What Survives Cross-Examination Becomes Your Room

Verified interview content survives cross-examination. Everything else gets struck. The editorial rule is simple and unforgiving: if a public source contradicts a claim, it is wrong and goes; if no source addresses a claim at all, it is unconfirmed and also goes. What is left, the set of facts that held up under a second adversarial pass, becomes the material you rehearse against. This is fact-checking applied to interview prep, and it is the single discipline that separates trustworthy preparation from confident guesswork.

The reason this matters is mechanical. In an interview, a strong claim with a weak foundation is a liability, not an asset. Generative AI is good at producing impressive-sounding metrics and specific-sounding detail. It is not good at making sure any of it is real. A single follow-up question from an interviewer destroys an answer built on a fabricated number, and it takes the rest of your credibility with it. So the safest content is content that has already been cross-examined before you ever open your mouth.

Why a second pass beats a first draft

Most AI prep tools stop at generation. They produce a plausible-looking set of facts about the role and the company, and they hand them over. The problem is that the first draft is exactly where fabrication lives. Models are trained and evaluated in ways that reward confident answers over honest uncertainty. OpenAI's September 2025 paper "Why Language Models Hallucinate" framed this precisely: hallucinations behave like errors in binary classification, and when grading rewards only whether an answer is right, a model that guesses scores better on average than one that says "I do not know." The model learns to bluff. A first draft is full of those bluffs, presented with the same fluent confidence as the true parts.

A second pass changes the job. Instead of generating, it interrogates. It takes each claim from the draft and asks two questions: does a source confirm this, and does any source contradict it? Claims that survive both questions stay. Claims that fail either one are struck. The output is smaller and less impressive-looking than the first draft, and that is the point. A short list of facts that are true beats a long list that is mostly true.

The two failure modes worth striking

There are exactly two ways a claim fails cross-examination, and both end the same way.

• Contradicted. A public source says something different. The team is described as using one stack when their engineering blog and job posting describe another. The product is positioned for one market when their own site says another. Contradicted claims are the most dangerous because they are specific and wrong, and an informed interviewer will catch them instantly.
• Unaddressed. No source confirms the claim either way. This is the quieter failure. The claim might be true, but you have no basis for asserting it, and if you assert it as fact you are gambling. Unaddressed claims should be demoted from "things I will state" to "things I will ask about." A good question signals curiosity; a wrong assertion signals carelessness.

Treating "unaddressed" as a strike, not a maybe, is the hard part of the discipline. It is tempting to keep a claim because it sounds right and would make the answer stronger. That temptation is exactly the bias the second pass exists to overrule.

What grounding in real sources actually requires

Striking the unconfirmable only works if you check real sources, not a model's memory. Training data has a knowledge cutoff, so a model has no inherent awareness of anything that happened after it was trained, and it cannot reliably tell you when its own knowledge stops. Recent reorganizations, new products, a leadership change, a shift in the team's tooling: all of it can be invisible to a model running on memory alone, and the model will still answer confidently about the stale version of the world it remembers.

Grounding fixes this by pulling current detail from the live web at the moment of research, then verifying each claim against what those sources actually say. The relevant sources for interview prep are concrete:

• The company's own site and engineering or product blog
• The live job posting and related public role descriptions
• Current documentation for any tool or framework named, because tools change and yesterday's behavior may be wrong today
• Reputable third-party coverage for anything about scale, funding, or recent direction

Cross-reference rather than trust a single hit. A claim confirmed by the company's own words is stronger than one inferred from a forum post, and a claim that appears nowhere is a claim you do not get to make.

This verify-then-strike loop is the editorial spine of how Mythic Intel assembles a room. It researches the exact role on the live web, then runs a second verification pass that removes anything public sources do not support, so the spoken questions, the course, and the fact-locked rubric all rest on content that already survived cross-examination. You are never rehearsing against a guess.

What survives becomes your room

The output of this discipline is a smaller, harder, more reliable body of material. Every fact in it has been contradicted-checked and confirmed-checked. Every claim you cannot support has been demoted to a question. What remains is content you can defend under pressure, because it has already been pressure-tested. That is the whole game: an interview is itself a cross-examination, so the preparation that survives a cross-examination first is the only preparation that will hold up when a real interviewer pushes back.

Once the content is verified, the work shifts from knowing to delivering. Take each surviving fact and the answer it supports, and say it out loud, repeatedly, until the true version comes easily and you stop reaching for the impressive-but-unconfirmed version your draft once offered. The facts that survived cross-examination are the ones worth saying with your full voice.