Put an expiry date on every conclusion

The most expensive thing in a strategy document is usually not a wrong claim. It is a claim that was right, was never dated, and is now being inherited by people who were not in the room when it was true. Nobody re-examines it, because it does not present as a claim any more. It presents as context.

Evidence decays at wildly different rates, and treating all of it as equally durable is the error. What a frontier model could not do a year ago is not a fact about models; it is a fact about last year. Integration conventions and tooling norms have been moving faster than the documentation describing them since 2025. Competitor positioning survives a quarter or two. But why a specific person behaves the way they do at work — the frustration, the workaround, the thing they will not admit in a survey — that decays over years, sometimes decades. This is the underrated reason that talking to ten users still outperforms a dashboard at the frontier: you are collecting the slowest-decaying signal available.

So assign an expiry at the moment of recording. A date, not a duration — durations never get computed, dates trigger. At expiry, you either re-test or retire. Re-testing is usually one call, one query, one search. Retiring is free. Inheriting is the expensive option, and it is what happens by default when no date was written.

The audit question for any review is short enough to ask out loud: when did we last actually observe this? A room that cannot answer is not making a decision from evidence. It is making one from memory, about a world that has since moved.

The honest caveat is that half-lives are estimates, not measurements. We publish ours as a calibration prior and mark them draft, because inventing a precise number for how fast a belief decays would be exactly the kind of false rigour this whole practice is supposed to prevent.