Fact-Locked Grading: How MI Refuses To Hallucinate

Fact-locked grading means the criteria you are scored on are fixed to verified facts before grading starts, so the grader cannot invent the standard it judges you by. This is how Mythic Intel refuses to hallucinate: the facts about your role are researched, run through a second pass that strikes anything it cannot confirm, and only then frozen into a rubric. The model that scores your answer compares it to that locked rubric instead of improvising one. Grounded AI feedback is trustworthy for exactly this reason, because trustworthy AI feedback requires a fixed source of truth the model is not allowed to rewrite.

To see why this is necessary, you have to understand what hallucination actually is and why an ungrounded grader is so dangerous.

What hallucination really is

Hallucination is when a language model produces output that is fluent and confident but factually wrong or unsupported by any real evidence. It is not lying in the human sense and it is not a rare malfunction. Surveys of large language models describe it as an innate consequence of how they work: the model predicts the most probable next text, optimizing for coherence rather than truth, so when it is uncertain it does not stop, it generates the most plausible-looking continuation. That continuation can be a fabrication delivered in the exact same tone as a fact.

The important consequence for grading is that the model has no built-in sense of "I do not know this." Faced with a gap, it fills the gap convincingly. That tendency is harmless when you ask for a poem. It is dangerous when the model is deciding whether your interview answer was correct.

Why an ungrounded grader invents its own criteria

Put an ungrounded model in charge of scoring and the hallucination problem turns inward. The grader does not just risk getting a fact wrong in passing. It can hallucinate the criteria themselves:

• It can invent a "correct answer" that no real interviewer would expect, then mark you down for not matching it.
• It can reward a confident-sounding wrong answer because the answer matched its own mistaken belief.
• It can shift the standard between two of your attempts, because nothing pins it in place.

This is the worst kind of feedback, because it is wrong and certain at the same time. You cannot tell a hallucinated criterion from a real one by looking at the score. You would adjust your real interview answers to satisfy a standard that exists only in the model's output. An ungrounded grader does not just occasionally misjudge you. It can quietly make up the test.

How grounding stops it

The fix that the field converged on is grounding: anchoring the model's output to retrieved, verified information instead of its internal guesses. The research is consistent that grounding a model in confirmed evidence reduces hallucination, because the model is now reasoning over facts it was handed rather than facts it imagined. Fact-locked grading is grounding applied to scoring. The rubric is the retrieved, verified evidence, and the grader is constrained to it.

Two structural choices make the lock hold:

• Verification before locking. The facts are not just researched once. They pass through a second, independent check, in the spirit of Chain-of-Verification, the 2023 method that showed a model produces more accurate output when it generates separate verification questions and answers them independently rather than trusting its first draft. Anything that cannot be confirmed is struck, so the rubric is built only from claims that survived an adversarial pass.
• Compare, do not invent. During grading the model's task is narrowed to comparing your answer against the fixed rubric. It is not asked "what is the right answer," which would invite a hallucination. It is asked "does this answer match these verified facts," which it can do far more reliably.

There is a known trap worth naming. Research on self-verification found that a model loosely critiquing itself can actually make things worse, and using a model as its own naive verifier can collapse its accuracy. That is why the lock depends on independent, retrieval-grounded checks rather than a vague "are you sure?" loop. Structure is what makes verification real instead of theatrical.

What the lock means for your feedback

When grading is fact-locked, the system can be honest about its own limits, which is the real marker of trustworthy AI feedback:

• A wrong claim is flagged because it contradicts a verified fact in the rubric.
• An unverifiable claim is surfaced as unverified, not silently accepted and not falsely blamed on you.
• The model answer you see afterward is built from confirmed facts, so you are comparing your attempt to a verified target rather than to something the grader wished into existence.

The result is feedback you can act on. A low score points at a real gap, not at a mood. You can change your behavior because the signal is grounded in something true.

The standard, stated plainly

A grader that can hallucinate its own criteria is not a grader, it is a confident guess wearing a number. Fact-locked grading exists to make that impossible: research the role, verify it adversarially, strike what cannot be confirmed, freeze the rest, and hold the scorer to it. That is what lets the feedback be trusted at all.

The way to feel the difference is to use it. Answer out loud against a fact-locked rubric, read the grounded feedback, and rehearse again until your spoken answer matches the verified truth rather than just sounding right.