Catching the Confident Error

AI output does not fail the way a tired employee's work fails, with visible sloppiness that invites a second look. It fails fluently: correct grammar, confident tone, professional formatting, wrapped around a fact that is not true. The polish is the camouflage, and the reviewer's instinct, "this reads fine," is precisely the instinct that misses it. This lesson is the field guide: the four errors that survive a casual read, and the check that catches each one.

Error one: the invented fact

The assistant needs a detail it does not have, and instead of stopping, it produces something plausible: a spec for a furnace model, a warranty term, a delivery window. Fluent, specific, and fabricated.

The check: trace-back. For any fact that would embarrass you if wrong, ask where it came from: the Brain, the pasted notes, or nowhere. You met this on Fridays; it generalizes to every yellow checkpoint. The system-level defense you have already built is the pair of skill rules from Chapter 3: [PRICE NEEDED] placeholders and ESCALATE. A skill allowed to say "I don't have this" invents less, but the checkpoint assumes some inventions get through anyway.

Error two: the stale fact

Nothing was invented; the fact was true in March. The old service-call rate, the promo that ended, the tech who no longer works there. Stale facts are the sneakiest of the four because trace-back passes: the fact did come from the Brain, an outdated one, or from an old conversation still in the assistant's context.

The check: date the source. The Brain's changelog and dated Right Now section exist for this moment. When a time-sensitive fact appears in a draft, the question is not "is this in the Brain?" but "is this in the current Brain?" And the deeper defense is the Chapter 2 ritual itself: staleness is caught in bulk, monthly, for fifteen minutes, rather than retail, at checkpoints, forever.

Error three: the plausible number

Arithmetic on wrong inputs, a subtotal that quietly excludes a line, a percentage computed against the wrong base. Numbers are where fluency does the most damage, because a column of figures with a total underneath looks like it has already been checked.

The check: one independent number. Do not re-verify everything; pick the number that drives the decision and recompute it yourself, from source, on a calculator. If the anchor number survives, the draft has earned the benefit of the doubt. If it does not, nothing else in the draft gets that benefit, and the run is redone from correct inputs.

Error four: the tone miss

Every fact right, and the whole thing lands wrong: too breezy for a customer whose basement flooded, too stiff for a fifteen-year regular, cheerful boilerplate under a subject line about a missed appointment. Tone misses are errors of situation, not of fact, which is why fact-checks sail past them.

The check: read it as the recipient. One sentence aloud, heard as the specific person receiving it, is usually enough. The system-level defense is the escalation rule's situational triggers, an upset customer already routes to you, plus voice rules in the Brain concrete enough to check against. "Warm but professional" catches nothing; "no exclamation marks in anything about money" catches something.

The checkpoint, armed

Notice the pattern across all four checks: none of them is "read it again carefully." Each is a targeted question that takes under a minute, which is what makes them survivable on a busy Monday, teachable to an employee, and honest in a runbook. A named check gets done; "review carefully" gets skimmed.

Try it now

Take three yellow-lane drafts your system produced this week and run all four checks on each, timing yourself. Two things tend to happen: at least one draft trips a check you would have skimmed past, and the full pass comes in under four minutes. Then update your runbooks' checkpoint lines to name the checks that caught something. Your checkpoints are now calibrated to your actual failure pattern, not a generic one.