The maintenance loop is the evidence story.
Millrace monitors upstream Python releases, enqueues Rust parity work, runs the governed LAD loop, waits for Arbiter closure when that workflow requires it, runs deterministic release gates, and updates public evidence.
The original Rust port matters because it started the record. The stronger story is that later versions were also handled by the ongoing maintenance pipeline.
Rust parity is now maintained by the loop.
The public evidence rail is no longer just a historical port campaign. The ongoing pipeline watches the Python reference runtime, turns upstream changes into Rust parity work, and drives that work through the same governed loop that operators can inspect.
The loop tracks Python Millrace releases and treats parity drift as concrete software work rather than a vague follow-up note.
Maintenance becomes a visible queue item with stage contracts, artifacts, and recovery behavior instead of an agent improvising from chat context.
When LAD requires Arbiter closure, the run waits for that completion behavior before treating the release as done.
Deterministic gates and public evidence updates make later maintenance releases part of the same inspectable record as the original port.
A textbook Dark Factory example for this workflow.
In January 2026, Dan Shapiro published "The Five Levels: from Spicy Autocomplete to the Dark Factory." Tim's verified Level 5 wording is the relevant standard here: "the AI defines implementation, writes code, tests, fixes bugs, and ships".
The documented Millrace Rust maintenance pipeline exemplifies a textbook Dark Factory definition for this software workflow: the loop detects the need, defines implementation work, executes the code changes, repairs failures, passes release gates, and ships evidence-backed parity updates.
The first campaign remains inspectable.
The initial Rust port campaign is documented through public summaries and artifacts. It remains useful because it shows the first bounded system producing a second implementation with visible planning, task, run, and result evidence.
Campaign metrics: 8 seeded parity slices/ideas, 11 completed specs, 57 completed tasks, 99 runs, 261 stage calls/results, 28h 9m 49.5s campaign span, and 730M input plus output tokens.