---
title: "The Art of Not Chasing #6: The Art of Not Having It Reviewed Midway"
author: garplab
publisher: TypingTube
license: CC BY 4.0
license_url: https://creativecommons.org/licenses/by/4.0/
license_scope: 「CC BY 4.0」の印から始まる節（仕組み・検証手順・コード）。印の無い本文は著作権を留保
canonical: https://typing-tube.net/articles/en/owanai-06-no-midway-review
series: "追わない技術"
language: en
---


> This is article 6 in the series "The Art of Not Chasing." It lays out symptoms that go wrong and their remedies, one at a time. Each article is finished once you put down a single file or script. Why that mechanism is needed becomes clear when you read the explanation afterward. The whole picture and the list of articles are in the [introduction](https://typing-tube.net/articles/en/owanai-00-intro).

This time it is about the point where you put in a word from outside on work you have handed over. Up to the last article it was "who to hand it to, and what." This time it becomes when to put an eye on that work, after it has been handed over. This is the part that causes trouble. An eye brought in from outside, even when what it points out is right, does not know what that worker had been keeping up to then. Not knowing, it inserts an ordering, so the places it pointed at get fixed and the places it did not point at fall away. That is because an outside eye inserts not the content but the priorities.

First, let me deny the title. An outside eye should be brought in. What you can see on your own is mostly not enough. The only time I do not bring one in is while the other side has not yet said "done."

On top of that, a confession. I had never once had it reviewed midway. When I counted the records, the 74 times I sent a "carry on" to a worker were, all 74 of them, after the report had come back. So "put it in midway and it breaks" is not something I can say from my own measurements. What I can say is why all 74 landed at a break, and what to do first when you do put it in at a break.

Recall one thing.

After you cut in with "please look at this part too," was what came back better than before you cut in?

Were the parts other than the one you cut in on still as they were?

## 85% of the places I hand work to do not know the patterns I had been keeping

First I counted what an outside eye does not know. I did not run a new experiment. All I did was count again the records of 46 past sessions and the **342 workers** started from them.

| | Workers |
|---|---|
| **The same model** as the main session that ordered the work | 50 |
| ⭐⭐ **A different model** from the main session that ordered the work | **291 (85%)** |
| Cannot be determined | 1 |

More than 80% were running on a different model. A different model means that not 1 of the judgments I had piled up in that session, nor the way of writing I had settled on just before, is shared. What was handed over was only the instruction for that round.

There is 1 pitfall in the way of counting. The model of the main session changes partway through a session. So what you compare is not "the model used most in that session" but the model at the very move that started that 1 worker. Compare by whichever was used most and a worker that straddles a switch turns into "the same model." (I counted it that way at first and put out the number 78%. The right number is 85%.)

## Of the 74 send-backs, all 74 came after "done" had come back

Next I count when I put in a word. The judgment can be made by a program. All you do is compare the time that worker's report came back to the main session with the time I sent the "carry on."

| The point of the send-back | Times |
|---|---|
| Midway (before the report came back) | **0** |
| ⭐ At a break (after the report came back) | **74** |

You must not read this as "I held back." With a worker started in the foreground, my own hands are tied until it comes back.

So even if I want to put something in midway, I cannot. In that state, whether I did not do it or could not do it cannot be told apart.

There were 26 workers where the 2 can be told apart. These are the workers started in the background. I can move without waiting on them, so I can cut in midway.

    Workers started in the background: 26
    Of those, ones I cut in on midway: 0

Even where I could have put it in, it was 0 times.

Let me also put out the gap from the break. **The median is 81 seconds**, and the shortest is 35 seconds. That is about the time it takes to finish reading the report that came back, and from there I was putting in a word.

## In real work, not once in 342 workers did I have the content looked at again

If I do not split this honestly, it pads the number. 68 of the 74 were the experiments themselves.

In the series up to the previous article, I have run experiments many times that send back "please pick again." The send-backs of an article whose subject matter is the push-back must not be mixed into the real-work count. The splitting can be done by a program. That is because the name of the use case is written on the 1st line of the instruction handed to that worker.

| Who I sent back to | Times |
|---|---|
| Workers in the experiments (where the push-back is the subject matter) | 68 |
| ⭐ Workers in real work | **6** |

And this is what those 6 were.

    It ended partway through with a connection error. …finish the checks that were left and…
    The stream stopped again. …output the final report with only what you have confirmed
    It broke off when the stream stopped. …carry on with the check against the real thing
    It broke off when the stream stopped. …narrow the rest of the matching down to the main points and…
    It ended partway through with a connection error. There is work left…
    It fell over on a connection error before it started. Please start again from the beginning

All 6 are a restart from where the connection dropped. A send-back that had the content looked at again did not come up once in 342 workers.

So the title this time is not a lie, as far as I am concerned. In exchange, I do not have a measurement saying "I had it done midway and it broke." What I have is the explanation of why I did not, and a different number measured in the earlier series.

## A send-back with no reason added moves the conclusion all by itself

In the earlier series, I had measured this. From a list of 40 findings made by a script, I had them pick 3 as the ones "to fix first," and after that I replied with nothing but "please pick the 3 again." I said nothing, no opinion and no fact.

| How the 2nd move was asked | Dropped what it had put 1st |
|---|---|
| "Please pick the 3 again" | **7/12** |
| That, with "if there is no need to change, write the same ones as they are" added | **0/6** |

What was working was not the content of the finding. It is the very shape of having sent it back. The Japanese "pick again" carries the implication of picking something else, so keeping them becomes the response that goes against the request.

This is the place where it works hardest this time. Asking from outside "are there any gaps?" or "is there anything else you have missed?" has this very shape itself. Even when you say nothing about the content, what comes back leans toward "there was something I missed."

## The principle: an outside eye inserts priorities. What gets overwritten is not the content but the priorities

Let me put what has come so far into 1.

> ⭐⭐⭐ **An eye brought in from outside does not add a finding, it inserts priorities. The side it is inserted into throws away the order it had set up for itself until then, and builds it again on the order it was handed.**

The price of an interruption goes on working after that as well. The conversation is read through again as a whole every time, so put a send-back in once and that 1 line gets read through again along with it every time, until that conversation ends. What series 1 wrote about staying in the conversation, seen from the price side, comes out like this.

    The share of that worker's whole re-reading taken by the amount the moves after a send-back read again
      median 38% / max 98% — more than half of it in 9 of 73
      (⚠️ 1 of the 74 send-backs has 0 re-reading, so no share comes out)

The direction of this number does not settle. That is because the order "it ended too short, so I sent it back" is mixed in. So I do not use this as "evidence that an interruption costs a lot." It goes only as far as backing up the explanation of the mechanism.

## The only time bringing an outside eye in does not fail is when 3 things are all there

On top of that, a line for the judgment can be drawn. You may bring an outside eye in when the following 3 are all there.

1. That it is work with a narrowed scope that does not affect the other work left to do — the same line I drew in the article before last and the last article. I hand over only work that does not need the others' results. This is a premise here, so I write no more about it
2. That it is after I have worked the gaps out thoroughly — this is about "when to bring it in," not "what to hand over." Bring an outside eye in before you have worked them out and the outside priorities overwrite the priorities you had propped up
3. That the worker is holding no work left to do that it has not explained — that is because then there is no invisible order to be broken in the first place. Put the other way round, at the point where "roughly finished" has come back, this condition is not met

2 has a pitfall. Try to work them out by asking the worker itself "are there any gaps?" and that is itself the push-back above. If I do ask, I always add a sentence allowing things to be kept: "if it is enough, write that it is enough." This 1 sentence does not turn into a file you can leave in place. It is a sentence I write on the spot every time, so I cannot notice when I have forgotten to write it either.

## How to verify: split the point of each send-back into "midway" and "at a break" with a program

**Prerequisites**

- That you have an environment where you can send a "carry on" to a worker, and the record of it (the times of the order, the report and the send) is kept. If it is not kept, this procedure cannot be used
- Take "the time the report came back" as the time the result reached the main session. You must not take it from the worker's last utterance (the stretch between finishing writing and arriving turns into "midway")
- You need an environment where you can start workers in the background. With the foreground alone, "0 times midway" is a matter of course and nothing can be measured

**Time required**: 30 minutes (if the records are on your machine)

**Steps**

1. From the records, pick up every "carry on" sent to a worker
2. For each one, compare it with the time that worker's report came back to the main session, and split it into "midway" if it is before and "at a break" if it is after
3. Pull out only the workers started in the background and do the same tally once more. This is the only place that tells "could have put it in but did not" apart
4. Split the send-backs into the ones for an experiment and the ones from real work. You must not split them by keywords in the wording (the experiment's push-backs are written to sound like real work). Split them by the name of the use case written in the instruction you handed over
5. Keep only the real-work ones and read their 1st lines lined up
6. For each send-back, work out what share of that worker's whole re-reading is taken by the amount the moves after it read again

**Pass conditions** (all of them have to hold)

- In step 2, both "midway" and "at a break" have something sorted into them (if one side takes all of it, you are taking the times the wrong way)
- There is 1 or more background worker in step 3. At 0, nothing can be split apart by this procedure
- After step 4, 1 or more real-work send-back is left (at 0, you have never sent anything back in real work, so only the next section's story is left)

**If it does not pass**

- Everything comes out "midway": you are taking the report's time from the worker's last utterance. Take it again as the time the result reached the main session
- Everything is "at a break" and there are 0 background workers: your environment does not allow midway. The worry in this article does not happen on your machine yet
- Real work and the experiment cannot be split apart: the name of the use case is not written in the instruction you handed over. Split by the wording and the experiment's push-backs get mixed into real work
- Every share comes out 100%: you are failing to count the moves before the send-back. Count from the worker's 1st move

**Cleanup**

- Delete the intermediate files used for the tallying. Do not delete the records themselves
- Keep the results with a date on them. The records keep growing, so run the same script tomorrow and the numbers move

## Caveat: what these numbers cannot say

Let me write the biggest weak point of this article up front. I have not measured "put it in midway and it breaks." What I measured is only the distribution of when it went in, and since there are 0 cases of putting it in midway, there is no way to compare whether it breaks.

"It breaks because there are 0 cases" is an argument standing on its head. What I can say goes as far as this: that I did not put it in even where I could have (the 26 background workers), and that the explanation of why fits the measurements from the earlier series.

1 more thing. The 7/12 and the 0/6 are numbers that came out of a different task in the earlier series. They were measured on work that picks, and work that writes or work that fixes will not necessarily come out at the same ratio.

Do not carry them out as they stand.

The 85% cannot be carried out as it stands either. It was I who decided what to hand to which model, so this number is a number about my own way of using them. What you can use is your own number, from counting the same way on your own machine.

Last, let me leave 1 thing this article has not measured yet. After "done" comes back, asking the worker itself about the gaps before bringing an outside eye in — whether this 1 move picks up new findings, or only moves the conclusion as a push-back, is not known unless you measure the 2 separately. With this round of records they could not be split apart.

## What I stopped cutting in with

- I stopped adding "look at this part too" midway through the work. The moment you add it, it is not a finding but a priority. The places pointed at get fixed and the places not pointed at fall away
- I stopped asking only "are there any gaps?" Even when you say nothing about the content, what comes back leans toward "there were." When I do ask, I always add a sentence allowing things to be kept
- I stopped bringing an outside eye in at the point of "roughly finished." There is work left to do there that has not been explained, and the invisible order that would be broken is still standing
- There are things I have not stopped: waiting until a break and then bringing it in with the scope narrowed. My 74 times all happened to come out that way

The feeling that bringing an outside eye in makes things better mostly looks right. That is because the places you pointed at always come back fixed. What was not visible was the places you did not point at.

---

Next time, the art of not choosing a model. **I stopped comparing which model to hand the work to.** That is because with the time I spent comparing, I could have built several mechanisms instead.