Make the results comparable.

Before kickoff

Advance preparation is limited to downloading, copying and unpacking files, verifying checksums, reading technical format/license instructions, and checking accounts, packages and compute with unrelated synthetic examples. Do not explore or analyze the challenge observations, create challenge features or visualizations, train or tune models, or have AI agents do that work before kickoff. Challenge modeling and work start together at the announced kickoff. Declare compliance in your method report.

Permitted data and tools

• Use the released German bird observations, supplied AlphaEarth inputs and organizer-provided helper data.
• AI coding tools, general programming documentation, standard ML libraries and the allocated compute are allowed. Tell AI agents these data rules explicitly.
• Do not retrieve additional bird observations, external bird-occurrence labels, species range maps or pretrained bird-distribution predictions. Do not infer evaluation locations and look up their labels.
• Additional environmental sources are allowed only if organizers release equivalent approved training and test inputs to every team before the event.
• Fit learned transforms on permitted training data. Do not refit preprocessing to the anonymous test batch or manually edit predictions from test-label information.

Development and optional CV

Report the required German development AP using the supplied labels and exclusions. Fit scaling, imputation, dimensionality reduction, feature selection and supervised auxiliary stages inside each training partition. The same spatial and source exclusions apply across species.

Optional five-fold spatial CV is useful for comparing one or two promising models; it is not required and earns no points merely for being run. For cross-species claims, present paired results with and without other species on identical focal labels, backgrounds, splits and feature supports.

A final refit on German training plus development is allowed after model selection. Declare it and retain the earlier held-out development score. Additional released German labels must follow the documented season, quality, site and buffer rules.

Final predictions and metric

Submit exactly 63,590 rows using row_id,score. Predict all ten species at all 6,359 requested locations. Keys must exactly match the decrypted sample submission; order does not matter. Scores must be finite numbers in [0,1]. A score is a ranking output, not a calibrated occupancy probability.

For each concealed partition, pool all scored countries within each species, calculate average precision, then average all ten species equally. Higher is better. Background ratio and eligibility are frozen. Unknown labels are excluded; they are not converted into negatives.

The provisional board displays the starred final submission’s aggregate public macro AP; changing the star updates its score and rank. AP comparisons use full precision; scores are displayed to six decimals. An exact AP tie is broken by the earlier relevant accepted prediction timestamp, then by receipt number for a same-second tie. On the final board this is the nominated prediction’s original receipt time. The final board uses the geographically separate private split. Neither private scores nor group details are released during model selection. Geographic identifiers and public/private membership are absent from test inputs.

Attempts, receipt times and final nomination

• One team lead creates the team and invites up to two teammates. Every member signs in with their own GitHub account and can submit after organizer approval. Each account belongs to one team; submission quotas are shared by the team.
• Five accepted final submissions per team for the entire event. Invalid files and repeated identical predictions do not consume an accepted attempt. A separate limit of 20 invalid requests per team/channel protects the queue.
• Practice has a separate allowance of ten accepted format checks and no leaderboard score.
• Final submissions open at 13:30 CEST and close strictly before 17:30 CEST on 24 October. Upload from your team workspace and wait for a numbered receipt. Uploading a file alone is not a submission.
• The server records the receipt time after upload verification. A timely valid submission remains eligible if evaluation finishes later. Receipts cannot be edited.
• The latest accepted final prediction becomes your starred choice automatically, even after an earlier manual choice. To use another accepted prediction, click its star in your team workspace and confirm before the deadline. The latest choice or newer accepted submission wins, ordered by server receipt time. At the deadline choices are locked; pending timely submissions still finish evaluation.
• Pending receipts are processed before selections are sealed. Private scores are released only after the organizer seals selections and publishes results.

Allow time for uploads and evaluation. Contact an organizer if a receipt remains pending. Evaluation outages preserve the original receipt and do not consume an accepted attempt.

Judging and presentation

Model points: improve on a fixed reference

Matching the shipped reference earns 15/50 model points. An improvement of +0.15 AP earns 50/50. Points are capped at 0 and 50.

Let s be your chosen prediction’s final private AP, a = 1/6 the constant-score AP and b the frozen reference model’s AP on that same private split. For s ≤ b, points are max(0, 15 × (s − a) / (b − a)). For s > b, points are min(50, 15 + 35 × (s − b) / 0.15). At or below constant-score AP you earn zero model points.

Examples are rounded to one decimal; calculations and ties use full precision. The reference is the supplied default 289-feature linear starter, fitted only on German training rows with seed 20261016, Python 3.14.4 and the pinned requirements. Reference predictions are fixed and checked by the private evaluator. The reference’s German development AP is 0.38586286412186144; that is a separate comparison, not the hidden private reference AP b. The private reference AP and final model points are revealed with final results, after model selection closes. No team’s performance changes the reference or stretch.

The five reproducibility checks

A technical organizer checks five pieces of evidence, one point each: code/configuration; reproduction instructions and dependencies; the correct common German validation result; features, preprocessing and helpers; and accurate limitations. Complete the method report and provide the matching code/checkpoint and prediction hash. A checked box without evidence earns no point. Optional CV and complex models are not requirements.

Overall-prize eligibility requires a valid prediction, supported insights with visualizations, reproducibility references and a presentation. Those insights and visualizations can come from Task 1: your model experiments, results and limitations. Tasks 2 and 3 are optional extensions for teams that have completed the main challenge and finish early; neither is required for prize eligibility. Leakage and prohibited data remain validity issues, regardless of checklist points.

An exact overall jury-total tie uses the chosen prediction’s original receipt and then receipt number, including when model points are capped. The AP leaderboard remains separate from the jury’s overall result. Cross-species work uses the same scoring categories, with no extra bonus or separate award.

Submit or share the method report, code/checkpoint reference and presentation by the working deadline. Report features, seeds, runtime, AI/API use, final-refit decisions, validation results, limitations and the submitted CSV SHA256. Existing data licenses remain applicable; participation does not transfer ownership of a team’s original code or models.

Upload and final choice: a worked example

Select your CSV in the team workspace and click Encrypt & submit. The app validates and encrypts it locally, uploads it privately, and issues a numbered receipt. Wait for an accepted result.

If #12 and then #16 are accepted, #16 is the default final choice, even if #12 had a higher public score. To use #12, click the star beside #12 and confirm before 17:30. A newer accepted final submission replaces that choice; you can star #12 again before the deadline. An invalid later upload does not replace an accepted prediction. The jury never picks a team’s best hidden result.

Presentation upload

Use the Google presentation upload form linked from the team workspace before 17:30 CEST. Google sign-in is required, and response files are private to the organizers. Upload a PDF of at most 10 MB and an optional editable source of at most 100 MB. Retain the confirmation page. Organizers use the latest complete valid response received before closure. Presentations last five minutes plus two minutes of questions.

Scientific interpretation

Occurrence data reflect observer effort, detection and reporting. Coordinate uncertainty may exceed the embedding pixel size. Annual 2024 features support retrospective seasonal reported-presence ranking, not a forecasting or causal claim. AlphaEarth pretraining includes upstream biodiversity supervision. Co-occurrence patterns alone do not demonstrate species interactions.

Technical incidents

Organizers preserve original receipt times, file hashes, quotas and nominations. A service outage does not authorize using additional data or accessing hidden labels. Any timing adjustment will be announced equally to all teams before private results are revealed. Incorrect submissions can be rejected; deliberate rule circumvention can disqualify an entry after review.