Evidence decays: why justified decisions can become wrong

A launch pack is a photograph. Production is a climate. Reliance has to notice when the photograph is no longer true.

8 September 2026 · 6 min read

The assessment you ran at launch is not false. It is dated. Teams that treat a launch pack as a permanent licence discover the gap after an incident, not before.

Evidence weakens for ordinary reasons: the model is updated, retrieval corpora drift, the user mix changes, a downstream system returns different fields, a human override path is quietly bypassed. None of these require malice. All of them can invalidate last quarter's conclusion.

What the 2026 Index adds

Source finding

Stanford HAI records 362 documented AI incidents in 2025, up from 233 in 2024, and describes a jagged capability frontier: systems that look extraordinary on one benchmark and brittle on another. Agent success on OSWorld improved sharply and still failed about one in three structured attempts.

Solarascope analysis

Decay is not a reason to freeze systems. It is a reason to give every conclusion a last-assessed date, a condition list and a trigger that can suspend Ready. Decision Coverage exists because approval is a moment and reliance is a period.

Sources

Stanford HAI. AI Index Report 2026: Technical performance (April 2026)
Stanford HAI. AI Index Report 2026 (incidents) (April 2026)