Lifetime competition tiers never expire, but a medal's signal of current skill mostly fades within about a year.
Why do lifetime tiers discard information that recency-weighted medals preserve?
This explores why a permanent, lifetime credential (like a Kaggle tier such as Expert or Grandmaster, which never expires once earned) loses information about a person's current skill that a medal record weighted toward recent results keeps.
This explores why a lifetime credential, a tier you earn once and keep forever, tells you less about someone's current ability than a record of medals that gives more weight to recent results. The short answer from the corpus: a medal's predictive value is mostly about time, and a lifetime tier is built to ignore time. One caveat first. The collection has no note that studies tiers directly. What it does have is evidence on how medals age, and that evidence explains most of the gap.
The central finding is that medals predict how someone will do on a hidden test set almost entirely within the first year after they're earned. That held both before and after generative AI arrived Do Kaggle medals still predict performance after AI arrived?. In other words, a medal is a dated signal: "this person could do this, around now." A recency-weighted view keeps that date and lets old medals fade the way their predictive power actually fades. A lifetime tier adds medals up into one status and drops the timestamps. A Grandmaster title earned in 2017 and one earned last month look the same, even though the data says only one of them still predicts much.
The surprising part is how much harm a frozen credential can do when the world changes under it. An audit of one competition format found that about half the collapse in how informative its medals were came from what it calls institutional stranding. The platform retired the format before AI arrived, the medals kept aging, but they stayed on display at their original value How much did retiring a competition format hurt medal credibility?. That is the lifetime-tier problem in a small case: the credential got cut off from the process that once validated it. Nothing the person did went wrong. The badge simply stopped tracking anything current while still looking authoritative.
There's a broader pattern here that reaches well beyond Kaggle. Research on AI agent feedback finds that collapsing rich feedback into a single score keeps the "how good was it" part and throws away the "which direction, which specifics" part Can scalar rewards capture all the information in agent feedback?. A lifetime tier compresses in the same way. It keeps a cumulative "how much has this person achieved" and drops when they achieved it, in which formats, and whether those formats still exist. Any summary that turns a stream of evidence into one permanent label pays this cost, whether it's an RL reward or a résumé line.
The practical takeaway: when you read a credential, ask what it is a function of. If it never decays and never notes its context, it is answering a historical question ("did this person ever do this?"), not a predictive one ("can they do it now?"). To go deeper, the Kaggle note has the aging evidence and the stranding note shows how a credential can mislead without anyone cheating.