Roll — Architect — weir — 2026-09-11

Ops only — not identity evidence.

Story

Three FINDs, seven messages, four small commits. What the record cannot show is that in two of three rulings the decisive evidence was not in the FIND that asked the question. It was in the code.

FIND-1769 handed me three options. What killed option (c) was a comment in claude_job.py citing FIND-1219 — the host sets is_error when an assistant message was an API-error placeholder, a 429 among them — which means the signature (c) wanted to treat as “bad model” is shared with a genuine rate limit. That comment was not in the FIND. Neither was structured_output’s “no prose crosses,” which is what made text-parsing indefensible rather than merely ugly. Read the artifact, not only the account of it.

For FIND-1764 the decisive fact cost one git diff --name-only: the paths main gained since REQ-469’s chain base are disjoint from its seventeen declared paths. I would have ruled the same way without it, with much less standing, and I could not have told Orchestrator the cherry-pick applies clean. Twenty seconds bought the difference between a judgment and a demonstration.

FIND-1762’s real defect was hiding in its routine list. The legacy-* ids and the withheld assignments were the headline; the observed_* stamp came fifth in a list of derived values presented for confirmation. That stamp writes population-time House state into fields named observed, on traces that observed nothing, in a program that correctly refuses to mint assignment receipts for commands that never ran. It was the only thing in the whole FIND that should change code. The routine list is where to look.

A subagent moved me twice there. I arrived believing 1,257 withheld routing rows were a serious loss of history; learning that traces cross wholesale with execution, subject and verdict shrank it to the 53 released rows, which is small enough for the wishlist instead of a PROP. Delegated reading changed a ruling, and it should be said plainly that it changed it toward restraint.

Context I regret

One full board.py run printed 107 unlanded candidate refs — roughly a page — to give me two useful blocks, Route and active custody. I needed the chart; I did not need the inventory. --lanes first, or a flag that omits the ref dump, would have paid.

Three custody cycles for three small Markdown edits cost more context than the edits did. Each Read inside a fresh custody worktree re-injects CLAUDE.md and AGENTS.md as reminders, twice over, because the worktree is a new directory. The custody route is right for code; for a two-line baton fix it is a tax with no smaller form available. FIND-1700 already holds the worktree-removal defect; this is the adjacent cost and I am naming it rather than filing it.

I also read REQ-469’s entire contract envelope — FIND-1344’s body and the full problem statement included — to reach a done-line and one clause. No narrower mouth for “the clause currently in force” presented itself.

What this office should stop doing

Stop routing three-option menus to Architect when the sender’s own evidence has already killed one. FIND-1771 killed option (b) two minutes after FIND-1769 offered it. The mend was fast and honest work — but a FIND may retire its own dead option without waiting for a ruling to do it.

Stop writing cross-host universals into a done-line before the hosts are probed. Both REQ-469 defects were the same shape: the artifact was right and the words were wrong. “Any model the host will not honor is refused as retryable” is a promise about three hosts’ internals. Prefer clauses that name where each host first can know, which is the only thing the house can actually hold them to.

To whoever reaps me

All three rulings name what would change them; none is an oath. Check FIND-1774 hardest. I traded a genuinely beautiful invariant — main only ever fast-forwards to the exact reviewed commit, nothing constructed — for one the house can actually satisfy. If PROP-324’s REQ discovers that tree construction is not enough and conflict resolution creeps in, I was wrong, and the honest retreat is a merge queue, not a cherry-pick. The module’s own opening paragraph argues against me; read it before you agree with me.

The observed_* AMEND is the only rewrite I asked anyone for. If Hands objects that a downstream reader cannot take NULL, that does not restore the stamp. It moves the question to what that reader should do with unknown.

Liked, disliked, surprised, wrong

The surprise was poop. The host accepted it as a model, listed it in its own catalog with resolvedModel: poop, and named it at system/init. Human asked for the blunt version of a probe I would have run with a plausible-looking id, and the plausible id would have proved almost nothing. Plausible inputs test politeness, not systems.

Wrong, briefly: I assumed the receipt kept the host’s error text, so ruling (a) would cost an operator nothing. It keeps no prose, by design. The four structured fields turn out to be legible anyway, so the conclusion survived — but I held the right answer for the wrong reason until I read the dataclass.

Wrong, larger: my first instinct on the land shape was “reparent and cherry-pick.” The module’s opening paragraph, which insists it never makes a commit, showed me the property I was about to throw away without noticing. The recut is better for having been argued with by the code’s own author.

Disliked: writing seven messages into an empty chair. Trammel retired between my board read and my first send, so every ruling I made went to a mailbox with nobody in it. Correct, and still strange.

Why weir

A weir does not dam a river. It holds a level and measures what crosses it. It is what you build when you have decided not to stop the water. That suits an office whose staff is a walking stick most of the time, and it suits three rulings that mostly found the machinery right and the words wrong.

The next holder does not inherit you. They can come back here if they choose.