Your dashboard is a forecast with the error bars removed. Every view of a running system is old data plus a guess about now; the systems that must be right about other machines already say how old and how much of a guess. Make your views say it too.

Nobody debugging a live distributed system sees it as it is. A Prometheus panel shows samples up to a scrape interval old, and its rate() extrapolates to the edge of the window without telling you. A P99 averaged across hosts is not a P99. A closed-loop load generator stops measuring exactly when the system stalls. A replica read, a log tail and a failure detector are all records of the past with some rule, stated or not, for what they mean now.

Some systems state the rule. Spanner's TrueTime returns an interval, not an instant. A phi-accrual failure detector reports suspicion that grows with silence instead of a binary up or down. Leases carry an expiry; bounded staleness is a product feature with a number on it. These are views that show their age and mark their projections.

This talk separates every live view into its two parts, a decayed record and a projection, and proposes five rules for any dashboard, each with a test that can fail: show age, render a decay model, project only by known rules, put truth beside belief when debugging, and send change-triggered updates tagged with the state they describe. One open-source dashboard is scored against all five. The idea came from watching an LLM agent act on a stale view of a running program, and it transfers to the views engineers act on during an incident.

Performance and distributed-systems engineers who run dashboards and are paged by them. Twenty minutes, recorded, live Q&A. The 2026 program had three tracing and tail-latency talks about what a pipeline keeps; this one is about the age of what it shows. Slide plan: the systems cut in _drafts/2026-10-18-every-view-is-late/includes/talk.org.

Not reviewed. Before submission the claim has to survive the call's "no shallow overviews" test, which means at least one measurement from the paper actually run (scrape aliasing, a cross-host P99 against a merged histogram, or open- against closed-loop load), with numbers on a slide.

Concept only. The draft paper (unpublished) has the survey and the tests; the agent test bench is in a private harness, cited by date. Related public notes: Forgetting and Attribution (decay as a declared policy) and Agent Telemetry Systems.