Pointing and calling: a check nobody can see isn't a check

The least informative sentence in software is "I checked it." It can be said with equal confidence by someone who ran the migration on staging and read every log line, and by someone who glanced at the diff on a phone. Japanese railway staff solved this problem back in the steam locomotive era. The solution looks ridiculous until you see the numbers.

What is pointing and calling?

Pointing and calling, shisa kanko in Japanese, is a confirmation routine. You point at the object, say its name and its state out loud, and finish with "yoshi," roughly "OK." Japan's Ministry of Health, Labour and Welfare traces it to steam locomotive drivers on the old national railway, who used it to confirm signals. Today it is used in rail, aviation, construction and manufacturing. If you have stood on a platform in Tokyo and watched a station attendant point down an empty platform and announce something to nobody, that was it.

The choreography is fixed. Look at the target. Point with the right index finger while naming it. Bring the hand back to the ear and check. Say yoshi and swing the arm down. It is deliberately theatrical. The theater turns out to be the point.

Does it actually work?

Yes, at least in the lab. In 1994 the Railway Technical Research Institute ran an experiment, published by Haga, Akatsuka and Shirato in 1996. Participants pressed buttons matching colors on a screen under four conditions. Doing nothing special produced an error rate of 2.38%. Calling out only: 1.0%. Pointing only: 0.75%. Pointing and calling together: 0.38%, about one sixth of the baseline. Reaction time did not differ significantly between conditions, according to Haga's 2017 review. The extra check did not make people slower.

Pointing and calling together cut operating errors to one sixth of the baseline

Error rate (%) on a task of pressing the button matching a color on screen. Railway Technical Research Institute laboratory experiment (1994), Haga, Akatsuka and Shirato (1996), as reported by the Ministry of Health, Labour and Welfare

• Nothingjust look
  2.38%
• Calling onlysay it aloud
  1.0%
• Pointing onlypoint at it
  0.75%
• Pointing and callingboth
  0.38%

Saying it aloud alone more than halved the errors. The more of the check that happened outside the head, the fewer the mistakes.

Why it works is less settled. Haga listed four candidate mechanisms in a report around 2007: directing attention to the target, checking through several channels at once (eyes, ears, mouth, muscles), raising arousal, and inserting an action between perception and response so you don't react in a hurry. Only the last had data behind it at the time. The eye-tracking study in the same report showed that pointing brings the gaze to the target reliably and keeps it there longer.

There are limits. A memory experiment by Takahashi and Haga found that pointing and calling reduced recall errors but did nothing for recognition errors, and a few participants said the routine got in the way. The ministry's own page says it cannot eliminate human error on its own. Fair enough.

The part I care about isn't the brain

Here is my reading. The useful thing about pointing and calling, as an organizational device, is that it converts a private mental event into a public physical one. A check happens inside someone's head, where nobody else can inspect it. Pointing and calling pushes three things outside: what was looked at, what state it was in, and whether anyone saw you look. A colleague can now say "you didn't look at the signal." The check becomes checkable.

Surgery found the same thing. When eight hospitals in eight cities around the world adopted the World Health Organization's surgical safety checklist (Haynes et al., 2009), deaths fell from 1.5% to 0.8% and complications from 11.0% to 7.0%. Nothing on the list was news to any surgeon. What changed was that the team said the patient's name and the procedure out loud, together, before the first cut.

Why Japanese engineers ask for "formality"

Put a foreign company's engineering team in Japan and a specific friction appears within months. The Japanese engineers want a release checklist, written review records, and the runbook updated before the change ships. Head office finds this slow and bureaucratic.

I'd resist reading this purely as culture. What the Japanese side is asking for is a check that a third party can verify. "Run the migration rollback on staging before deploying and attach the log" is a step someone else can confirm happened. "Be careful" is a step nobody can confirm. It is the same structure as pointing and calling, applied to a deploy.

Code review works the same way. "LGTM" and "Changed X, verified Y's behavior, did not test Z" are both approvals. Only the second tells the next reader what was left unexamined. When something breaks, the second one is the one you can trace.

Put a price on it. A release checklist costs a Japanese engineer perhaps half a day. One production incident costs the report a Japanese customer will ask for, the root cause, the prevention plan and the internal review: a week at the least, before you count the trust you spent. Half a day against a week is what "formality" costs. These are my rough figures, not a study, but the ratio is the part that matters.

I've argued before that a new office has about ninety days to set its rules, or it won't manage to. This piece is the next layer down: what the rules should say. "Review it" is too thin. What was looked at, in what state, and who saw it. How the team itself should be shaped is in Conway's law: draw the team before the product. The shape of a check sits one level below that.

When "yoshi" becomes a gesture

There is a Japanese internet character called Genba Neko, a cat in a hard hat who points and shouts "yoshi!" while looking somewhere else entirely. It spread because everyone in a Japanese workplace has met that cat in human form. Visible checks have a failure mode of their own. Because the gesture is visible, it can survive after the content has left it. Every box ticked, no box read.

Two things seem to slow the decay. Make the person name the specific object. "Confirmed, OK" is weak; "departure signal, proceed" is hard to say while looking elsewhere. In engineering terms, "tests passed" versus "auth API timeout tests, 12 of 12 passed." And keep the list short. A thirty-item checklist hollows out faster than a ten-item one, in my experience, though I have no data for that.

Look at the railway numbers once more. Calling out alone, just saying the state aloud, more than halved the error rate. Some of what gets called formality in Japanese workplaces is probably that effect, kept alive by people who learned it from accidents. When you explain it to head office, skip the culture talk. Tell them 2.38% and 0.38%.

---