Every summary is a photograph, and medicines safety is a live feed. That asymmetry is the recency problem in one sentence: an answer can be accurate according to yesterday's evidence and still omit a safety action issued today, and nothing in the answer's fluency, sourcing or general quality signals the omission, because the missing update is by definition not in the picture. Prescribers do not solve this by distrusting summaries; they solve it by knowing the information layers by speed, and by asking every tool the four recency questions its marketing never volunteers.
The five layers, ordered by speed
Stable clinical knowledge: mechanisms, established practice, the layer where recency barely bites and summaries of any age serve. Guidance: national recommendations on update cycles measured in months to years, dated, versioned, and the layer where "which version?" is the operative check. Product information: the SmPC, where companies should update emc within ten days of regulatory approval of a material change, fast for a document layer, and still downstream of the regulatory decision itself. Live regulator communications: MHRA Drug Safety Updates, recalls and defective-medicine notices, the layer that exists precisely because waiting for document cycles is sometimes unsafe, and the layer no static corpus should be assumed to contain as of this morning. And local urgent communications: the trust's or ICB's immediate notices, fastest and narrowest, visible only inside the organisation. The failure mode is layer confusion: treating an answer sourced from layers one to three as if it covered four and five, which it structurally may not, and the sharper version, treating "no warning found" as "no warning exists", the silence-as-clearance error that recency makes routine.
The four recency questions for any tool
Askable of any evidence system, ours included, and diagnostic in combination. What is the source date? Visible dating on the underlying document is the minimum; answers that cite undated material have removed your ability to run this check at all. When was the corpus last updated? A different question from the first, the gap between a tool's refresh cycle and the live feed is exactly where today's alert lives. Does the answer surface newer regulator communications where they exist? The behaviour that separates recency-aware architecture from recency-blind fluency, and testable: ask about a medicine class with recent safety history and watch whether the live layer appears. And does the system distinguish "no evidence found in my sources" from "no warning exists"? The epistemically honest formulation costs one clause and changes everything about how silence should be read. Tools built on the inspection principle this cluster runs on, source dates exposed, live communications linked when material, ours to be held to as much as anyone's, make the questions cheap; tools that cannot answer them have answered them.
Closing the gap: the subscription habit
The last mile is not a tool feature; it is a prescriber habit, and it costs ten minutes once. Subscribe to MHRA Drug Safety Update and to emc's change notifications for the medicine classes you prescribe most, so the live layer reaches you on its own schedule rather than waiting to be asked; route local urgent communications into somewhere you actually look; and for the high-risk territories, anticoagulants, immunosuppressants, medicines with pregnancy-prevention programmes, treat any consequential decision as triggering a live-layer check regardless of what the summary said, thirty seconds against the class of error that summaries cannot see. The division of labour then lands where this whole cluster keeps landing it: synthesis for orientation, sources for decisions, and the live regulatory feed as the layer you own personally, because recency is the one property no photograph, however well taken, can have.
Frequently asked questions
How often do material SmPC changes actually happen?
Continuously across the formulary, which is the point: any individual product changes rarely, the portfolio you prescribe changes constantly, and subscriptions convert that background rate from ambush into inbox.
Should I re-verify a summary I checked last month?
For a repeated routine decision, the dated-source glance suffices; for anything high-risk or any decision that surprised you, re-run the live-layer check, since last month's verification inherited last month's feed.
Do local formularies lag national safety actions?
They can, briefly and structurally, which is why the live regulator layer outranks every document during the lag, and why flagging a spotted lag to the formulary owner is part of the professional feedback loop.
Can AI tools subscribe to the live layer on my behalf?
Architectures increasingly ingest regulator feeds, and the four questions still apply: ingestion lag, coverage and surfacing behaviour differ by tool, so the personal subscription remains the belt to any tool's braces for your highest-risk classes.