Every number with a receipt: how the Casebook is made
Each case in the Casebook ships the prompt that produced it, the queries behind every figure and what rerunning them costs. Here is the process, and the rules that keep it honest.
On this page
One rule: the public API only
Every case in the Casebook is made by an AI agent working through the same public Metix AI Platform our customers use: the REST API, the MCP server or the agent skills, with a normal key. No internal database or export feeds a case. If someone with a key cannot reproduce a number, the number does not belong in the Casebook.
That rule is what makes the rest possible. Because every figure comes from a public call, every figure can ship with the call that produced it.
Count before you read
A search that asks only for the total costs 1 API Credit. Reading a record costs more. So every case counts first: it builds its population one condition at a time and looks at each total before reading anything. Most cards never read a record at all. The cheapest replay in the Casebook is six counts and 6 API Credits.
Counting first also catches mistakes cheaply. When one condition removes far more than the others, it is usually measuring something other than what we meant, and a count shows that before any money goes to reading records.
The receipt
Each case has a small script, fetch.py, that replays its committed queries without an agent and writes the aggregate files the page is built from. Before its first call and after its last, it reads the key's balance. The difference is the receipt: what the replay cost, measured by the Platform rather than estimated by us.
The page shows two costs and labels them: what the replay costs, and what an agent following the prompt spends, with the ceiling the prompt stops at. What it cost us to explore before settling on the queries is recorded too, and never presented as the price of a rerun.
Counting people without naming anyone
Cases that count people follow a small-cell rule. A group of fewer than 10 people is published as <10, never as an exact number, a share or a bar length. We also check the differences a reader can form: if a total minus its parts would give back a group under 10, one of those numbers is withheld.
This came up while writing this piece. A table of which skills a group lists had one row that, subtracted from the group's total, left four people. We removed the row before publishing. No case names a person, shows a profile or keeps records in the repository; the records a run reads stay on the machine that ran it.
The prompt is the source
Every case ships the prompt that produced it, in ten numbered steps under the question: what to read before querying, how to define the population, what to count, the API Credit budget, an audit of the results, what to output and what the numbers cannot say. The site checks at build time that the prompt names the same API Credit ceiling the page promises, so the page cannot say one thing and the prompt another.
The prompt is also the part a reader keeps. Change the title, the city or the company in one step, and the same method answers a different question on your own key.
What making a case costs
The four use cases published with this piece cost between 9 and 31 API Credits to replay, and every one fits in the 100 free API Credits a new account gets. Create a key and pick one from the task list.