Supervise exceptions, not output

The loom that knew it was broken

In 1924, Sakichi Toyoda built a loom that stopped itself when a thread broke. The innovation sounds small and was enormous: before it, one worker watched one loom, because catching a broken thread late meant yards of ruined cloth. After it, one worker oversaw dozens of looms — not by watching harder, but because the looms now announced their own failures and only failures needed a human.

Toyota grew a manufacturing philosophy from that loom and named it jidoka — usually translated "automation with a human touch." Two commitments define it: machines detect their own out-of-spec conditions and stop rather than continue producing defects; and humans, freed from inspection, concentrate entirely on the exceptions and on improving the process that produced them. Paired with the andon cord — any worker's right to stop the whole line on sight of a defect — it turned quality from an inspection activity into an architecture.

Your AI production line has no loom-stop

Now look at the modern knowledge worker's line. A machine produces drafts, summaries, answers, and analyses at a rate no human review can match — and, out of the box, it never stops itself. There is no broken-thread detector in a chat window. Every output flows onward equally confident, and the human either inspects everything (impossible at volume, as the variety law proved) or inspects nothing (unacceptable, as trust-but-verify taught).

The missing piece is not a better model or more diligence. It is the loom-stop: mechanical gates between the machine and everything downstream, checking every output against an explicit spec and halting the flow on violation. With gates in place, the arithmetic flips exactly as it did in 1924 — coverage becomes total because it is mechanical, while human attention concentrates where it is irreplaceable: the exceptions.

Out-of-spec must be written down

Jidoka's quiet precondition is the spec. A loom can only stop on a broken thread because "thread continuous" is a checkable condition. Most people supervising AI have never written their equivalent — they carry a felt sense of "good output" and apply it in real time, which is inspection, not architecture.

Converting the felt sense into conditions is the real work of this lesson, and the artifact fields from Phase 2 hand you the starter set: provenance present, falsifier stated, confidence declared, expiry set, links resolved, no unresolved contradiction. Add domain conditions — figures match the reference sheet, claims carry citations, banned patterns absent, format intact. Each condition must be checkable without your judgment; anything requiring judgment is either sharpenable into a condition or genuinely reserved for the exception review, where judgment belongs.

The verifier is not the generator

One implementation rule prevents a subtle collapse. If you ask the producing AI to also certify its own output — "and check that you followed the rules" — you have rebuilt self-inspection, the pattern quality engineering spent a century abandoning. Generators grade themselves generously.

Keep the roles separate: a checking pass that only verifies, whether a script, a schema validator, or a second model instructed to hunt violations and forbidden from fixing them. The separation is what makes the gate's word worth something — the same separation of concerns behind test suites that do not rewrite the code and auditors who do not book the entries.

Stops are sacred

The whole architecture rests on one behavioral covenant: when the line stops, someone looks. Toyota's insight was that honoring stops is what keeps stops meaningful — a plant where andon pulls get waved through is a plant where workers stop pulling, and defects resume their silent passage.

The personal version: a gate violation gets prompt attention, every time, even when — especially when — it turns out to be a false alarm. False alarms are information about the gate (log them; the next lesson runs on that log). Ignored alarms are corrosion of the covenant. If you find yourself routinely overriding a gate to keep the flow moving, the gate is mistuned and must be retuned or removed — never silently bypassed, because a bypassed gate provides the feeling of coverage with none of the fact.

What your attention becomes

Follow the discipline for a month and notice where your time went. Not to reading output — to reading stops, and to a second activity jidoka considers the human's true job: improving the spec. Every exception that reaches you is evidence about the process — a condition missing, a prompt inviting a failure class, a reference sheet gone stale. Fixing the process retires whole categories of future exceptions.

That is the trade the loom offered a century ago, now offered to knowledge work: stop being the inspector of everything; become the designer of what stops the line.

Sources

• Ohno, T. (1988). Toyota Production System — jidoka, andon, and stopping the line.
• Liker, J. (2004). The Toyota Way — build quality in; separate detection from production.
• Shingo, S. (1986). Zero Quality Control — source inspection and poka-yoke over judgment inspection.