Sens は、外側に置きます。あなたの側に、言葉がどう使われているかの記録を持つ。AI に渡すのは、そのつど、必要な分だけ。これを観測層と呼んでいます。
On the home page we said: what's missing isn't knowledge. It's context. So — where do you put that context?
One way is to put it inside the AI. Fine-tune it, embed it, write it into the system prompt. But models change. You switch. Whatever you put inside goes with them.
Sens puts it outside. You keep a record, on your side, of how words are actually used. The AI gets what it needs, when it needs it. We call that the observation layer.
Data =「この文書がある」 Context =「この文書は、この人たちにとって、この状況では、この意味を持つ」Data — "this document exists." Context — "this document means this, to these people, in this situation."
The first already travels. Files can be sent. Search works. The second is usually written down nowhere. "Here, this is how we treat it" lives in someone's head.
The usual way to align what words mean is to decide first: write a glossary, then have everyone use it that way. Sometimes that works. It has a cost.
決めるやり方Deciding
誰かが、正しい意味を決める必要がある
決めた後で現場の使い方が変わると、ずれていく
揃わない言葉は「間違い」として扱われる
Someone has to settle what the right meaning is
When usage on the ground drifts, the glossary drifts out of true
Words that don't match get treated as mistakes
記録するやり方Recording
決めません。すでに使われている言葉を記録します
使い方が変われば、変わったことがそのまま残ります
同じ言葉が部署で違う意味を持つこと自体が、記録すべき事実です
We don't decide. We record the words already in use
When usage changes, the change itself stays in the record
That the same word means different things in different teams is itself a fact worth recording
揃えるのが目的ではありません。どう使われているかを、そのまま渡せるようにするのが目的です。
The goal isn't to make everyone agree. It's to make how it is actually used something you can hand over.
③ 観測の図What is recorded
何を記録しているのかWhat is actually recorded
縦の線は、ひとつの言葉が、いくつもの文章に現れること。 結び目は、その言葉を含む文章が、ひとつの観測として記録された点です。A vertical line is one word, appearing across many texts. A knot is a point where a text containing that word was recorded as one observation.
Sens がするのは、この結び目を打つことだけです。線をまっすぐに直したり、結び目を揃えたりはしません。
The knots sit at different heights on different lines. Texts containing the same word differ with who wrote them, when, and in what situation — and so does what gets recorded.
Tying those knots is all Sens does. It does not straighten the lines, and it does not line the knots up.
④ できること/できないことCan and cannot
できることと、できないことWhat it can and cannot do
ここを曖昧にする説明は、検討の材料になりません。先に書きます。
An explanation that blurs this line is no use to anyone deciding. So here it is first.
できることWhat it can do
言葉がどんな話題の構造で使われているかを記録する⚠︎ 体験版では、観測はあらかじめ用意した例です
観測した一つひとつに、どう測ったか(道具と設定)が必ず残る
AI に渡す前に、何を渡すのかを確かめられる⚠︎ 体験版が開くまで、公開している面にこの経路はありません
渡した後に、何を、いつ渡したかを辿れる⚠︎ 体験版では、この記録はブラウザを閉じると消えます
Record the topical structure a word is used in⚠︎ in the trial, the observation is an example prepared in advance
Keep, for every observation, how it was measured — the instrument and its settings
Let you check what you are about to pass before it goes⚠︎ until the trial opens, this path is not on the public site
Let you trace, afterwards, what was passed, and when⚠︎ in the trial, this record goes when you close the browser
It is not robust to rephrasing. Close paraphrases hold together; once the wording changes substantially, it can stop recognising them as the same thing.
What has been recorded can sometimes be guessed from outside. Not the content itself — but assume that "this topic is in there" may be inferable.
We don't answer with a single number. Different axes reach different depths, and averaging them loses the meaning. Any figure we publish comes with the date it was measured and how.
この 3 つを隠さないことが、観測層の設計そのものです。 私たちを信じる必要はありません。確かめられる状態を渡すことが、私たちの仕事です。Not hiding those three is the design of the observation layer. You don't have to trust us. Handing you something you can check is the job.