Reading pack for an outside reviewer, September 2026

For a reader from the Erlang and OTP world, and one who has built a capability system (Pony, Wyvern, Newspeak, E). One sitting, in this order; the appendix is there so a reviewer does not re-suggest what was already tried. Everything is in this repository; nothing is private.

Read, in order

  1. The thesis: spec/design-v0/01-premise.md. Three layers, the runtime first; the BEAM as the null hypothesis. Ten minutes.
  2. The failure model: spec/design-v0/03-semantics.md, the “What a crash discards” section onward. What a timeout leaves, what restart loses, when a reply is durable, poison, overload. This is the page an OTP reader should attack.
  3. Capabilities and packages: spec/design-v0/06-packages.md, then directions/d34-packages-are-recipes.md and research/concepts/ on supply-chain attacks. This is the page a capability-systems reader should attack: authority as a parameter, obtained only at main, recipes at spec altitude, bricks as audited platform code.
  4. The two reviews so far: outside-review-2026-09-13 with outside-review-2026-09-13-response, and outside-review-2026-09-14 with outside-review-2026-09-14-response. The second reviewer’s replies are summarized in the decision log rows of 14 Sep afternoon.
  5. The evidence: control-run-6 (the round Mo lost on the old measure, all four predictions) and control-run-7 (the round on the new measure, all four held, reliability level). Read the Result tables, not the prose, first.
  6. A program: examples/programs/jobq/ (nine modules, the durable job queue; eleven by 16 Sep, after the changes) with examples/README.md’s two rules. Read queue.mo’s update, the nevers, and the verified: lines; skip bodies as Robert does.
  7. The working agreements: mo-wiki/SCHEMA.md, the “Working agreements with Robert” section, and the decision-log conventions. The process is part of what is under review.

What we want from you

  • OTP reader: where does chapter 3 promise something the BEAM learned not to promise? Where is a deadline on every wait, a bound on every mailbox, or “a bug crashes the process” going to hurt a long-running service in a way our programs have not shown? Is the runtime surface (directions 37 and 40) something you would use, or a debugger by another name?
  • Capability reader: is main as the only source of authority, with narrowing (fs.read_only, scoped) and the checker following a capability through parameters, start arguments, message fields, and branches (MO0404), sound? Where does it leak? Is “a recipe’s bodies are generated and checked against its tests” a package model or a hope?
  • Both: what would you measure that rounds 6 and 7 did not?

Appendix: tried and rejected

Pulled from the decision log’s overturned, dropped, and narrowed rows, so they are not suggested again.

triedwhenwhat happened
Memoizing pure calls in the interpreterstep 12reversed in step 12b: the reference interpreter runs every body every time
Checking never blocks only under --simstep 12reversed in 12b: every never runs at the end of every test
Contracts off by default in mo buildstep 13overturned on acceptance: contracts run in every build, --no-contracts for measurement only
A handle can never live in a valuestep 18narrowed in step 24: a process’s state field may hold a Handle(T); structs, enums, and returns still cannot
No timers: a lease is a deadline checked at the next lookprogram 1kept, and send(msg, delay:) added in step 24 after three programs asked
The one-line if refused, with a mo fix back to the block formstep 21reversed in step 25: the one-line if is a value (pick 16); step 26 made it the value in tail position
state, result, old as reserved words everywherev0narrowed in steps 25 to 27: keywords only in the positions the grammar reserves
A 500-line file lawv0dropped after round 6: a shape law cost a loop and no check earned one; the honesty laws forbid warnings, so it went rather than softened
A never reading every value a run held, a var between statements includedv0narrowed in step 27: a never reads values at rest
Agent time to green as the control run’s measurerounds 1 to 6dropped by Robert after round 6: no model has seen Mo; the measure is reliability, speed, the loop, dependencies
”Agents write better code with the compiler as teacher” as the premisev0demoted in session 6 to the language layer’s hypothesis; the runtime is the thesis
Rust or Gleam braces, def/end, -> arms, :: paths, a 40-line function limitsessions 1 to 3rejected by Robert on sight; fn ... end, case ... end, Module.Path, 70 lines
Phone-fit as a design criterionsession 3rejected firmly; never used since
Notion as the vaultsessions 1 to 2retired; the Obsidian vault in mo-wiki/ is the record
A registry with third-party packages as codeQ17not built; packages are recipes, code comes from bricks or is generated against the recipe’s tests