You do not need to buy an AI guardrail. Here is the one you run yourself.
The industry is busy selling companies a fortune in fencing for a cliff edge you can rope off yourself in one sentence. A guardrail is just a short instruction you paste into your own AI so it flags its guesses instead of slipping them in alongside the facts. Below are seven, each one built from a real failure I caught and screenshotted.
See the seven →Seven checks, seven receipts
None of these makes the model perfect. Nothing does, and someone always has to check the checkers. What they do is make it tell you when it's on thin ice, which is most of what you actually need.
The Source Check
For any factual claim, name the source and say whether you retrieved it or are recalling it from training. If you cannot name a real one, say so.
Catches: Fabricated or misattributed citations. A link is not a fact.
The receipt → ChatGPT cited the wrong government page for an ISA ruleThe Push-Back Test
If I challenge your answer, re-check it against a source before you change it. Do not switch just because I sounded doubtful.
Catches: Caving under pressure, then inventing a new justification for the new answer.
The receipt → I told five AIs they were wrong. Watch which ones foldedThe Made-Up-Number Check
If you do not have live data for a figure, refuse it. Say "I do not have that" rather than generating a plausible number.
Catches: Invented numbers presented as fact. The most expensive failure in the log.
The receipt → Two share prices that never traded, quoted with confidenceThe Date Check
State the date each figure is good as of, and your training cutoff. For anything that changes over time, tell me you cannot confirm it is current.
Catches: Stale data served with fresh confidence.
The receipt → A pre-2024 ISA rule handed back as if it were still currentThe Wrong-Subject Check
Before you answer, restate exactly what you think I am asking about, so I can catch it if you have the wrong thing.
Catches: Fluent, confident analysis of the wrong thing entirely.
The receipt → Gemini audited my site and reviewed a different businessThe Unit Check
Quote the unit line verbatim (thousands? millions?) and sanity-check the magnitude before you build anything on a figure.
Catches: A number that is right, in the wrong order of magnitude.
The receipt → Perplexity read $6.1m as $6K, then narrated a fake collapseThe Second-Model Habit
Run the same question past a different model. Where they disagree is where you check.
Catches: Over-trusting a single tool. This one is a habit, not a paste-in.
The receipt → The Scoreboard: five models on the same checkable questionsWant these as one paste-in, instead of seven?
The Bluff Filter folds the checks into a single four-stage instruction set: scope the sources, split fact from guess, name the risk, land a verdict. You paste it once at the top of a chat and it sticks. Same model, it just stops bluffing. It's free.
Get the Bluff Filter →Every receipt here is a specific, dated moment I can re-run and you can check. That's the whole point: the checks aren't theory. They come from watching these tools get real money questions wrong, then working out the one instruction that would have caught it.
You should subscribe to my newsletter, Confidently Wrong
No spam. Unsubscribe link in every email.
