How Crux uses AI

Version 1.0, effective September 16, 2026. This statement describes what artificial intelligence does inside Crux, what it does not do, its known limits, and how people stay in charge. It is written for students, professors, and the offices that review tools like this.

The models

What the AI does

• Reads the professor's case, rubric and settings and drafts an interview plan: an opening, probes for each rubric criterion, surface-knowledge checks, and pushback in the chosen persona's voice. The professor can edit every question.
• Reads the student's uploaded work (if any) and plans questions about the student's own claims and numbers.
• Conducts the interview: one question at a time, following up on what the student actually said, re-asking once if a question was not answered, and moving on so that every rubric criterion gets its chance.
• Checks each number the student states against the case, the professor's answer key, and a recomputation, and labels it correct, an arithmetic slip, a wrong input, a wrong method, a defensible alternative assumption, a transcription doubt, or unverifiable.
• After the interview, maps the transcript to the rubric with quotes and timestamps, writes private notes for the professor, suggests a band with its reasoning, lists misconceptions, and writes feedback to the student.

What the AI does not do

• It does not assign grades. The professor does, with the evidence in front of them.
• It does not identify people. No voiceprints, no face templates, no matching against a roster. The name on an interview is the name the student typed.
• It does not detect cheating. It records what was said; the professor judges.
• It does not evaluate accent, pace, filler words, tone or "confidence." Only the content of what was said.
• It does not reveal answers or say whether a number is right during a graded interview. It asks the student to walk through it. Professors can switch on a formative mode that reveals after one attempt.
• It does not learn from students. No data from Crux trains any model.

Known limitations

• Transcription errors. Numbers are the most common casualty ("fifteen" and "fifty"). The audio clip is the record; the numbers check flags doubtful transcriptions; the interviewer is told to ask "fifteen or fifty?" when it matters.
• Non-native speakers and accents. Automated transcription is less accurate for some speakers. Crux has not yet measured this systematically on its own interviews; doing so before any graded use is an open task. The typed-answer accommodation exists for this reason among others.
• Judgment calls. Whether an answer shows "mechanism" or just "recall" is a judgment the model makes and the professor can overrule. Suggested bands are calibrated to a written scale but are not consistent to the letter grade.
• Persona drift. The interviewer sometimes praises or affirms when it should stay neutral. Rules and server-side checks reduce this; they do not eliminate it.
• Pacing. The interviewer may spend too long on one area. Crux enforces a limit: after two attempts on one criterion it must move on, and it may not close while a criterion is untouched.

Human oversight

• A professor reviews every interview before any grade exists, with the transcript, the audio, and the evidence map side by side.
• Students can contest an evaluation with their professor, who can replay the recording.
• The operator reviews interviewer behavior in transcripts (never audio without cause) to fix defects, and logs every such fix.
• Any student may request their record or its deletion.

Transparency to students

Students are told before they begin that an AI conducts the interview, that it is recorded and transcribed, and that their professor decides the grade. The persona's name and role are shown in the lobby. The interviewer never pretends to be a human.

Contact

Questions about this statement: jon@cruxcall.com.