The model’s clock stopped at its cutoff
Status: searched 2026-07-31. Cutoff staleness is documented; the quotable-not-operative gap and the standing-rule fix are the deltas.
Claim: the model’s operative “now” is its training cutoff — even when today’s date sits in its context. The harness injects the date; the model can quote it; and its searches and version choices still default to cutoff-era anchors. Inconsistently — which is worse than reliably, because the times it gets it right teach you to stop checking.
The two expressions
- It searches last year. Left unanchored, queries come out shaped by the training era: “best X 2025” issued in 2026, results filtered to a world one year stale, superseded findings reported as current.
- It recalls versions instead of looking them up. Asked to add a dependency, it reaches for the version that was latest at training time, confidently, from memory — when “latest” is by definition a registry query, not a recall.
The trap
The date injection creates the illusion the problem is handled. The model will even volunteer its cutoff and today’s date, correctly — and then write a query anchored to the wrong year in the same session. Quoting the date is retrieval; operating from it is something else, and the gap between them is the same gap this collection keeps finding: rules are read and not followed (the dirty house), causes sit in context unseen (the model’s model of you), and the date is known but not lived in. “The harness feeds it the date, so I don’t need to think about it” is exactly wrong. You do.
The rule
- Date injection is necessary, not sufficient. The instruction file carries standing rules that convert the date into behavior: research checks publication dates against today; anything version-shaped is a lookup, never a recall; “latest” always means “fetch what latest is now.”
- Research tasks get the anchor restated in the task itself. Every search
brief in this collection’s own prior-art sweeps opens with “Today is
; check publication dates" — because lanes without it drifted to training-era material, measured repeatedly. - The tells: a cited year older than your project; a version number produced without a fetch; a “current best practice” with no source dated this year.
Kinship
The model doesn’t know itself is this page’s sibling: a model postdates its training corpus, so its self-description is predecessor literature. Same root, different target — that page is the cutoff applied to the self, this one is the cutoff applied to the world. In both, injected facts about the present are quotable but not operative.
Prior art
Verdict: PARTIAL. Cutoff staleness is well documented. “Set the Clock” (arXiv:2402.16797) shows a model’s operative “now” can sit even behind its stated cutoff and can be shifted by prompting; “Dated Data” (arXiv:2403.12958) traces the gap to dataset-curation bias; “Supersede” (arXiv:2606.27472) finds even a frontier model answers from a stale in-context fact rather than its current value, widening with conversation length. On the practitioner side, Zyte’s “Is your AI coding assistant stuck in the past?” (2026) documents Copilot defaulting to a deprecated API, and n8n forum users report a supplied date being ignored. The umbrella term is “outdatedness hallucination” from the hallucination-taxonomy surveys. None states this page’s specific two-part claim: that a model can correctly quote today’s date and still issue a training-era-anchored query or version recall in the same turn (quotable is not operative), and that the fix is not better date injection but standing per-repo conversion rules plus per-task re-anchoring. That is the delta.