Answer real ops questions from kitchen CCTV.

Submit one reproducible program. It receives CCTV-style kitchen videos and a JSON file of questions, then writes answers and a run log. No manual video inspection during scoring.

python answer.py --videos ./videos --questions questions.json --out answers.json --log run_log.json

The useful skill is not captioning every frame. It is finding the few moments that answer the question, using OCR only where it helps, and saying not_visible when the footage does not show enough.

Start with the long fixed-view kitchen clips below. Scoring uses hidden questions and an answer key on the stated video set; if any held-back clips are added later, that will be stated before they are used.

These are examples so builders understand the job. Final scoring uses hidden questions on the same public videos, with an answer key.

Open-source or local models are recommended, because this should be cheap enough for a small kitchen to run often. They are not mandatory.

Cloud models are allowed if every call is logged and the normalized cost stays under the cap. Local runs count as $0 model/API cost, but still have to fit the same runtime and frame limits.

The baseline is a simple coarse-to-fine video agent: sample the whole clip cheaply, identify likely time windows, inspect only those windows with a vision model and OCR where useful, then answer in JSON with evidence.

Total score is 100 points. A run is valid only if it stays under the hard cost and runtime caps. If it goes over, it does not rank.

• 80 points - answer accuracy: hidden yes/no, counts, event order, timestamp, duration, and not-visible questions.
• 15 points - cost efficiency: lower model/API cost helps only after the answer is valid and useful.
• 5 points - reproducibility: one command, useful evidence, frame/model call log, and repeatable outputs.

Hard cap: $0.30 estimated model/API cost per 60 minutes of source video, 25 minutes wall-clock, and about 1,500 sampled frames per 60 minutes unless the evaluator states an equivalent frame budget.

If the scored set is longer or shorter, the dollar and frame caps scale with video minutes. The cap is deliberate: a useful restaurant tool has to find the right windows cheaply, not send every frame to the biggest model.

A restaurant or cloud-kitchen operator does not need a general video captioner. They need cheap, repeatable checks from footage they already collect: whether cooks wore caps or hairnets, whether food sat too long at handoff, whether a container was sealed, where orders backed up, and what the camera could not show.

The cost cap matters. A system that spends heavily on every frame is less useful to a small kitchen than one that samples intelligently, answers the actual operations question, and admits when the video does not show enough.

Public video-QA benchmarks usually separate examples from the held-out scoring answers. This round follows that idea, but makes it practical for kitchens: fixed-camera footage, objective questions, a cost cap, and logs that show what the agent actually looked at.

Sample questions are public so builders understand the job. Hidden questions prevent agents from hard-coding answers to the visible examples.