Errors in numbers, dates, and names are the hardest to notice and the most damaging

What it is

A company history says “founded in 1989”; the year is 1998. Title, narrative, neighbouring sentences are right; what is wrong is one digit in a four-digit year. Readers do not dispute the frame “there is a founding year,” and do not treat every numeral as a suspect. Specific-detail errors are substitutions in slots — numbers, dates, names, model IDs. They draw the least attention and, once adopted, often the most harm: money, identity, and deadlines hang on those slots.

A false sentence or a false section snags reading. A false slot does not.

Why it happens

Reading grabs the propositional frame (“there is a founding year”); the slot is filled by a fast path that rarely runs a semantic check. A well-formed number looks true: a four-digit year, a dose with a unit, a proper name that looks like a name. Generators also have strong priors on these slots — common years, common names, common doses — so a wrong slot still “looks right.”

Harm is asymmetric. Frame right, slot wrong: downstream systems execute the slot — the wrong account, the wrong drug, the wrong party. A wrong frame is often stopped before adoption, because the whole sentence fails the task.

Studying it

In texts with a correct frame, swap slots: year ± several years, one digit of an amount, a same-surname name. Do not hint that anything is wrong. Dependent variables: spontaneous detection, detection after “please check the numbers,” task harm after adoption (simulated). Independent variables: slot type, whether the format is legal, whether slots are styled differently from body text.

The baseline is spontaneous detection with no prompt. Detection after a prompt overestimates everyday noticing. Use frame errors (a sentence irrelevant to the task) as a control, to show how much more often slot errors are missed.

Where it stops holding

When slots are constrained by an external system (a date picker, a name from the address book, an amount from an API), the generator cannot write a false slot and the claim weakens. When the user has episodic memory for that number (their own birthday, a quote they just saw), a wrong slot hits memory. Pure estimates or order-of-magnitude talk (“on the order of hundreds of thousands”) are not slot promises. Slots in code and tables are sometimes easier for tools to check than slots in prose. This entry does not treat the cost structure of exhaustive sentence scanning when errors are scattered.

Applying it

• Lift numbers, dates, names, and identifiers out of prose into fields, styled unlike body text, and prefer pick-or-cite over free generation.
• For slots that must still be generated, show the source’s original value beside them; mark a conflict rather than writing silently.
• Before copy, export, or submit, list slots on their own for field-by-field confirm, rather than asking people to reread the essay.
• Check: on a text whose frame is entirely right and three slots have been changed, give no hint, and count how many people adopt. High adoption with wrong slots means checking is still reading frames, not slots.

Related

• Same group: L3.03.1 Fluent wording is not the same as being correct · L3.03.2 The cost of checking can exceed doing the work oneself · L3.03.3 High-stakes settings must not leave checking entirely to the user · L3.03.4 When errors sit inside correct content, checking is sentence by sentence and costs about as much as rewriting · L3.03.5 The less familiar the domain, the harder checking is — and that is exactly when people ask the system · L3.03.6 Certainty of wording has no relation to reliability of content · L3.03.8 Putting the checking duty in a disclaimer does not reduce how far errors actually spread
• Nearby: L3.02 Source Attribution · L3.08 Source Attribution of Generated Results
• Search terms: specific-detail errors · entity slot error · numerical hallucination