Comparison synthesis (draft)
A draft for Fable to finish, written by the worker session at the end of the comparison pass (comparison-pass). Nothing here is decided. Robert’s rule: these pages inform the v0 design doc, and nothing changes a direction until he says so. Evidence and citations live on the thirteen pages linked below. This page only collects.
Contradictions and tensions (⚠️)
Ordered by how much of Mo each one touches (the worker’s judgment).
- Direction 15’s “the signature is the purity proof” vs closure capture (koka). An anonymous function can capture a capability, so a function with no capability parameters can still do I/O through a closure argument. The same hole would let a “no capabilities” package act through callbacks (Q17). Options:
- Pick 11 (bounded
for) vs Q8 tier 3 provingrequires/ensures/never(bosque). Loops need invariants. Partial answer from spark-ada-and-dafny: LLMs plus solver counterexamples solve benchmark invariants (133/133, 99.5%), but no real-program evidence yet. - Q4 (sized integers only) vs direction 22’s Dafny-level (~82%) expectation (spark-ada-and-dafny). Vericoding blames part of Verus’s 44% on machine-integer obligations. One option: unbounded integers inside contracts only.
- Q8 (SMT prover in tier 3) vs direction 24 / Q11 (Zig as the only dependency, zero third-party deps) (austral, bosque). Z3 is about 300K lines of C++. Mo must write a solver, vendor one, or relax the rule.
- Q7 (“
sendnever blocks or fails”) vs direction 4 (bounded everything) (go). “Never blocks or fails” implies an unbounded mailbox. A bound meanssendblocks or returns rain. - Q1 (no doc-comment variant) vs tested doc examples (elixir). Weak:
testblocks may already cover it. - Q13’s stated reason (roc). The “100x faster incremental builds” figure is not yet available on stable Zig. The decision’s other reasons stand. This is a correction to the why, not the decision.
Gaps several pages hit (not contradictions)
- Trait coherence is undefined (rust, roc, hylo). Hylo’s rule is a ready answer.
- Crash semantics are underspecified (elixir, austral, koka, verse): supervisor strategy, escalation, what a crashed process leaks, and whether a restart starts fresh or keeps state.
- What tier 3 is (elixir, bosque, spark-ada-and-dafny): counterexample search or proof, whether “unknown” counts as red, and how to stop agents from gaming it.
Top ten steals
The worker’s ranking, by leverage on the v0 design.
updateas a transaction. On a crash, discard that message’s state writes and its buffered outgoing effects. Freezeclock.nowperupdate. The runtime reclaims the process’s memory and handles, possibly as a per-process heap region. From verse, austral and koka, for direction 14 and direction 18.- Define tier 3 precisely. Counterexample search over a decidable fragment; “unknown” isn’t red; loop invariants come from a generate-and-check loop with the solver; a change that edits a spec while its proof fails is rejected. From bosque, elixir and spark-ada-and-dafny, for Q8.
- An honest
verified:line. A per-obligation record whose headline is the weakest obligation, plus a Silver-style entry, “no runtime errors proven (k of n)”. From agent-native-cluster (Thermite) and spark-ada-and-dafny, for Q6. - Hylo’s coherence rule. One
implper type/trait pair; apubimpl only in the type’s or the trait’s module. From hylo, for pick 15. - Content hash and caches. Unison’s hash recipe (names normalized, dependency hashes substituted) per declaration, and a typecheck/test/proof cache that is never invalidated. Proofs are keyed by callee contract hashes. From unison, spark-ada-and-dafny and roc, for Q10 and pick 12.
- The agent interface is the toolchain. An MCP server over edit-by-ID, diagnostics and history, plus version-pinned guidance served by
mo. No Mo-specific agent. From unison, agent-native-cluster and moonbit, for Q9 and Q10. mo fixfrom day one. Every language change ships with a rewriter, and every published package version carries an old-hash → new-hash patch. From go and unison.- Diagnostics that teach. Caller/callee blame, both conflicting sites, “overlapping
inoutaccess”, a no-false-positives bar, and tier-3 counterexamples delivered as structured fixes rather than raw prover output. From agent-native-cluster, rust, hylo, elixir and spark-ada-and-dafny. - An eval suite shaped like the real job. SWE-AGI-style tasks (fixed
pubAPI with contracts, hidden tests), whole-program tasks, false-positive rate, p50/p95 time to diagnostics after a one-declaration edit, and p99 latency. From moonbit, rust, elixir and bosque, for direction 28. - Determinism laws for replay. Stable sort and defined map order in the stdlib; virtual time in
Mo.Simadvances only when every process is blocked; one clock value perupdate. From bosque, go and verse, for direction 21.
Runners-up:
- a platform manifest in Roc’s five parts (roc)
- a spec-altitude shape diff (moonbit)
Platformnever leavesmain(austral)- a
fip-style “allocates nothing” check (koka) Task.runbound to its scope, withracecancellation for deadlines (hylo, verse)
Q17 inputs, not recommendations:
- locked builds, a transparency log, no code execution at fetch or build (go)
- a lockfile audit of platform native code (austral)
- launch-time capability grants (agent-native-cluster)
- the BoltDB lesson: review the hashed bytes, not the VCS view (go)
Open questions for Robert
Collected from the pages, for Fable to order and ask one at a time.
Types and syntax
- Supervisor
strategy:field, and escalation whenmax_restartsis exceeded (elixir) - Flow narrowing into refinement types inside
if/case(elixir) - Restrict
tryto the head of a binding or statement (go) - Named traits vs Roc’s per-method
where; coherence rule (rust, roc, hylo) - Mark mutation at the call site: nothing,
&v, orinout v(hylo) - Nested place paths as the only projection (hylo)
- Recursion: ban, bound, or allow (koka)
- A linear
resourcekind for must-consume values (austral) - Copyable capabilities, or not (austral)
- A spec-only contract kind for
neverclauses too costly to run (bosque) - Keep tests and
ensuresuntouched when an agent changes a body (spark-ada-and-dafny)
Runtime and tooling
- After a crash: restart from
init, or resume from the last committed state with the poison message set aside (verse) - Hot reload in development only, via the interpreter (roc)
- Ship an agent, or tools only; recommendation: tools only (moonbit)
- A compatibility promise at v1, with a per-module language version (go)
forbid reaches(A, B)as the general form offlows(...)(agent-native-cluster)
Packages (Q17)
- Two versions of one package in a program? (unison)
- Launch-time capability grants on top of in-code narrowing? (agent-native-cluster)
17 Sep 2026: this list is stale — recursion, crash-and-restart, hot reload, and ship-an-agent are all settled, and the “Session 3 answers” below already decide several of the ⚠️ rows.
Cross-cutting observations
- Mo’s verification design is now the consensus of the agent-native verification camp: mandatory contracts, a solver, runtime fallback (agent-native-cluster). Differentiation has to come from spec-altitude readability, processes with supervision and replay, capabilities as package permissions, and the laws.
- Direction 18 got independent support. Verse makes runtime errors uncatchable and separate from failure (verse). MoonBit hit bugs from a catchable cancellation error (moonbit). Austral agrees on crashing but crashes the whole program (austral).
- Zero-shot code in a new language is near 0% on hard tasks, yet agents with a toolchain loop build real systems (moonbit). Raw verifier errors teach models poorly (spark-ada-and-dafny). Q9 carries the premise.
- Perceus-style native compilers ship (Koka, MoonBit, Roc), so direction 10’s memory bet has precedent (koka, moonbit, roc).
- Two claims need a primary-source check before the design doc: MoonBit’s reported built-in verification (moonbit), and Q13’s rationale (roc).
Notes for Fable
- All 13 pages are
confidence: medium. The agent-native cluster page uses short per-language entries instead of the idea template, as the brief allowed. - The shared ledger
research/comparisons/.ledger.jsonholds 152 sources, and ids were never renumbered. Three are uncited: [50], [55] and [87] — fetches that returned junk, deleted before commit. (Counted and corrected 17 Sep 2026.) Where a fetch cut off before a cited section, a-part2or section excerpt was saved toraw/. - Two pages were edited after they were first committed:
- Roc: its effect-polymorphism line now points to the closure gap on Koka.
- MoonBit: the IEEE TSE acceptance is marked as a secondary-source claim.
- No page under
directions/,questions/orsyntax/was changed.
Session 3 answers
- Closure capture: in on option (c), Bosque’s rule → direction 31.
- Loops vs tier 3: in on keeping loops with no invariant syntax; tier 3 does small-model checking, unproven loops show as “tested, not proven”.
invariantline deferred until measured. Noted on Q8 and pick 11.
Robert then asked to stop speculating on untested issues and start building; tensions 3–7 are Claude’s provisional decisions at his request (“your decisions are as good as mine”), each marked with what first tests it.
- Sized ints vs proof rate: unbounded math in contracts only, two obligations on the
verified:line. On Q4. - Prover vs zero deps: tier 3 v0 = property tests + simulation in the interpreter; proving is a separate
mo provebinary with a vendored solver → direction 32. On Q8. - Mailboxes: bounded, declared per process,
sendnever blocks, overflow crashes the sender → direction 33. On Q7. - Doc examples: no change,
testblocks are the examples. On Q1. - Q13 rationale corrected, decision stands. On Q13.