# Timeline | Date | Milestone | |---|---| | ~~Sep 1, 2026~~ (23:55 UTC+0) Sep. 8 2026 | **Development Phase** starts; training data released | | ~~Nov 1, 2026~~ (23:55 UTC+0) Nov. 8 2026 | **Testing Phase** starts; blind test data released | | ~~Dec 1, 2026~~ (23:55 UTC+0) Dec. 8 2026 | Submission deadline | | Dec 14, 2026 | Results announced | See CodaBench's timeline progress bar for the exact time of each phase in your local region. **Please carefully look at the TIME ZONE displayed in Codabench and make sure you are in the right zone. Otherwise, please do a manual conversion between different UTC zones. It is YOUR responsibility to submit the results before the test phase ends.** Please refer to the Codabench timeline display to accurate time and date in your time zone. ## Phase guide ### Development Phase (Sep 8 23:55 UTC+0 – Nov 8, 2026 23:55 UTC+0) Use this phase to download the training data, run the baselines, and tune your model. You can use any cross-validation protocols to test the effectiveness of your method. We **suggest** you to follow the competition's protocol and test your methods on **held-out datasets**. That is to say, you leave one dataset out as the validation set, and use the remaining datasets to train your own decoders. It is **optional** to loop this **held-out datasets** procedure for each single training dataset. But keep in mind that, this is **suggestions** only. You can use **ANY** cross-validation protocol as you wish. **We will start the submission channel to allow you to test the submission format when the Development Phase is close to end.** ### Testing Phase (Nov 9, 2026 00:00 UTC+0 – Dec 8, 23:55 UTC+0 2026) The blind test set (with novel subjects and stimuli, no labels) is released at the start of this phase. Run your final model on it, produce the predictions, and submit. You may submit **at most 2 times**; the **latest** submission, **NOT the best**, will be used for final ranking. Follow the competition rules to keep your method qualified. **MAKE SURE the submission label sequence is at 128 Hz**, no matter what sampling rate you use during inference. A brief summary of rules you should follow during test: - A **maximal look ahead of 1000 ms** is allowed. To predict label at time t, you are allowed to use data ranging from [0,t+1000ms] in the same trial. Make sure to take all lags into account (decoder's post-lag, post-processing's non-causality, and any others.) - **Do not finetune** your model with test samples with any method (e.g. self-supervised learning on test set). Your algorithm can only use test samples from the trial under test and taking the causality requirement (max 1000ms look-ahead) into account. - **Do not** try to utilize the dataset/subject/trial identity to boost your decoder with any method. Your decoder should take the **EEG** and **(optionally) envelope** as the only inputs. - **Submit** your label sequence for **EVERY** test trials, with a sampling rate of **128Hz**, regardless of the sampling rate you use for inference. ### Validation on the leaderboard results This phase will start once test phase is completed. And we will announce the validation results as soon as possible. We will manually check the submitted codes and reproduce the reported results for the top-5 ranked teams. We will check the following terms: - Is the look-ahead-restriction followed? A maximal look ahead of 1000 ms is allowed, counting every possible lags in the submitted method into account. - Does the solution utilize finetuning, unsupervised/self-supervised learning, subject/trial-ID positional encoding, or any other technique that violates the competition rules? - Is the final result reproducible? - Any other possible violation of the competition rules. The top-5 ranked teams will be disqualified if violation of competition rules is found. **Submitting your code does not transfer the copyright to any other parties, including the organizers (us).** In addition, we highly encourage , no matter what rank you get, to make your code public after the competition finished. **You may choose the type of license when you decide to make your code public**. Submissions with published code will be highlighted on the leaderboard as verifiable submissions (i.e. verifiable by the AAD researcher community).