Comparison pass: brief for the worker session

Thirteen pages in research/comparisons/, one per shortlist entry in language-landscape. Written by a worker session (Opus) with fresh web research; reviewed by Claude (Fable) and Robert. This page is the full brief. Robert’s rule: nothing here is decided until he says so — these pages inform the v0 design doc, they don’t change directions on their own.

Orientation (do first, every session)

  1. Read HANDOFF.md, SCHEMA.md, index.md, the tail of log.md.
  2. Read research/concepts/language-landscape.md (the shortlist and why each entry is there) and deep-dives/steal-list.md (what session 1 already took).
  3. Skim directions/ titles (30) and questions/ titles (17) so you know which Mo decision each language touches. index.md has one line per page.

Tools

  • Web research: source ~/.zshenv && python3 tools/exa.py search "..." -n 10 / contents URL / answer "..." / research "task". Also plain WebFetch/WebSearch. Prefer primary sources: language site, spec, repo, papers.
  • Papers: arXiv API (https://export.arxiv.org/api/query?search_query=...) — wait ≥3s between calls, it rate-limits hard. Semantic Scholar https://api.semanticscholar.org/graph/v1/paper/search?query=....
  • Citations: python3 .claude/skills/grounded-citations/scripts/sources.py --ledger research/comparisons/.ledger.json add URL --title "..." returns [n]; cite inline as [n]; finish with render --cited-in PAGE >> PAGE and verify PAGE. One shared ledger for the whole pass; never renumber.
  • Raw sources: save any page or paper you lean on to raw/articles/<slug>.md or raw/papers/<slug>.md with frontmatter source_url, ingested: YYYY-MM-DD, sha256 of the body. raw/ is immutable once written.
  • Lint: python3 tools/lint.py must be clean before every commit.
  • Commit after each finished page: git add -A && git commit -m "Comparison: <Language>" && git push. Commit trailers per the session’s system reminder.

Page template (research/comparisons/<slug>.md)

---
title: "Mo vs <Language>"
created: 2026-09-12
updated: 2026-09-12
type: comparison
tags: [research, <one or two from the taxonomy>]
sources: [raw/articles/<slug>.md, ...]
confidence: medium
---
 
# Mo vs <Language>
 
**One line:** what this language is and why it's on the list.
 
## What it is (status as of Sep 2026)
Release stage, activity, who's behind it, domain. 3–6 sentences, cited.
 
## The ideas, one by one
For each distinctive idea (3–7 of them):
- **Idea** — what it does, in two sentences with a ≤8-line code snippet if syntax matters.
  - *Mo today:* which direction / question / syntax pick already covers this (link it: `[[d15-effects-via-capabilities|direction 15]]`).
  - *Verdict:* **steal** / **already have** / **reject** / **open** — and one sentence why.
 
## What it gives up
The costs the language accepts, and whether Mo accepts the same ones.
 
## Evidence
Anything measured: performance numbers, adoption, study results, LLM benchmarks. Cited. "No evidence found" is a valid line.
 
## What Mo should take from this
Bullet list, each item a concrete proposal or a question for Robert. Mark anything that would **contradict an existing direction** with ⚠️ and name the direction — do not resolve it, surface it.
 
## Related
- [[language-landscape]]
- links to the directions/questions/picks this page touched (≥2)
 
## Sources
(rendered by sources.py)

Keep pages under ~180 lines. Snippets under 10 lines. Three-column tables max. Phone-readable.

The thirteen, with what to look for

  1. elixir — Elixir v1.20’s gradual set-theoretic types (inference-first, no annotations, “verified bugs”); OTP supervision shapes vs Mo’s declared supervisor (Q7); what Elixir 1.20+ shows about typed BEAM ergonomics. Touches d7, d14, d18, q7.
  2. go — what Go’s “simplicity by law” is precisely (gofmt, no generics for a decade, stdlib policy, module proxy + checksum DB for Q17); if err != nil vs Mo’s try; goroutines vs Mo processes. Touches d12, d23, q11, q17.
  3. rust — traits, enums, exhaustiveness, Result/Option, the borrow checker’s cost to humans and agents (find studies), compile-time causes; what Mo keeps (d22) and drops. Touches d22, d23, q5, p15.
  4. roc — platforms in detail (how a platform is written, the app/platform boundary, the effect interpreter), Perceus-style memory, the Zig rewrite numbers, Roc’s abilities vs Mo traits. Touches q11, d24, d25, q13.
  5. koka — effect types + handlers vs Mo’s capability parameters (d15); Perceus with reuse (the memory model Mo wants); compile-to-C. Touches d15, d16, d10, d24.
  6. austral — linear types, capability-based security, the “no magic” spec; how capabilities are obtained at the root and threaded; what linearity would buy Mo for resources/deadlines. Touches d15, p13, q16, q17.
  7. hylo — mutable value semantics precisely (projections, inout, set, sink), how it avoids aliasing, generics; new compiler status. Touches d13, d10, p3.
  8. unison — content-addressed definitions, codebase-as-database, exact dependencies by hash, hash-cached tests, why text files were dropped and what that cost adoption. Touches q10, d29, d23, q17.
  9. moonbit — the shipping “AI-native toolchain”: what MoonBit Pilot does, the ICSE 2024 semantics-aware sampling paper, multi-backend, what its language actually looks like, adoption. The closest competitor; be specific. Touches d1, q9, q10, q13.
  10. bosque — regularized programming (no loops, no mutable state, no reference equality), Mark Marron’s papers, why the project stalled (find the history), what Bosque 2.0 / recent commits look like. Touches d4, d10, q12, p11.
  11. spark-ada-and-dafny — contracts as language with prover + runtime semantics (SPARK); Dafny’s 82% LLM proof rate (vericoding benchmark 2509.22908), Dafny-as-IL (2501.06283), Marmaragan (2502.07728); what contract shapes LLMs discharge well. Touches d3, d22, q8, p8.
  12. agent-native-cluster — one page for Intent, Vera, Thermite, Vow, Aver, AILANG, Hale, Tacit, Zero, Codong, Axis, plus Neam and NTNT (not in the catalogue; verify they exist). For each: two lines + what overlaps Mo. Then: what did 42 fresh attempts converge on? Touches d1, d3, q9, q10, q12.
  13. verse — functional logic + transactional semantics (every function inside a transaction that can roll back), structured concurrency, Simon Peyton Jones’s papers; could transactions replace or complement crash-and-restart (d18, d21)? Wildcard; be skeptical. Touches d18, d21, d14.

Order

Do them in the order above (1 → 13). One page at a time, commit after each. If a page needs more than ~90 minutes, write what you have, mark confidence: low, and move on.

When done

  • Append to log.md: one entry per page (## [date] ingest | Comparison: <Language>).
  • Add each page to index.md under Research.
  • Write research/concepts/comparison-synthesis-draft.md: the ⚠️ contradictions across all 13, the top ten steals, and open questions for Robert — a draft for Fable to finish, not a decision.
  • Update HANDOFF.md “Where we stopped”.

Progress

Finished pages, in order (worker session):

  1. Elixir
  2. Go
  3. Rust
  4. Roc
  5. Koka
  6. Austral
  7. Hylo
  8. Unison
  9. MoonBit
  10. Bosque
  11. SPARK Ada + Dafny
  12. Agent-native cluster
  13. Verse

Synthesis draft for Fable: comparison-synthesis-draft