AI News

Reels-RSI treats an explainer video as an executable program consisting of a script, storyboard, and per-frame animation code, rather than as a set of generated pixels. The open-source reel director can run with any agent harness and improves the director at the program level, without changing model weights or prompts. Each revision is logged with a code snapshot; retrospectives link check findings to fixing diffs and propose lessons. Lessons that can be checked statically become rules with must-fail and must-pass examples, making the memory unit-tested, while newly learned rules are also run against earlier reels. A bounded, benchmark-gated procedure may rewrite the director's playbook, but a rewrite is retained only when blind pairwise judgments on held-out topics prefer it, and human decisions are required before anything enters memory. Across 111 reels covering a field study, two playbook evolutions, a 2x2 memory ablation, a weaker director, a 48-reel learning curve, and a judge-reliability study, the authors report several qualifications. A rule learned in the field detected four latent defects in older frames and none in frames made afterward. Rules found no violations in eight unguarded reels from a strong director, but reduced violations for a weaker director from 30 to five. Lessons shortened revision loops without compounding over time or changing judge scores. The first rewrite's apparent 11-1 preference was attributed to vote-order bias: judges selected the second-shown reel 62-81% of the time, obscuring scripts that presented 13 facts in exchange for craft. The paper documents four preference-signal failure modes, the checks that detected them, and the corresponding fixes. Code, rules, lessons, and data are public.