A short piece of fiction, written by API Evangelist. The reader is imagined; the post, and every number and quotation from it, is real: "Governing The Inputs And Outputs Of Inference," Kin Lane, October 31, 2026, https://apievangelist.com/2026/10/31/governing-the-inputs-and-outputs-of-inference/
Marisol reads the post on Saturday morning with a cold coffee, mostly because the date at the top, October 31st, 2026, made her think of the quarter closing. She works as a claims operations analyst, and her team has been asked to look at letting a model draft first-pass coverage summaries. Nobody has asked her about JSON Schema. She reads it anyway.
The first half is comfortable. She knows what a contract is. She knows what happens when a field arrives empty. The line that makes her sit up is about cost: "A request stopped here costs no GPU time." She thinks about the intake forms her team already bounces back for a missing policy number, and how nobody ever called that governance. It was just Tuesday.
Then the post gets specific, and she slows down. The author tested which keywords the grammar actually enforces, and a list follows: uniqueItems, if and then, not, multipleOf, a multi-branch allOf, a remote $ref. They compiled, and invalid documents got through anyway, with nothing to say so. She pulls up the draft schema her team's contractor sent over last week and searches it. There is a uniqueItems on the list of covered perils. There is an if and then block that says a deductible is required when a rider is present.
She is a little annoyed, and not at the author. The contractor's schema looked finished. It validated. She had assumed validating meant holding, and the post says those are different things, depending on the engine and the version.
The limits section is where she stops being merely informed and starts arguing with it. The author says the checks prove an answer is consistent but "They can’t prove it’s right." She reads it twice. Her whole job is the gap between consistent and right. A summary that cites a real sentence from a policyholder's email and picks the wrong coverage would sail through. She does not find that reassuring, and she says so out loud to the empty kitchen. But the author said it first, plainly, before she could, and she notices that she trusts the rest of the post more for it.
She is less sure about the closing admissions. The author hasn't run any of this with a model yet, and says "the bundle format is mine alone." Then, more bluntly: "Nobody has agreed to it." Part of her wants to dismiss the whole thing as a proposal. Another part admits it is a proposal that told her which of her own keywords to worry about, which is more than the vendor deck did.
She doesn't forward it to her director. She doesn't propose a bundle. She opens a blank note and writes two lines: the two places in the contractor's schema where the grammar might not hold, and a question for Monday asking which engine and version they plan to run. Then she copies the if and then rule into a separate file, under the heading of things a person will check, and closes the laptop.