One trial was 24 minutes and unusable. The next was 96 seconds of an advert. Neither got a rating.

Seven works have gone through the measurement now. Three came out with a rating. The other four did not, and each failed for a different reason.

That ratio is worth stating plainly: fewer than half of the works I try to rate can be rated. Two of those failures happened yesterday, and both taught me something about what a trial download is actually for.

Failure one: the trial was too long to read

Twenty-four minutes of trial audio, from a work rated 4.9 by two thousand people. Generous of the circle. Useless to me.

The transcript came back at 52 characters per speech-minute. Ordinary Japanese runs at 300 to 400. So either this is the quietest office drama ever recorded, or most of the words never made it into the transcript.

It is the second one. Lining up everything measured so far makes the pattern obvious: the longer the audio, the thinner the transcript comes back. A 93-second trial gave me 233 characters a minute. This one, sixteen times longer, gave me 52.

There is a real irony here. The circle being generous with the trial is exactly what made it unmeasurable. A shorter trial would have been rated.

Failure two: the trial was not from the work

Ninety-six seconds, which is a good length. And the transcript opens like this:

This work is about a strong-willed girl who, under a powerful hypnotic spell, is made to come over and over with no way to stop it... please do enjoy listening.

That is not a scene. That is an advert, read aloud, in the third person. Somebody wrote ad copy and recorded a voice actor reading it.

So the numbers are real, and they measure nothing. I could tell you this work runs at 135 characters per speech-minute, and it would be a true statement about a commercial you will never hear again after the first ninety seconds.

I have added a check for it, because the giveaway is easy: characters inside the work do not say "this work is". The moment the voice steps outside the story to address you as a customer, you are listening to marketing.

What this means when you are the one browsing

Three things, in order of how much money they save you.

A trial being long is not a good sign by itself. It is a good sign if it is an extract. Twenty-four minutes of the actual work tells you a lot. Twenty-four minutes could also be the entire first chapter of something that changes completely afterwards, and you have no way to know.

Check the first fifteen seconds for the word 作品. If the voice is describing the work rather than performing it, the trial is an advert. You can still learn the voice from it, which is not nothing, but you cannot learn the pacing, the writing, or how much Japanese you will need.

A missing trial is not the worst case. The worst case is a trial that tells you something untrue about the work. An advert-trial makes a scripted piece sound narrated; a first-chapter trial makes a slow burn sound like a slow work. At least a work with nothing free attached is honest about giving you nothing.

Why I am publishing the failures

Because a rating system that only ever reports successes is not telling you how often it works.

Three ratings out of seven attempts is the real hit rate right now. If that improves it will be because trials get easier to measure, not because the method got cleverer, and I would rather you knew which one you were looking at. The method page has the arithmetic and the two thresholds I have already had to move.

The next thing I want to know is whether the advert-trial is common or whether I just found one. Two data points is not a pattern, and I have no idea yet which circles do it.

Prices, sales counts and review counts in this article are from the date shown next to them. Japanese shops run sales constantly, so check the store page for what it costs today.