The 3-Check Method: how to trust AI in high-stakes work
A confident, perfectly-formatted, completely wrong answer is the most dangerous thing an AI can hand you. Here's the sixty-second discipline that catches it before it reaches a decision.
Most advice about AI lands in one of two camps: use it for everything, or don't trust it at all. Both are wrong, and both are useless to someone who actually has to get work done in a regulated environment.
The truth is narrower and more useful. A large language model is a very fast analyst—usually right, sometimes wrong, and always worth checking. The skill isn't trusting it more or trusting it less. The skill is knowing exactly where it breaks, and building a quick habit that catches the break before it costs you.
That starts with naming the failure modes.
The three blindspots
These aren't bugs. They're built into how language models generate text. Knowing the names is half the defense.
1. Hallucination
The model fabricates a regulation with the correct format and complete fiction. A legal team once filed a brief citing cases that didn't exist; the judge was not amused. For contracting and compliance work, this is the dangerous one.
2. Context drift
The assistant forgets the rule you set 2,000 words ago and quietly reverts to its defaults instead of your requirements.
3. Inconsistency
Same prompt, three different answers, all plausible. For compliance work, that's not a quirk—it's a governance problem.
The blindspots aren't a reason to avoid these tools. They're a reason to use them with verification, structure, and your judgment in the loop.
The best AI users aren't the ones who trust the technology most—they're the ones who understand where it breaks.
The 3-Check Method
Of the three blindspots, hallucination is the killer. The model can produce a Federal Acquisition Regulation citation with the correct formatting, the right numbering style—and a completely fabricated dollar threshold. It will sound right. It won't be right.
So here's the routine. Run it every time the model cites a regulation, a CFR section, a dollar threshold, or any binding number. It takes about three minutes—sixty seconds a check. It's the difference between a productivity tool and a liability.
Source Check
Ask the model to cite the exact part, subpart, and clause. If it can't—or if the citation looks off—stop and verify before you go any further.
FAR Part Check
Open the source and verify the actual regulation text. Sixty seconds, non-negotiable for compliance work. Heads up: there are now two FARs in active circulation—the legacy FAR and the FAR Overhaul. Both are valid; which one applies depends on your agency's transition status. Verify against whichever your office uses.
Own Eyes
Read it yourself. Thresholds, effective dates, and recent amendments are exactly where the model drifts. If you sign it, you own it.
Why it works—three jobs at once
The 3-Check isn't just a one-time checklist. It does three things:
- →It's the defense against the hallucination blindspot—the one that can actually hurt you.
- →It's the tool you run in the moment—every citation an AI surfaces gets the 3-Check before you trust it.
- →It's the takeaway for your team. Source Check · FAR Part Check · Own Eyes. Put it on a card next to the monitor.
And while the example here is federal acquisition, the method travels. Anywhere AI hands you a binding citation, number, or threshold—legal, finance, healthcare, insurance, tax—the same three checks apply. Get the source. Verify the source. Read it with your own eyes. The domain changes; the discipline doesn't.
This is what we mean when we say disciplined AI, not AI hype: not slower work, not more fear—just a habit that lets you move fast and trust what you ship.