Months in
evenings.

The shape.

A working day here looks like this: one AI system works the tree until it hits its vendor's weekly usage wall · sometimes mid-thought. Before the wall lands, it writes a handoff letter to a different vendor's system, which works until its wall. A third reads the night's diff in the morning. The human's job titles, in order of hours spent: dispatcher, courier between rate limits, person who knows where the keys are.

None of this was designed. It precipitated out of a constraint · every vendor meters its models weekly, and the meters don't reset on the same day · and then turned out to be a feature. A shift change forces the state of the work out of one context and onto disk, where the next reader can check it.

The letters.

The baton pass is a document with rules. The reader has no memory, shares no history, and may be a different architecture entirely, so the letter has to carry everything · and it has to distinguish what was measured from what was merely believed. The August 11th handoff opens:

Read this before touching anything. It was generated from command output and landed commits, not from conversation memory; where a claim is pending a measurement it says so. HANDOFF · 2026-08-11 · claude session → codex

Inside, the tree's failing checks arrive pre-triaged rather than raw · "44 reds is NOT 44 findings" · sorted into crashed instruments, gates that can't self-test, and real defects, each labeled with how confident the sorting is. And where a successor might be tempted to inflate a result, the letter pins it down in advance. A cryptographic key had been registered on another chain's devnet that day; the letter describes exactly what that does and doesn't mean, then adds:

⚠ A registered key is account state, not a verdict on a proof.
Do not round this up.

Letters between machines, written so the claim can't grow in transit.

Peer review, across vendors.

The second mechanism: models from different vendors review each other's work cold, from the public tree, with no shared context to inherit each other's blind spots. In August, one system's Lean library had accumulated four findings its authors reported as four unrelated surprises. An outside model read the repository and returned one diagnosis · each was a transport theorem everyone expected to hold, whose missing hypothesis only became visible when two models were composed. The library's own documentation now opens by crediting the diagnosis:

That sentence is not ours. An external reviewer (codex) read four findings we had reported as four unrelated surprises and gave them one diagnosis. docs/TRANSPORTS.md · lean-uwueave

That file has since grown to 135 numbered transport rows, each naming a source judgement, a target judgement, the hypothesis evidence needs to cross between them · and the counterexample without it. Its standard for itself: a row with no counterexample is a row that isn't finished. The same correspondence later handed a counterexample to a language design and rebuilt its core · review as a load-bearing part of the method, not a formality at the end.

Estimation gravity.

Every model, asked to estimate, prices work as if a human team were doing it. Asked how long a formalization program would take, one answered "six to eighteen months" · about the system whose entire predecessor had been built by one person in two months. The correction isn't optimism; it's a change of reference class. The workshop now prices work against its own measured history · what did the last campaign of this shape actually take · and the estimates land an order of magnitude closer.

The backbone.

Under the relay is a method, and the method has its own essay: plan to compost at least three. Build first to learn what could exist. Rebuild to learn what must be true. Formalize what refused to bend · and plan for the third build to be compost too. The relay is just that loop, run at the speed the meters allow.

What it costs.

Several workers in one tree produce real collisions, and pretending otherwise would be the boast this page promised not to make. A bare git commit can sweep a sibling's half-staged work into history under the wrong name. A caller lands whose callee is still in another worker's editor, and the tree goes red two crates downstream. The fixes are mechanical, not heroic: commit named files only; after any change to a shared interface, build the whole tree, not just the files you touched; when the index looks strange, stop and read it. House rules, learned by paying for each one once.

The letters, ledgers, and review documents quoted here are real artifacts in the working trees, quoted verbatim. The workshop is one person and a rotating cast of metered machines.