Your product, sold in Japan.Your product, sold in Japan.

THE BRIDGE — InstrumentsRecords as of Oct 2–3, 2026 (JST) · every figure comes from our own logs

MAKE — Working with factories5

1. SYS 01Ingest

   Supplier chat ingestion

   ~10 minAbout every 10 minutesHow often chats are pulled in Running · last pull Oct 2, 20:59 JST

   Every message in the factory chats we have set up for tracking — original text, Japanese translation and attachments — is copied into our own records without anyone opening a chat window. The reply meter and the question ledger both run on these records.

   Cadence

   About every 10 minutes (scheduled at :00, :10, :20 …)

   People decide

   Pulling in is automatic. Adding a new factory chat to the list is done by our team.

   On record

   Last 7 days: 106 received, 71 sent (as of Oct 2, 21:00 JST; attachments and links count as one each)

2. SYS 02Cross-check

   Draft and cross-check

   3 companies3 companiesAI companies involved. The one that writes and the one that checks are never the same (OpenAI, Anthropic, Google) Running · latest ruling Oct 2, 16:54 JST

   For each new factory message, one AI drafts an English reply and an AI from a different company rules on it: send, fix and send, or stop. A machine checks that ruling against the original chat and the requirement sheet, and only then issues a send ticket. We also feed in test messages with planted mistakes, on a schedule, to measure whether the checker catches them.

   Cadence

   A ruling for every request · planted-mistake tests every 10 min · the checker’s work tallied hourly

   People decide

   Prices, quantities, lead-time commitments, orders, payments and contracts are approved by a person. AI never makes a payment.

   On record

   Manual reply route — last 24 h: 14 checks, 0 failed to return a ruling (measured Oct 2, 20:07 JST) · last 30 rulings: 27 send, 3 fix or stop (Oct 1–2)

3. SYS 03Gates

   Pre-send gates and human sign-off

   6 checks6 checksChecks before and after sending Running · latest send Oct 2, 16:59 JST

   Before anything reaches a factory, it passes a row of machine gates: (1) no new message from them left unread, (2) numbers and dates that match the chat record, with no stray promises, (3) no question asked twice, (4) a ticket from the second AI, (5) text unchanged since it was approved — approvals expire after 24 hours. (6) After sending, we check that the text on the factory’s screen matches the draft.

   Cadence

   Every send

   People decide

   Contracts, prices, quantities, lead-time commitments, orders, payments and closing a project are approved by the founder on an A/B card on his phone. Routine messages — receipts, thanks, follow-ups from the requirement sheet, factual answers — go out once they pass the gates and carry a ticket.

   On record

   Latest sends: two messages on Oct 2, at 16:49 and 16:59 — both with a ticket, both matched against the factory’s screen after sending

4. SYS 04Reply meter

   12-hour reply meter

   00Factory chats waiting more than 12 hours for our reply (Oct 2, 21:00 JST) Running · reading Oct 2, 21:00 JST

   Every hour we count the factory chats that have waited more than 12 hours for our reply. The target is zero. Anything waiting 6 hours is raised to the team by an in-house scheduled notice, so we move before the 12-hour line. A message leaves the count as “no reply needed” only when the reason is written down.

   Cadence

   Every hour (since Sep 27, 2026; once a morning before that)

   People decide

   Counting and flagging are automatic. Replies are drafted by AI and go through the gates; anything with a commitment needs a person’s approval.

   On record

   0 at the Oct 2, 21:00 reading. Nine messages marked ‘no reply needed’ with a reason, and 16 closed chats, are not counted.

5. SYS 05Question ledger

   Supplier question ledger

   146146Questions marked answered (all time, as of Oct 2, 21:00 JST) Running · matched Oct 2, 20:37 JST

   Every question we put to a factory is kept as its own line. Twice an hour, new replies are matched against open questions and marked answered. A message that re-asks an answered question is stopped before it goes out. Each product’s requirements live in a numbered sheet that tracks, item by item, what each factory can do.

   Cadence

   Every 30 minutes (at :07 and :37)

   People decide

   AI marks a question answered; people spot-check it against the original text. Specs and requirements are decided by people.

   On record

   Of 14 pre-send checks, 4 stopped the message (all time; last used Sep 30) · requirement sheets: 3 projects, 70 items in total (as of Oct 2, 21:00 JST)

6. How one message reaches a factory

   1. 1Factory message
   2. 2Logged, original + translation
   3. 3Drafted by AI
   4. 4Ruled on by a second AI
   5. 5Pre-send gates
   6. 6Only with prices, quantities or ordersApproved by a person
   7. 7Sent
   8. 8Matched on the factory’s screen

SELL — Selling in Japan3

1. SYS 06Search articles

   AI-run search articles

   3×53×5Each morning: three AI models asked for five ideas each; returned ideas scored, the best queued (up to three) Running · morning run Oct 3, 05:38 JST

   Each morning, search data comes in; Claude, Gemini and GPT are each asked for five improvements, and an AI scores the ideas that come back and queues the best of them (up to three). During the day, scheduled runs edit existing articles, add internal links, improve titles and descriptions, and draft new articles. Before any change goes live, the old version is saved and the way to measure the change is set.

   Cadence

   Ideas at 05:36 daily · work runs at 09:06, 15:06 and 18:06 · watchdog three times a day · health check every 2 hours (12 items)

   People decide

   Direction, budget and what may be published are set by people. A new article goes live only if it clears five gates — claims sourced, quotes marked, no boilerplate, numbers and lists checked, a full read by a second AI — and never more than one a day. Miss one, and it goes back to a person.

   On record

   Oct 3, 05:38 run: 3 ideas returned → 2 queued as work · 80+ articles published on the blog in total (as of Oct 3)

2. SYS 07Video line

   Video production line

   3 slots a day3 slots a dayDaily publishing slots (two shorts, one explainer). On Oct 3, all three were uploaded by morning Running · latest upload Oct 2, 18:35 JST

   For our own brand’s videos, once an episode gets the go-ahead, scheduled runs carry it through script → review by four AIs → voice → stills → motion → assembly → machine checks. GPT, Claude, Gemini and Grok review as a beginner, a supervisor, an editor and a skeptic. The finished audio — exactly what viewers will hear — is transcribed by two speech models and matched against the script, word for word and number for number. Once a video is scheduled, a timer job (every 5 min) uploads it ahead of its slot; scheduling itself is done by the AI in charge after it has inspected the finished video.

   Cadence

   Production steps every 10 min · machine checks every 30 min · uploader every 5 min · metrics and the day’s plan at 05:35

   People decide

   People set the voice, format, spending limits and direction. Within those limits, an AI that looks at the actual output gives each episode its go-ahead and reviews each finished video after it passes the machine checks.

   On record

   100 videos uploaded (Aug 14 – Sep 29) · latest upload Oct 2, 18:35 JST

3. SYS 08Ad decisions

   Daily ad decisions

   05:3005:30 every morningRuns every morning Running · ran Oct 3, 05:30 JST

   Every morning, yesterday’s ad results are read and matched against real sales, and each product gets a proposed call: scale, hold, wait or stop. Eight written rules make the call — they weigh early buying signals and budget bands, and one bad day is not enough to move. Ratios and dates are computed by the machine, never by mental arithmetic.

   Cadence

   Every morning at 05:30 (a scheduled job on our office machine; if it was asleep, it catches up on waking)

   People decide

   Pausing, budget changes and going live all wait for a person’s approval. New ads are built switched off, checked by machine character by character — copy, links, budget — and previewed by eye before they run.

   On record

   297 automated tests around the ad-decision pipeline, all passing (Oct 3) · covers several ad accounts