opening
the reading
the reading
runningChoosing a company to open on
Explaining Direction Unstated
There is a state in Teniq that reads *direction unstated*.
It appears when a filing has clearly moved something — a reason is no longer sitting where it was — but the filing does not say which way. The engine will tell you the reason moved. It will not tell you whether that is good or bad for the argument.
People find this unsatisfying, and they are right to. It is the least useful thing the product can say. It is also the reason the rest of it can be trusted.
An engine that always produces an answer will produce one when the evidence does not support it. Not by malfunctioning. By doing exactly what it was built to do, on an input where the honest output is nothing.
This is the ordinary failure mode of language models applied to documents. Asked what a filing means, a model will say what a filing like that usually means. The answer will be fluent, plausible, correctly shaped, and unattached to the document in front of it. It is very hard to catch, because a confident wrong answer and a confident right answer look identical on the page. You only find out which one you got by going and reading the filing yourself, which is the work you were trying to avoid.
So the design question is not how to make the engine right more often. It is what the engine does when it does not know.
Teniq is built to stop rather than guess. When the filing supports a conclusion, the engine states it and shows you the sentence it came from. When the filing does not, the engine says so and stops. There is no path in the system that produces a conclusion with nothing behind it.
Concretely, that means three things you will see in the product:
A reason can move without a direction. The engine reports the change and withholds the interpretation. This is the *direction unstated* case.
Absence renders as absence. If nothing in a filing bears on a reason, the engine says nothing happened. It does not manufacture a small observation to fill the space. A quarter where nothing moved is a real result and gets reported as one.
Nothing appears in the filing register without provenance. Every quoted span carries the company, the form, the accession number, and a live link. If a sentence cannot be traced to a document, it does not render as one.
Fail-closed systems are quieter than their competitors and look less capable in a demo. A tool that always has something to say about every filing will produce more output than this one. Some of that extra output is real. Some of it is the tool talking.
You cannot tell which from the outside, and neither can we, which is why we would rather not be in that business. The output you get from Teniq is smaller than it could be. Every piece of it is attached to a document you can open.
Every company building on language models says it is careful about accuracy. That claim costs nothing to make and cannot be checked, so it carries no information.
What can be checked is a rate. We measure how often the engine produces a claim that the cited document does not support, and we publish the number rather than describing our process.
A measurement with a stated limit is more useful than a polished claim with none. It also means we can be wrong in public, which is the only version of accountability that means anything.
Teniq does not tell you what to buy or sell. It does not look at prices. It does not have a view on whether a company is cheap.
It reads filings, writes down the reasons a company is worth owning at the moment those reasons are derived, and tells you what each subsequent filing does to them. The derivation date is pinned so the reasons cannot be quietly rewritten later to match what happened. That is a separate design decision and it gets its own piece.
The short version of this one: when the filing supports a conclusion, you get the conclusion and the sentence. When it does not, you get silence. Both are answers. Only one of them is the one everybody else is willing to give you.