TITLE: A model's derivation counts only when it reproduces the reference
  A model writes the {admiral} program for one ADaM variable
  The program runs in isolation on public SDTM data
  Its column is accepted only on an exact match with the reference
  Recorded from demo.adam.chernov.ca
Two models each wrote one {admiral} program for every one of 159 ADaM variables
  Each column was accepted only if it equals the published reference exactly
gpt-5: 118 of 159 accepted. A 7B coder model on a laptop GPU: 35 of 159.
A derived date: the program gpt-5 wrote, accepted on an exact match
  Screen failure date, from the disposition records
The 7B model's program for the same variable stops with an error
The reference leaves AEACN empty. gpt-5 filled all 1191 rows anyway.
  An action taken for every adverse event, inferred from the dosing records
The comparison counts the differing rows and lists them: empty in the reference, filled by the model
Re-run it here: R starts in the browser with webR, on the same data and the same check
Your browser reaches the same verdict: 1191 of 1191 rows differ
  All 318 recorded verdicts reproduce this way
CARD: What the check makes possible
  A model in the first-line seat, with QC as the judge
  Every program and every verdict on the record
  Any verdict re-checked without trusting the machine that made it
  Public pharmaverse data. The protocol was committed before either model ran.
END: demo.adam.chernov.ca
  Every program and verdict, re-run in your browser. More at adam.chernov.ca