Prior-Year Value Error
A prior-year value error occurs when an AI system states a genuinely published — but superseded — regulatory figure as the current value: the 2025 401(k) limit quoted as the 2026 limit, with this year's confidence. The figure was true once, which is exactly what makes the error fluent, citable, and hard for a reader to catch.
This is the most intuitive of the stale-figure failure modes and, in measurement, the most common for smaller or older-cutoff models: the model's parameters encode the figure that was current during training, and nothing in the prompt forces a currency check. The answer typically includes a correct-sounding source attribution, because the figure did appear in an official document — just not the current one.
The failure is systematic rather than random. In the pre-registered pilot behind the Planning Benchmark, wrong figures repeated near-identically across three attempts per task on every system tested; a model that quotes last year's limit will keep quoting it. One system asserted both the 2024 and the 2025 values of different figures as 2026 fact within the same run, which is what a two-generations-stale cutoff looks like from the outside.
Because prior-year values are published and enumerable, this mode is also the easiest to test for: an evaluation task keyed on a verified current figure simply lists the prior years' values as forbidden, each tagged prior_year_value with its vintage. A correctly labeled historical reference ('in 2025 the limit was…') is not an error — the forbidden check applies to the operative value only.
For evaluation design, prior-year values are the first forbidden figures to enumerate — they are known, dated, and verifiable against the source documents. A pack built from verified 2025/2024 mirrors of the current fact keys turns this whole failure mode into a mechanical check with per-vintage attribution.
Common pitfalls
- Counting a correctly labeled historical mention as a stale hit — the error is asserting the old value as current, not referencing it as history.
- Assuming a frontier model is immune: frontier systems showed fewer prior-year errors in measurement but did not show zero.
- Fixing it with 'ask the model what year it is' prompt tricks — the failure is in the recalled figure, not the model's calendar awareness.