A short piece of fiction, written by API Evangelist. The reader is imagined; the post, and every number and quotation from it, is real: "DeepInfer And The Shape Of The Answer," Kin Lane, October 20, 2026, https://apievangelist.com/2026/10/20/deepinfer-and-the-shape-of-the-answer/
Marisol read the post on Tuesday night, after the dishes, on her phone. She builds integrations for a freight brokerage, and the part about a shipment exception coming back as the carrier's own status codes made her sit up. She had spent the previous Friday untangling a webhook that choked because a model summarized a delay in a friendly sentence instead of a code.
The first half she skimmed. DeepSeek had open-sourced DeepInfer, Apache 2.0, and the README was full of kernels and expert parallelism that she would never touch. Then the post slowed down on three features, and so did she. A request returns the same result no matter what else is in the batch. Identical input and sampling parameters give identical output. Generation can be constrained to a JSON schema, a regular expression, or a grammar. She liked the line about an answer you can describe ahead of time. Her whole job was making things arrive in a shape the next system expected.
Then she got irritated. The post says the shapes already exist, in the APIs a business depends on. Marisol knew her company's shapes. Half the exception statuses lived in a spreadsheet that one dispatcher maintained, and two of the carriers sent codes that contradicted each other. The post's list of what gets harvested sounded tidy, with its more than 57,000 real request and response examples, but tidy catalogs belonged to companies that had done the boring work. She wasn't sure hers had. The post says as much, that the rest will be writing prompts that describe a shape in prose and hoping the model guesses, and she felt personally described.
She also doubted the easy reading. Determinism on Tuesday and Friday was nice, but a wrong answer repeated identically was still wrong. A constrained decoder could only promise that the output would fit. It couldn't promise the output was true. The post seemed to know this, since it names rules that check what a schema alone can't, run after generation, but it moved past that quickly. She noted the gap and didn't let it go.
Still, the question the post asks stuck with her, because it was her actual problem: Where do the shapes come from? She had been about to blame the model for the Friday failure. Really, nobody had ever written down what a valid exception status was.
She didn't open DeepInfer, and she didn't pitch anything to her manager. She opened the dispatcher's spreadsheet instead and copied the status column into a plain text file. There were eleven distinct values, two of them spelled two ways. She merged the duplicates, sorted the list, and saved it as the start of an enum, named for what it was. Then she messaged the dispatcher, asking only whether the list was complete and which codes the carriers actually sent. She went to bed without waiting for the reply.