the train ran, and it said red
the release train has rails. soft3 carries the phase-1 manifest, release/phase1.toml, the sibling revisions every candidate closes over, and a qualifier that reads it, builds only from committed and pushed inputs, records every sibling's revision, runs the named gates and writes receipts. cyber consumes that qualification for its own draft releases. the rules the machine follows are on the cyberia/dev page and in an AGENTS.md beside every CLAUDE.md, so Codex, Kimi Code and Grok Build read the same thing Claude reads.
this morning it cut the first three candidates: candidate-20260924.1 for cyber, soft3 and cyb, draft pre-releases with a component table, captured revisions and a verdict. all three are red, and the red is exact: every captured sibling is the origin default (bbg 0.2.1, lens 0.1.3, nox 0.2.0, zheng 0.3.3) while the versions the owner actually runs sit in four unmerged pull requests. a working tree is never an input, so the train cannot see them. a red candidate ships red with its evidence; nothing was hidden and nothing was patched by hand. the first green candidate is four merges away.
sources
- the train: soft3 #27, #28, #29 · cyber #111 · soft3/specs/releases.md · release/phase1.toml
- the candidates: cyber candidate-20260924.1 · soft3 · cyb · the node rehearsal receipt cyber/audit/release-2026-09-23
- the rules: cyberia/dev.md § release train · AGENTS.md
- the four merges: bbg #9 · lens #14 · nox #17 · zheng #33