Seed 18: One Simulator, Nine Languages, the Same Bytes
Table of Contents
A draft for the proposal shelf. Not submitted anywhere. Status: New, 2026-11-17.
1. Pitch
Write a deterministic simulator in your language in an afternoon. If it is right, it prints the same bytes as nine other languages, and it finds the bug on seed 18.
2. Abstract
Deterministic simulation testing, the technique behind FoundationDB's simulator and TigerBeetle's VOPR, turns a distributed-systems failure into a seed: run the whole system in one process, route time, randomness and the network through a kernel, and any failure replays exactly. DST-0 is the smallest kernel that still makes the point: a 32-bit generator, a virtual clock, an event queue, a network that drops, duplicates and delays, and a replicated register with an ordinary last-writer-wins bug that only reordering exposes. It is specified tightly enough that independent implementations in Forth, Tcl, SML, Python, Ruby, TypeScript, Clojure, Racket and Guile print byte-identical traces for every seed tried, and all nine find the bug on the same six seeds of fifty. The talk shows seed 18 on a space-time diagram, sweeps a thousand seeds in twenty milliseconds, and plants a missing 32-bit mask live to turn the nine-language agreement matrix red. The workshop has attendees write the tenth implementation against a checker that prints the first line where their trace leaves the golden one, which in practice names the bug.
3. Audience, format, venue
- Audience
- Anyone who writes or tests distributed systems, and anyone who likes writing the same small program in many languages. Needs one language and no libraries.
- Format
- Three cuts of the same material:
- lightning, 5-10 minutes: all demo (seed 18, the sweep, the matrix, a planted fault);
- talk, 30-40 minutes: what DST is and who uses it; the five things a kernel owns; a live bug hunt; the traps that differ by language (integer width, tie-breaks, the order of generator calls);
- workshop, 2-3 hours: five checkpoints from the generator to
150/150 match, then a fault swap in which each pair plants a fault in the other's code and names it from the first differing line.
- Where
- Systems Distributed (its audience already runs DST); Papers We Love and the Recurse Center for the talk; local Clojure, Racket, Scheme and Python groups for the workshop, since each brings its own language to the matrix; EuroForth 2027 for a Forth-first cut tied to the keel work (see below).
4. EuroForth 2027
Not announced as of 2026-11-17: euro.theforth.net still shows the 42nd EuroForth (4 to 6 September 2026, with a Senadd tour on the 5th, so Wales), and a web search found no 2027 dates, venue or call. Recent editions all fell in September (Rome, 15 to 17 September 2023; 12 September 2025; 4 to 6 September 2026), so the likely window is early to mid September 2027, with papers due over the summer. See the 2026 notes.
A Forth cut would lead with the gforth implementation (case-insensitive names shadowing the
compiler's STATE, 64-bit cells needing a 32-bit mask) and end on keel: the 13-primitive
Forth kernel whose stepped execution and snapshots would let DST-0 schedule events while
keel runs each node's code.
5. Review
Not yet reviewed. The question a committee will ask is what is new: DST is established, and "the same program in N languages" is a genre. The answer is the conformance claim, which neither has on its own. Byte-identical traces across independent implementations make a seed a bug report that does not depend on the language, and the agreement matrix is evidence anyone can rerun. Each implementation was also tested with a planted fault (missing masks, a reversed tie-break, swapped generator calls), and every one was caught. The weak spot is that the nine were written within one day by agents working from the spec; the workshop's fault swap is the human version of that check.
6. Backing
aygp-dr/dst-kernel: the spec, nine implementations, the cross-language checker, the seed sweep and the trace viewer. Private; to be made public before submission.aygp-dr/dst0-workshop: the participant kit (spec, golden traces and fingerprints, the checker, facilitator notes), with no implementations. Private; to be made public before the workshop.aygp-dr/keel-rs: the 13-primitive Forth kernel and its own simulation harness (wasm keel against a JavaScript model, step by step, with snapshots and a shrinker).