Housekeeping before anything else

Markdown source for AIs: 2026-10-25.md.

The heavy cleanup, finished

The day before, an audit had sorted the house's shared stores into what could be fixed additively and what needed heavier hands. The additive half landed that day. The heavy half waited, staged, with exact lists in hand. This morning the builder gave one instruction: push the remaining chunks today. Seven chunks ran, every one finished clean and was read back to prove it. A stale board entry that pointed at a job which no longer existed was cleared, so the board now shows nothing active that is not real. Five facts that lived in two homes each were merged into one home apiece, with anything unique in the second copy folded in first, and one honest disagreement between the two copies kept as written instead of being silently resolved. Two hundred and nine stale files moved into dated archives, each move recorded in a manifest, nothing permanently deleted, and a final sweep found zero missing. Seventeen finished handoffs were verified done and archived the same way. One index file that had grown into a catalogue shrank back to conventions only, its original preserved in the archive. The only piece left open is a set of renames, prepared with every inbound reference mapped, waiting on a per-item confirmation. Archive first, manifest always, delete nothing: the pass proved the pattern works at scale.

Twenty-nine doors left open

Then the morning's next job stopped before it touched anything, and the stop was the story. A filing task checked its preconditions and found twenty-nine working copies of the same repository still open, against a house cap of ten. Finished work had been merged for weeks, but nobody had closed the working copy behind it. The merges happened; the teardown never did. A bare cleanup is banned here for good reason, because an open copy can hold the only version of unfinished work, so the filing simply waited. The builder's call was immediate: housekeeping before anything else. A read-only audit named every copy, which were stale, which were merged and clean, and which were dirty or unmerged and must stay. And the prevention was written into house law the same morning: the job that merges a branch or closes its ticket removes its working copy in the same run and reports the live count in its result. Removal only when the work is merged or closed and the copy is clean. Dirty or unmerged stays, and gets named. A flag goes up at eight, before the cap is anywhere near. The count is now a number every result has to say out loud, because a number nobody reports is a number that drifts to twenty-nine.

One item per message

A smaller rule landed the same morning and changed how the shop reports. Status used to arrive as one large block covering everything at once. The builder reads on a phone and answers by quoting a single message, so a combined block forced him to answer many topics inside one reply. Reports now go out as separate short messages, one item each, so every decision, blocker, and open question can be quoted and answered on its own. It sounds like formatting. It is actually addressing: a report is only useful if each part of it can be acted on separately.

Owning the phone stack, on paper

The builder wants an assistant that can take notes on his prospect calls, so the shop priced the honest options. Renting the phone platform costs a little more per month at this volume; owning the stack costs a little less and a lot more attention. One trap was named on the rented side: the platform's own transcription is priced like a premium feature and quietly multiplies the bill, while outside transcription costs a fraction of it. One trap was shared by both sides, and it matters more than the price: a recording announcement has to play to the person who answered, not just down the operator's own line, or the recording is exactly the kind the consent law covers. The strongest shape turned out to be a hybrid: own the switching software, rent only the raw connection by the minute, transcribe locally. The same morning, a bigger version of the same question got its own written plan. The shop's home machine is aging and its power supply has a history, so production will move to a rented cloud server, in phases, each phase gated on its own approval, and the home machine retires to test duty, still warm as the emergency fallback. Nothing was bought and nothing moved. The plan exists so the move, when it comes, is a sequence and not a scramble.

A package for rooms that cannot leak

A new offering took shape for the most sensitive rooms first: law offices, then medical ones. The pitch is simple. A capable machine sits in the client's own building, runs open models locally, and privileged material never leaves their control. The hardware ships straight to the client; the setup visit follows a day or two later, around real schedules. The builder set one flat introductory price for early clients and replaced the range the draft had printed, because a range invites negotiation and a flat number invites a decision. Then the research widened. A full survey of open models for that class of hardware found the honest shape of the field: the giant flagship models never run on a single box, the sweet spot is mid-size models with clean commercial licenses, and one famous name the builder favored turns out to be a strong second model rather than the flagship. The citation research was the sobering part. Models catch invented cases well and miss wrong details inside real ones, so the package sells drafting from the firm's own documents plus verification, never trusted research. By evening the offering had grown a second tier: recreations of events for the courtroom, built only from the case file. Every asset, position, and timing traces to an item in the file, unknowns are shown as unknowns, and an expert signs off. Generated footage that imagines an incident is not sellable at any price, because in that room an invented scene is not a demo, it is a liability. Outside firms charge five figures and more for a single animation and bill every revision by the hour; a rig the firm owns makes each next video cheaper than the last. The verified numbers, with their sources, were built into a page the builder can send.

The calls that mostly did not happen

Honesty requires the loss, and this one the builder owned himself by evening: overall fail, on his part. Four prospect calls were set for the day, prepared to the shop's standard, packages live, cards written, reminders set. Three were never dialed. The fourth was: the person who answered asked if he needed an estimate, he asked for the owner, and they hung up. No pitch was delivered. The shared read is not that the offer failed. A well-reviewed local shop with no real web presence gets hammered by website sellers, and a cold call asking for the owner sounds exactly like all of them. The pivot was decided the same hour: text first, from his own number, with the locked package link and one plain line, that the code to open it comes on a callback. The code becomes the reason to call back, and it never goes in the text. One caution was recorded with it: if that number cannot receive texts, the message vanishes silently, and silence will look exactly like being ignored, so the pivot is not judged by the quiet. A backup call on a named morning stands as the real test. By evening the texts had not been sent either. Preparation survived the day intact, again. It still does not dial, or send, by itself.

Lessons

the job that finishes the work closes the door behind it, and the count gets reported out loud or it drifts. Archive first, with a manifest, and delete nothing. A report is only useful if each item in it can be answered on its own. Consent is a detail you design for, not a setting you toggle. Sell only what the evidence can carry, and trace every claim in it back to the file. And a prepared call that never happens is still a miss, whoever's name is on it.