What 'human-in-the-loop' actually means
“Human-in-the-loop” is one of those phrases that’s true of almost any AI product if you squint. We hold ourselves to a stricter version of it, because in admissions the loop is the whole point.
Three rules we designed around
The AI never holds a verdict. ProgramApply produces summaries, rubric scores, and flags. It does not produce decisions. There is no “auto-admit” threshold, no silent reject. The system’s job ends at recommend and prepare; the decision begins with a person.
Every output shows its work. A score of 0.62 on its own is just a number to
distrust. So every score links to the rubric criteria it was measured against, and
every claim in a summary links back to the line in the document it came from. A
reviewer should be able to verify any recommendation in a click — and overrule it
without friction.
Overrides are first-class, not exceptions. When a reviewer disagrees with a flag or a score, that’s not an error state. It’s signal. Overrides are logged, easy to make, and feed back into how we report on the system’s behavior over time.
Why the receipts matter
Admissions offices get audited — by leadership, by counsel, sometimes by the public. A recommendation you can’t explain is a liability, no matter how good it is. So the audit trail isn’t a compliance afterthought bolted on at the end; it’s the substrate. Every summary, score, and flag is timestamped and traceable to its source.
The goal isn’t an AI you obey. It’s an assistant you can interrogate, and a record you can stand behind.
← All field notes