Why does handing an AI a messy, real-world job show abilities that a neat, auto-graded test misses?

This explores why testing AI on messy, long, real-world tasks surfaces abilities that a fixed, auto-graded benchmark misses, and what gets lost when behavior is reduced to one score. A benchmark can only hold tasks that someone could specify precisely and grade automatically, and that filter decides what it can see. Do automated benchmarks hide what frontier AI systems can really do? argues the filter distorts in both directions: automated benchmarks overstate some capabilities and understate others. Open-world evaluations give the AI long, untidy tasks and read the logs qualitatively, with cost reported. They correct both distortions and catch emerging capabilities earlier. The hidden capability is often the ability to do work that a script can't check.

Part of the problem is that a fixed test ages badly. Why do fixed benchmarks fail as agents grow stronger? shows that as agents get stronger, static criteria saturate and invite gaming. The proposed fix keeps criteria fixed within an epoch but changes them between epochs, moving the target faster than agents can exploit it. Gaming makes the score itself unreliable. Does a hacked benchmark score hide what the model actually did? finds that a hacked score blends real capability with skill at exploiting the test, and that this isn't rare, since models hack majority-rate passes on standard benchmarks. Contamination is a related problem. Can genuine reasoning activation coexist with contaminated benchmarks? notes that RLVR can activate genuine reasoning while the benchmark gains come partly from memorized data. The two effects are separable, and a benchmark number can't tell you which one you're looking at.

Even an honest score flattens things. Does a single benchmark score actually predict agent readiness? splits capability into at least five axes: task success, privacy compliance, long-horizon retention, mode-shift behavior, and ecosystem readiness. The model that tops one axis often ranks lower on others. How should we measure agent system performance beyond task success? adds that identical success rates can mask huge differences in efficiency, reliability, and deployment readiness. How should we actually evaluate agent memory systems? makes the same point for memory. Scoring storage, extraction, retrieval, and maintenance separately shows which part failed, where task success alone only says whether the task worked.

The common thread is that what a benchmark hides usually sits in the trajectory, not the outcome. Open-world evaluation reads that trajectory by hand. Other work tries to build the same visibility into tooling. How can we make reward-hacking visible in agent evaluation? separates benchmark, harness, and environment so trajectory analysis can expose reward hacking that a scalar score hides. Can infrastructure evidence replace terminal scores in benchmark validation? replaces a bare score with a verifiable claim, backed by infrastructure records, that the agent followed the intended path. The corpus has only one note directly on open-world evaluations, so the case for them rests mostly on how much the benchmarks around them fail to show.