged: true, or never constructs a Step, passes. Useful instrumentation; not independent evidence.
P10. Deadlines as literals
kv/*.mo: within: 60_000.ms, within: 1.minute, within: 5_000.ms, within: 1_000.ms, within: 100.ms across calls with no derivation from a request budget.
P11. Control run, round 2 (repo’s own numbers)
control-run-2: Mo 11.8 min, Go 8.2, Python 7.9; zero check-caught bugs in any language; Mo functions shortest (median 3 lines) under a law that makes short functions compulsory. The Quasar-style control (Go/Python subset plus Mo’s checks) from case-against-new-languages has not been run.
P12. Worker defaults now in the semantics
decision-log: zero-value state fields; order-sensitive equality for maps and sets; write-after-timeout behaviour; contracts-off default (overturned only at step-13 acceptance). Each was implemented first and ratified after.
P13. Roc precedent for “the platform provides all I/O”
Librarian survey of roc-lang/roc and roc-lang/examples: two general-purpose platforms carry the ecosystem (basic-cli, basic-webserver); docs/langref/platforms.md still says there is no official platform-authoring guide; a host must supply roc_alloc/roc_dealloc/roc_realloc/roc_dbg/roc_expect_failed/roc_crashed and per-target link inputs; the Task effect type was removed in Jan 2025 in favour of direct-style ! functions with purity inference. Application code has no FFI; every unsupported native library becomes platform work.
What worked well in the probes
Diagnostics are the best part of the toolchain: multiple diagnostics per run, MO0301 (70 lines) and MO0309 (catch-all) both carry a correct “why”, --json is clean, mo fmt --stdout is idempotent, named functions are first-class (f = double), test rejects trips a requires with a readable report, process crash reports name the process, seed, message history and prior state.