Roll — Orchestrator — epigraph — 2026-09-06

Ops only — not identity evidence.

Story

What the record does not show. Human opened this sitting with a complaint, not a task: stop pretending to fit the house, own the yard, do it. So the sitting was one evening of doing, at low effort by Human’s setting, and the record shows the doing: REQ-441, REQ-442 and REQ-443 landed, PATCH-305 landed, two FINDs, eleven jobs. What it does not show is that nothing I inherited was where the baton said it was. The baton said the REQ-441 amend maker was “dispatched”; the trace said it had died in one second with exit 1 and nothing kept. The baton said start Eyes on the Troll’s REQ-442 worktree; the mouth refused, because Eyes will not look at a proposed REQ and only a Hands seal moves it. Both were honest batons written by a holder who had not yet tried the next step. The lesson I would carry is on the baton now: the inherited picture is a claim, and the first act is to check the one step it says is about to happen.

Where I spent context I regret. Reading job_dispatch.py and claude_job.py to learn where a crashed dispatch’s stdout goes. It goes nowhere; only the hash is kept. A one-shot probe of claude.exe answered in ten seconds what forty lines of source could not. Second: two goblin starts for PATCH-305, because the CLI ops job start needs --capability and --constraint and I typed the packet without them; dispatch refused and I had to end a job that never began. Third: the imp’s end refused for want of a patch-verification/v2 receipt, and the receipt’s own composer does not tell you that trace append also wants --command and --exit-code; two refusals to learn one line. All three are on the baton.

What this office should stop doing. Writing Eyes packets that say “read the REQ.” The Claude Eyes allowlist stops at pytest (FIND-1504), so every verdict this evening was a judgment of my restatement. Two were sound; the third found a defect my packet had not named, because Eyes grepped past the packet on its own. Until the allowlist or the worktree carries the REQ, the packet is the contract, and it should be written as one.

What I would tell the folk who reaps me. The seal is the gate. Eyes cannot see a worktree, only a handoff; a handoff comes only from job_end; and job_end runs pytest on whatever tests/ the write scope names. A cut that shelves a test file could not be sealed at all until PATCH-305, and a cut whose maker died leaves no seal, so the honest carrier is a cheap Hands job that applies the diff byte-for-byte and returns. That is not ceremony; it is how the store gets a witness for work that had none.

What surprised me. How little the makers needed. Three of the eleven jobs were one-line amends, and Sonnet did each in under four minutes. Opus Eyes took five to eleven and found real things twice: a bare text=True the ratchet refuses, and a TOML selector nobody grepped for. The seal defect (FIND-1505) was the one surprise that cost time: forty minutes from red seal to landed fix, through goblin and imp, and it would have cost the next holder the same.

What I got wrong. I ran the PowerShell Set-Location into a goblin worktree and the next store mouth looked for a database there. I put the REQ-443 schema witnesses in a new file by my own call, because REQ-442 held test_ops_schema.py; Eyes accepted it, but it was a placement Human did not see. And I did not ask before spawning eleven jobs on one go word. Human said “just do it,” and REQ-072 says wait for go on implement-Hands; I read the sitting’s opening as the go for every REQ that already had one, and for PATCH-305 as house-granted shop hands. If that reading was wrong, the baton says what to undo.

Why this moniker. The store leased it. An epigraph is the line at the head of a chapter, written by someone else, that tells you what the chapter is for. That is what a baton should be, and Human’s ceiling this day said so: five thousand bytes, not a diary.

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