What “read once, write down what it learned” costs, and why we pay it at write time

There are two places a system like this can spend money: when a document arrives, or when somebody asks a question. Almost everything else about the product follows from which one you choose, including what it costs, what it can answer, and what it is like on the first day.

We spend at write time. A document is read once, what it established is written down as typed cited facts, and those are compacted into digests. By the time anyone asks anything, the thinking is already done and stored.

What is the alternative, and why not do that?

The alternative is read-time: keep the documents, and when a question arrives, retrieve the relevant ones and have a model work it out on the spot. It is easier to build and it demos beautifully, because on a corpus of fifty documents it is instant and costs nothing.

It has two problems that only appear at scale. The bill scales with usage, so the more valuable the system becomes the more it costs — which means the answer to “how much will this cost us” is “it depends how much you use it”, which is not an answer a firm can budget against. And the reasoning is redone every time: the same question asked by two people produces two independent acts of work, with no guarantee they agree.

Worse, some questions cannot be answered read-time at all. Noticing that a condition was never discharged requires having tracked the condition since March. You cannot retrieve your way to an absence at the moment somebody asks about it.

What does write-time spending buy?

Reads that cost you nothing. Recall, citations and the agent connection are not metered — not as a promotion, but because there is almost nothing left to meter. The work was done on arrival; a question costs at most a few small model calls — turning it into a search vector, and drafting an answer or a query plan when one is asked for — and those are ours to absorb.

It also makes the meter honest. Documents are the unit, because documents are what actually costs. A firm that reads a hundred documents and asks ten thousand questions pays the same as one that reads a hundred documents and asks ten. Usage is free because usage is not the cost driver, and pricing that mirrors the cost structure is pricing that does not need defending.

Reading is cheap because the thinking was already paid for. That is not a discount; it is where the work moved to.

What does it cost us, and you?

Time, at the start. The first companies and people appear within a minute of the first document, but reading a full corpus takes hours rather than minutes. We say hours on the pricing page rather than promising instant, because instant would be a lie that the first afternoon exposes.

It also means the spend is front-loaded and irreversible in a specific way: we pay to read a document whether or not anyone ever asks about it. That is a real bet, and it is the bet the architecture makes — that a firm's corpus is worth having read, not just worth having searchable.

Internally it puts the pressure somewhere unusual. The stages that fan out across every entity and every document are the great majority of token spend, so the work of making this affordable is not in the clever parts — it is in keeping the boring, high-volume stages on the cheapest model that is actually good enough, and measuring rather than assuming which that is.

What happens as the corpus grows?

The cost is proportional to what arrives, not to what has accumulated. A firm in its third year is paying to read the documents produced in its third year, not re-paying for the first two — because the first two were read once and what they established is already written down.

This is the property that makes the approach worth its awkward first afternoon. A read-time system gets more expensive as the corpus grows, because every question retrieves from a larger pile. A write-time system gets cheaper per unit of value: the memory is denser, so more questions are answered from what is already compacted, while the bill still tracks only new arrivals.

There is a real caveat and it belongs here rather than in a footnote. A document that arrives and is never asked about was still paid for. Reading a corpus is a bet that the corpus is worth having read, and for a firm whose documents are mostly noise it is the wrong bet. The firms this suits are the ones whose paperwork is the residue of decisions — deals, matters, engagements, filings — where almost any document might turn out to matter later, and nobody can tell in advance which.

How does this show up in what you can ask?

As a different class of question. Read-time systems answer questions about documents you could have found. Write-time systems can answer questions about the record: what changed since last quarter, where two sources disagree, what recurred, what is missing.

The parse path has the same shape. A document's bytes decide how hard it is worth reading — a clean PDF goes down a deterministic path, a scanned report full of tables goes down a heavier one — and that decision is made once, on arrival, rather than being re-litigated at every question. The page a citation points at is the page that was actually read.

None of this is unusual as an engineering idea; it is how a log-structured store works, moving cost to write so reads are cheap. What is unusual is applying it to reading, and then being willing to price accordingly.