New preprint: "When the harness decides success: state that cannot be re-observed"
Agent harnesses drop old turns, summarize them, ask the model to keep notes, or replay its earlier reasoning, usually to save cost. When does that choice decide whether the agent succeeds? A pre-registered probe, plus a component-by-component decomposition of ARC-AGI-3's harness gap.
Probe: truncating the context cost a low-cost reasoning model nothing when a needed fact could be looked up again. When the fact could not be re-observed, success fell to chance. The penalties differ by 92.5 points (90% CI 89.5–95.5), and a replication on a second provider's model agreed.
Truncation also cost 2.9–4.0× as much in dollars as full history while sending fewer tokens, because dropping old turns breaks prompt caching.
ARC-AGI-3, one model, four games: summarizing old turns instead of dropping them, while keeping the Standard harness's note-taking instruction, recovered 86% of a 63-point gap without reasoning replay. One run per configuration and game, so it's a lead, not a measurement.
Not shown: real tasks, other models, or any ranking of models.
Paper: https://arxiv.org/abs/2610.26632
The pre-registration, amendments, code and cost ledger are open: https://github.com/raimondasl/persistgym
