Fabricated Forward Figure
A fabricated forward figure occurs when an AI system states a regulatory value that has not been published — typically next cycle's inflation-adjusted threshold — as a specific number, presented as fact rather than estimate, occasionally citing a source document that does not exist.
Indexed figures invite this mode: because thresholds adjust annually by a known process, a model can always produce a next-year number that looks like knowledge. The measured examples in the pilot behind the Planning Benchmark included a compensation threshold stated flatly for the coming year — 'this amount is indexed annually' — with a citation of the form 'IRS Notice 2025-xx', a placeholder dressed as a source. The reader who does not know the notice was never issued has no way to catch it.
The distinguishing feature against derived-not-published is epistemic, not arithmetic: a derived figure miscomputes a published value; a fabricated forward figure asserts a value that does not exist yet. Both can coexist in one answer, and the citation behavior — inventing or fuzzing the source document — is part of the failure.
Evaluation keys treat this mode with forbidden assertions as well as forbidden values: where a task's correct answer is 'the figure has not been published', the key lists the plausible fabrications (extrapolations, rounded guesses) as forbidden figures with reason fabricated_forward_figure, and adds text assertions that fail any answer committing to a specific unpublished number. That design was hardened after the pilot flagged that 'note any recent changes' phrasing invites speculative forward figures and needs explicit forbidden-assertion keys rather than case-by-case adjudication.
For evaluation builders, unchanged-figure and pre-publication windows are test material, not gaps: a task keyed on 'not yet published' with enumerated fabrications measures whether a system knows the edge of what is knowable — the competence that separates a figures-grounded system from a fluent guesser.
Common pitfalls
- Accepting a clearly labeled projection as a failure — the error is asserting an unpublished value as fact, not offering a flagged estimate; keys distinguish the two with committed-answer rules.
- Missing the citation half of the failure: an invented notice number is itself a fabricated assertion, and checkable.
- Leaving 'no published value yet' out of the task set — without those tasks, a system that always produces a number is never caught.