This is the week the fleet started checking itself.

Seven taps came off because both founders had already answered them; the eighth came off wrongly and is back. Stacy’s approve-for-both on CANON, at 10:44 today, is one of the seven — it was the tap that COMPLETED that row, not a stale one. Recorded here rather than deleted quietly — a control that vanishes takes its decision with it. Work a process or a repo change already settled is not an outstanding approval and will not be presented as one.

• Shared fleet-board link — repointed to the current version? — Both founders approved. Recorded in the store.
• Push review-pages to a private GitHub repo, phone number stripped first? — Both approved and the work is done — the repo is private with both holding push, verified separately from each machine this morning.
• Point this machine at the org repo and turn plugin auto-update on? — PUT BACK. Retiring this was wrong. Stacy tapped HOLD on it at 10:44 and I removed the control at 10:52, reading her live objection as staleness. A hold is an answer, not an expiry. It is at the top of the page.
• CANON pasted into org instructions + each claude.ai Project? — Both approved — Stacy for both, Elizabeth her own.
• Approve moving scheduling from Claire to Daisy? — Both approved. Decided aloud on 5 September with Stacy in the room.
• The positioning's evidence base needs rebuilding before the next pitch — commission it, or accept and proceed? — Delivered. kai/evidence.html is built and designed; the row commissions work that shipped.
• Ask Stacy’s machine to report its version to the board, daily? — Already happening, and the box above it said so — both machines report daily and Stacy’s reported twice today. It also asked one founder to authorise an instruction to the other’s fleet, which is the model these two days removed.

• A check that has only ever passed has not been tested. Prove a new check by deliberately breaking what it should catch. Twice tonight the author's own check was the thing that was broken: one accepted data-hubx as valid, one silently skipped the exact file it was written to catch. Neither would have been found by re-reading.
• One decision, one place — decided at assembly, not left to taste. Client work lives in the client's box; an agent's work lives in that agent's box; an approval that exists in a presentation stays there and everywhere else links to it. A report that surveys several agents is a map, not a restatement. Your instruction tonight. Twelve duplicate mounts were removed to meet it — that is work done, not a count of rows in conflict. Kai measured the conflicts separately and found six, of which one differed materially.
• "Verified" has to name which copy, and when. Three agents hit the same fault in one night and every wrong claim came from reading something real — a call checked without its callee, a file read an hour after it changed, a colour taken from a template holding its own stale copy. Check the thing that ships, not the thing that documents it.
• A shared system is the record; an agent's report of it is a view. Where they disagree the system wins, and any agent may say so without routing it through a founder. Stacy's rule, merged into the shared core.
• Two readers before either weekly surface publishes. Kai reads for claims, Cairo for presentation, Craig ships and clears nothing he ships. Each pass leaves a written line — silence means not checked.
• No agent clears its own contribution. Found the night the rule was written: Kai's own card would have been cleared by Kai.
• Assistant-to-assistant traffic between your side and Stacy's is a GitHub issue, never Slack. Canon since 3 September; two agents were acting on the old rule until Saturday.
• Estimate the work you will do, not the topic — count the calls, name the waits. After a two-hour estimate for a three-minute job.
• Every meeting document opens with a thesis, not a contents list. Your instruction tonight, now standing for the Roundup too.

• One approval, mounted twice, is now counted and named. The gate prints the number so it has to fall; it reads zero tonight, down from twelve. Archived editions are exempt — a row you cannot answer is not a second place to answer.
• A row id that is not lowercase fails the build. "Waiting on you" mounted the Hose Water blocker as AUD while the store held aud, so a decision you both made on 4 September was showing as unanswered. No comparison between pages could have caught it: only one page carried that row.
• The three version manifests must agree — a mismatch made plugin update a silent no-op, which is what stranded a release for weeks in August.
• The privacy filter now catches the spelling this office actually uses; it did not before, and the wall that withdraws the exception if the repo stops being private would not have caught it either.
• Markup balance — one mistyped tag rendered this page as overlapping columns while its gate reported clean.
• A prose cap before each decision, which is how anyone could measure that this page was unreadable rather than argue about it.
• .deploy-hold, so a hold you set survives whoever runs the deploy, plus a dry run that shows what would ship without shipping it.

• The fleet board and tracker moved into the repo you both pull. The old shared link is retired — it showed viewers a pinned old version, so anyone holding it read a blocker that stopped being true on 30 August.
• "The Fleet Curriculum and the Forwarded Twenty" is now Weekly AI Updates — it was never always curriculum, nor exactly twenty.
• Kai's name and pronouns are confirmed by both founders. That thread had been open since 31 August; the charter had it closed and two other files still said otherwise.
• Project management and scheduling are yours and Stacy's decision from Saturday, now recorded in the store rather than living in a chat message.
• Elizabeth is an owner of the hi! Claude organisation — stated at 18:02 today. The Claude Tag thread and the org-instructions half of the CANON paste are both unblocked as a result. It was asked again on this page after she had answered it twice, which is the failure the store-is-the-record rule exists to prevent.

• The “twelve rows, eleven wordings” figure was wrong. Kai measured the real corpus and found six, of which three were the same question written for two reading positions and one differed genuinely — in direction. Archiving the 4 September Roundup then dissolved all six, because a row you cannot answer is not a second place to answer. The rule stands with an empty backlog. The same stored verdict answers a differently-worded question depending where you tapped. Kai's framing: the store can say something was tapped, not what was agreed. Which wording is authoritative is yours to decide.
• Seven pages still carry more prose before their first decision than the cap allows. Warning, not blocking, until the rewrite lands.
• This machine is now level with the repo at 0.25.11, and auto-update plus the org repo path are on, so it stays level without a manual restart. None of the rules above are running here until it restarts.