A fixed camera. A budget of K remembered frames. Three objects that may have changed — a frame ago, a minute ago, or ages ago. The task is knowing when. Slide K down and watch every schedule except log lose a timescale. ✅ = with these memory ages, the change time is pinned to one bucket; ❌ = no possible model could know — the schedule, not the model, is the bottleneck.
Trained result on this benchmark (128K-param model, 3 seeds, K=8): log 0.90 vs best hand-built baseline 0.76, sliding window 0.48 — and the schedule retunes on frozen weights. Code, paper draft & the tri-axis (zoom) results: github.com/resoajoe/loglens · benchmark generator: resoa/triclock