A short piece of fiction, written by API Evangelist. The reader is imagined; the post, and every number and quotation from it, is real: "Healthcare: A Bundle For The Shape Of The Answer," Kin Lane, November 4, 2026, https://apievangelist.com/2026/11/04/healthcare-a-bundle-for-the-shape-of-the-answer/
Marisol worked as a pharmacy informatics analyst, and she read the post on her lunch break with a cold coffee beside the keyboard. She had expected another piece about how models get medicine wrong. It was about the check that runs after generation.
She liked the narrowing first. The status enum was cut down to draft, and the post called that the strongest safety property in the file, one word of work. She thought about how many meetings she had sat through trying to get a human step written into a workflow, and here it was a single value that the decoder could not step around. The patient and requester were bound from the request rather than extracted, and that struck her as obvious in the way good ideas are after someone else has had them. The post put it this way: "Anything the system already knows should come in as context and be bound into the shape, not extracted."
Then she got to the fixtures and her mood changed. A draft with the 20 mg code, a dosage of 40 mg twice daily, an impossible February date and a fabricated quote was accepted by the decoder. The rules caught the quote and the dates. They did not catch the code. She read the admission twice, because vendors rarely hand you their own counterexample. Change only the code to 314077, cite a true passage, and the post says "every check in this file passes while the record says 20 mg and the note says 10".
She felt uneasy, and the unease did not go away. The post said it would add a rule comparing each candidate's strength to the cited dose before anyone used this, which was sensible. But she noticed she did not trust the sentence about failures nobody has imagined yet. In her work the imagined failures were never the ones that hurt someone. She also wondered if a tidy file with retries and a human at the end might make a busy clinician sign faster, because everything upstream looked so orderly. The post did not claim otherwise, and it said it would rather show a small shape with a person at the end than a big one that implied more certainty. Still, she wasn't sure how much that honesty would help on a Friday afternoon.
She disagreed with one thing, quietly. Free-text dosage was a fair choice for a demo, but in her world that was exactly where the trouble started, and she did not want anyone reading the demo as a green light. The post agreed with her on that, which made it harder to stay annoyed.
What she did was small. She opened a blank file and wrote one fixture by hand: a note saying ten milligrams, two candidate codes, and a draft that chose the wrong one while citing a sentence that really was in the note. She titled nothing and promised nothing. She just wrote down what the record should say and what it said instead, and saved it next to the other test cases her team kept.
Then she sent the file to a colleague with one line asking whether anything in their current checks would catch it. She didn't know the answer, and the post was the reason she thought to ask. It had told her the difference between constrained and correct, and she wanted to see which side of it her own systems were on.
She finished the coffee cold, and went back to work.