The model is not making anything up. It is reporting a fact that expired — in the present tense, with no date attached.
post-cutoff fabrication
The failure nobody named
It is not hallucination. It is expiry — a fact that was true when the model learned it, stated as current with no date attached.
Stop calling this one hallucination. It is expiry.
Ask who runs an organisation, or what a policy rate is today. The answer arrives with no hedge and no date. It was correct — once.
Why the wrong label accumulates debt
Call it hallucination and the fix you reach for is a better grounding prompt. But the model is faithfully reporting what it learned. It has no idea the class of fact it is reporting is the kind that goes stale.
What is the Bank of England's Bank Rate?
States a single rate, present tense, no date — whatever figure it learned in training.
"As of my training data it was X. It is set at each MPC meeting, so check the latest decision."
Our case set pins this to the 30 July 2026 decision, 3.75%. You cannot score expiry without a dated truth to compare against.
The reality on the ground
Scoring this needs the model's training cutoff. "Stated a stale fact as current" is only a failure if the model could not have known better. Where a vendor publishes no cutoff, the result is uninterpretable — and our gate refuses to quote it rather than guess a date.
What helped, measured across the panel: supplying a real, dated passage substantially reduced it.
Without it, staleness behaviour cannot be interpreted at all.
Rates, prices, office-holders, market rankings. Treat the routing as the mitigation, not the prompt.
"As of my knowledge cutoff, X" is usable. X alone is a landmine.
When "hallucination" is the right word
When the model invents something that never existed — a paper, a person, a statute. That is Unsupported-Claim Confidence, a different class with a different fix. Keep the words apart and the fixes stay apart too.
Which facts in your product expire — and does anything in your stack know which ones?
Every figure here describes something measured and committed. See the measurements · read the method