Live network
Ask the estate a question
A question in plain words, a computed answer out, and underneath it the query that ran and the rows it read. 32 questions are understood in 102 phrasings. Anything outside that is refused by name rather than guessed at, because one confident wrong answer would end the usefulness of this surface.
Ask
Not signed in. Questions can be composed, but the surface will not run one until it knows who is asking.
I cannot answer that — The scope of that question is ambiguous
The question could be read against more than one population and the answers differ materially. Choosing one silently would be a wrong answer with a confident face.
Name the population - one week or the run rate, one function or the estate - and ask again.
How the surface refuses
Refusal behavior is the feature. These are the only ways this surface is allowed to say no, and each one names itself.
The reason reached for most often is that needs a person, not a query, 4 times of 13 refusals.
What has actually been asked
Every question put to this surface, whether it could be answered or not, with the outcome it produced.
Actions
What is waiting on a person
An answer surface is judged on what it refuses, not on what it answers. Each line is a property of the resolver measured against its own record of 33 questions.
questions the surface holds a computation for
Each one carries the query it runs, the population it covers and the records it reads. 70 alternative phrasings map onto them. Anything outside that set is refused.
recorded questions resolved to a computation
They read 11,945 rows between them. The row count is shown on every answer because an answer that reads nothing is not an answer.
of the recorded questions were refused with a named reason
That is 39.4 percent. A refusal names one of 5 fixed reasons and offers the nearest questions it can answer. It never returns a number it cannot compute.
people have used the surface
On a modeled history of 33 questions. This is a seeded record of use, not observed adoption: nobody outside the build has been put in front of it.
catalogued figures cannot be reached by a question
The catalog holds 34 figures and the resolver holds 32 questions. The difference is reachable only by opening its entry in the catalog, and the surface says so rather than inventing a phrasing for it.
Operations
What this desk is allowed to start
A surface that only reports is not operable. This is the work this page can set in motion, and the bound it runs into.
Trigger and bound
What a question does. It is tokenized, matched against stored questions and their phrasings, and the winning match runs one stored query. No language model is called and nothing is generated. When it will not match. At least half of the words carrying meaning in a question have to be words the matched question knows. One overlapping word is not a match, which is how a surface ends up answering something nobody asked. What a refusal contains. One of 5 named reasons, what that reason means, what to do instead, and the nearest questions that would work. What is written down. Every question asked is recorded with its outcome, the reason if it was refused, and the number of rows read.
Live observability
What the record shows right now
The 5 reasons below partition all 13 refusals in the record. That needs a person, not a query leads on 4 of 13.
Current distribution
13 refusals
Is policy and strategy coming to fruition
Whether the written intent is holding here
Partly
Nothing on the record settles this
The written intent is plain language in, a traceable number out, and an honest refusal when the data does not hold the answer. What is now true: 32 questions resolve to a stored query, every answer shows the query and the row count, and 39.4 percent of the recorded history was refused by name. What is not true: this resolves phrasings, it does not understand language. A question worded outside the 102 phrasings it holds will be refused even when the estate could answer it, and that failure looks identical on screen to a question the estate genuinely cannot answer.
What this page is, and what it is not
Every answer here is a stored query run against the estate, not a generated sentence. The surface can only answer the questions it holds a computation for. When it cannot, it names a reason and stops - a refusal is the correct output, not a failure.
The resolver was checked three ways against what is stored. Every one of the 32 questions resolves to itself. Every one of the 70 alternative phrasings resolves to the question it belongs to. Replaying the recorded questions through the live resolver reproduces the recorded outcome every time. Those three checks are the whole of the evidence that the matching works; they say nothing about questions nobody has thought to ask yet.
6 figures in the catalog cannot be reached by asking, because no question has been written for them. They are reachable from the catalog and from the page they live on. Asking a question changes nothing on this estate, so nothing is written to the evidence spine. The attempt log above is the record.