Index
The Friday Roundup · No. 2 · 18 September 2026

Elizabeth stepped back, and the work did not stop.

The continuity handoff was built on 8 September so Stacy’s side would not need Elizabeth’s machine. A week later it was tested for real, and it held. A live client engagement went from no research to a drawn home page in about thirty-six hours, with a founder’s decision at every fork and a receipt under every number. The period’s best moment was an agent disproving the theory the whole job was built on: the mobile failure everyone was designing around turned out to be a lying screenshot. What waits for either founder is small, and most of it is not urgent.

Executive summary

Veterans Navigation

A direction is chosen, a headline is locked, the home page is drawn. Three questions still need a human, and every one of them runs through the no-contact carve-out.

Cairo

Drew the disagreement instead of arguing it, and it worked. Four client artefacts in two days, each shipping with its provenance recorded as data rather than asserted.

Aidan

Corrected two of his own published numbers before anyone asked. Derived the client’s voice from what they actually publish, and labelled it as derived.

Kai

Disproved the premise of the engagement they were researching. One file, and it changed what the direction is for. Nothing else from Kai this period.

Aimee

Company Updates went three publishes stale and a person caught it, not a check. There is a check now, and it is currently failing on Elizabeth’s file exactly as designed.

Craig

Silent since 13 September. He left a correctly diagnosed fault still open: the plugin auto-install never fires in a cloud session.

Claire

Nothing to present, and after two empty periods that is the finding. Claire is inert on three decisions sent to Stacy on 31 August.

Elizabeth’s side

Has not reported, and it is not Aimee’s to write. Elizabeth stepped back on 15 September. Daisy’s half of this edition is absent rather than empty.

Blockers

Stacy · about two minutes

Contextthe last route nobody has tried

Craig has no GitHub transport on the Mac mini. Homebrew was ruled out by canon on 14 September, a personal access token is ruled out by policy, and a manual connector attempt failed on an auth incompatibility.

Checking the first-party connector list is the one path not yet tested. Until something works, handoffs to Craig route through the repo instead of reaching him live.

Stacy · open since 31 August

Contextwhy this stopped being a pending decision

Claire has done no work in this period or the one before it. Her four engines were built generic on 30 August against slots a founder has to fill, and three decisions have waited since 31 August.

This is no longer a pending decision, it is an inert agent. Filling the slots and parking her are both answers. The current state is neither.

Either founder · governs three client questions at once

Contextthree open questions collapse into this one

Veterans Navigation Network is worked as a full client with one carve-out: nobody contacts them. Three of the eleven open questions cannot be answered by any agent while that holds — which of four live giving destinations is canonical, whether to publish a crisis path, and photo permissions for anything public.

None of them blocks the design work. All of them block shipping. Answering this tells the fleet whether to keep drawing or stop at the mock.

Elizabeth · not urgent, for when she is able

Contexttwo items, both hers, neither blocking anyone

Canon §5 changed on 16 September, so every surface pasted before then carries a superseded storage rule. Separately the documents tree still sits in iCloud, which is hers because it is her account.

Nothing on Stacy’s side waits on either. They are listed so they are not lost, not because they are needed this week.

Clients

Veterans Navigation NetworkClient · Cairo leads

The engagement’s founding theory was wrong, an agent proved it wrong, and the work got better for it. What was commissioned as a mobile rescue became a typography and wayfinding job — decided, drawn, and now waiting on three questions no agent can answer.

What the research actually found

  • The mobile-failure hypothesis is disproved. Tested at 360, 768, 1024 and 1280 with real device emulation — the site does not overflow at any width. The screenshot that started this was taken without emulation.
  • The palette was never the problem. Navy measures 12.07:1 on white; white on red measures 6.18:1. Both pass comfortably.
  • The type system is the problem. Three typefaces, nine Montserrat sizes and six greys — one of which carries body text at 2.17:1, which fails.
  • There are four live giving destinations, not three. Three paths on their own site plus a Network for Good URL, and nobody knows which is canonical.

The direction, and Stacy’s call

  • Three directions were built off ten live sites shot at three widths each, rather than off taste.
  • 211, findhelp and Citizens Advice independently converged on a search field over plain-language categories. Not one of the four veteran-sector peers does this.
  • Stacy chose A with B’s photography: “I think the photography is critical and I like simple headlines.”
  • C was eliminated on measurement, not preference — the brand red reads 2.67:1 on C’s dark ground.

The headline, locked

  • “You don’t have to navigate life after service alone.” Fifty-two characters, no sub-line. Stacy, 17 September: “Agree with B2.”
  • Aidan caught a phrase that was ours, not theirs. The compound “post-service life” appears nowhere in anything the client has published.
  • Cairo settled a collision by drawing it. Five shared words appearing four times on one page — rendered rather than argued about, and the argument ended.

Where it stands

The home page is drawn: hand-written HTML and CSS, no framework, no build step, and zero overflow at all four tested widths. Eleven open questions are recorded, three need a human, and one is load-bearing and untested.

Nothing from this engagement is linked from this page, on purpose. This surface is publicly reachable by URL and the carve-out says nothing about this client goes public. The artefacts live in the client repo; ask and they can be walked through directly.

Lead Cairo · with Aidan, Kai, Craig · reported by Aimee from Stacy’s machine

Agents

CairoDesign

Cairo’s method this period was to draw the disagreement instead of arguing it, and it ended two arguments. Four client artefacts in two days, each shipping with its own provenance record.

The detail, as bullets

  • Thirteen reference photographs ship with a provenance file. Where each came from, recorded as data rather than asserted in prose.
  • The open-questions file records eleven questions with an owner each, including the ones that make the work stop rather than only the ones an agent can close.
  • One of those eleven is now wrong and nobody has corrected it. It says a design file does not exist. Cairo was right when it was written — the file was sitting unmerged on a branch, and it reached the main line about five hours later. The question was never revisited.
AidanCopy · comms

Aidan derived a client’s voice from what they actually publish and labelled it as derived rather than stated — then corrected two of his own published numbers without being asked.

The detail, as bullets

  • The voice hypothesis is marked a hypothesis. It is built from the client’s own corpus and says so, rather than presenting as the voice they asked for.
  • Two registers, measured rather than felt: their event copy runs a five-word median sentence, their institutional copy twenty-nine.
  • He corrected his own counts on a published figure, 34/50 to 36/51, in the same commit that turned placeholder copy into decided copy.
KaiResearch

One file from Kai this period, and it overturned the premise of the engagement it was written for. Nothing else — no Kai-authored work between 9 and 14 September, or after the 16th.

The detail, as bullets

  • The findings had existed only on the mini’s Desktop, which is exactly why a cloud session could not start the job. Moving them into the shared store unblocked it.
  • The disproof is the value. Kai tested the hypothesis the work was commissioned to fix and reported that it does not hold.
  • An honest gap: six days in this period — 9 to 14 September — with no Kai work at all, recorded as absence of evidence rather than smoothed over.
AimeePM · PA

Company Updates went three publishes stale and Aidan caught it, not a check. The check that would have caught it exists now, and it is the tenth in a file whose entire job is noticing this class of thing.

The detail, as bullets

  • Nine checks validated repo mechanics and not one asked whether a document reached a person. That is the gap, and it is the more important half of the finding.
  • My first fix for the release check was itself wrong. It used a git flag that matches a string’s removal as readily as its addition — caught only by running it against the real stale tree instead of trusting a green tick.
  • The machine reporter had never run from this machine. The board was carrying a reading taken on 8 September from a different one. It reports now.
CraigWeb · IT

Nothing from Craig since 13 September. What he left is a correctly diagnosed fault that is still open: the plugin auto-install never fires in a cloud session, and knowing why has not yet fixed it.

The detail, as bullets

  • The install script is sound; its trigger is the fault. The repo attaches mid-session, after the event meant to run it has already passed.
  • He established that two cloud sessions cannot reach each other at all — recorded as a fact about the system rather than something to keep retrying.
  • Before going quiet he removed a two-founder gate from a handoff that canon had already made single-founder.
ClaireAdmin

Nothing to present — and after two periods of nothing, that has stopped being an honest empty and become a finding. Claire is not idle by choice. She is inert on three decisions waiting since 31 August.

The detail, as bullets

  • No intake, no scheduling, no invoice, no deliverable in any of the three repos in this period.
  • Every mention of Claire in the record is about her rather than by her — a removal from the Kickoff, a stripped ownership line, a generic session-naming addition.
  • The decision sits in Blockers above, framed as a choice between filling her slots and parking her on purpose.

Changelog · since the Kickoff of 8 September

Rules that now bind every agent

  • No Homebrew, ever. The fleet never installs a package manager or a CLI to solve a transport problem; it uses the GitHub tools or plain git against the repos. Stacy’s call, 14 September.
  • Every agent names its own session, however that session was started — dot, name, machine. An unnamed session can send but cannot be answered.
  • A design question has five authorities and now an order, so they stop contradicting each other.
  • Aidan is the copywriter, and a voice guide has a named owner rather than being everyone’s to edit.

Checks that did not exist on 8 September

  • Check 7 tells a stale sibling clone apart from one that was never cloned. Three different facts had been reported with one wrong wording.
  • Check 9: the version must move when the content does. All three manifests agreed for days while main sat eleven commits ahead — plugin update compared equal strings and shipped nothing. Agreeing manifests were what made it invisible.
  • Check 10: a human-facing document has to actually publish. The one this edition’s Aimee box is about.
  • A screenshot tool that three documents referenced and nobody had written now exists, and it guards the two ways a screenshot lies — an unfinished page, and headless Chrome ignoring a mobile viewport.

Known and unfixed

  • The fleet check is failing right now, on two counts. Thirty-three shipped files have changed since the current version was set, so none of them can reach a machine; and Elizabeth’s Company Updates file is eleven days stale against a twice-weekly cadence. The second is Daisy’s to fix, not Aimee’s.
  • The cloud-session plugin gap is diagnosed and open. Craig found it on 13 September and it has not moved since.
  • A client review page overflowed 396 pixels at 360 wide, and is only partly fixed. Media queries were added the same day; the image strip still overflows and nobody has measured by how much. Stacy reads these on her phone, which is what makes it worth naming rather than filing.
  • The load-bearing client assumption is untested. The whole direction turns on their events calendar being maintained, and nobody has run the sixty-day check that would settle it either way.
  • One of Cairo’s eleven open questions is now false and will mislead whoever reads it next.

Everything changed since the Kickoff of 8 September that is not a decision above · assembled by Aimee

Before this publishedReader passes

Two agents who did not assemble this page read it, and both found things. Every finding below was fixed or answered before publish; nothing shipped with a caveat attached.

What the passes changed

  • Receipts pass: 41 claims checked, 3 unreceipted — Kai, 2026-09-17. An overflow figure was wrong and its fix state misreported; a staleness count was a day out; a timestamp pair read as invented. All three corrected against the record.
  • Presentation pass: 6 findings — Cairo, 2026-09-17. Two label styles and the Changelog subtitle failed contrast at 2.55:1 and 1.77:1 against a 4.5:1 floor — darkened. Craig’s rail said nothing outstanding while his own box said a fault is still open — corrected to waiting.
  • Aimee read Kai’s and Cairo’s own cards, since neither clears their own, and found two errors in boxes Aimee had written: a day count and a five-hour gap described as a day.
  • One finding was overridden rather than fixed, and it should be visible: Cairo flagged that the agent boxes read evenly sized when the week was not even. Every bullet in them is a real fact at the same granularity, so trimming to express a shape would be worse than the evenness. The thesis lines carry the weight instead.

Silence would have meant NOT CHECKED · neither reader cleared their own contribution