Teams begin with access to 30% of the dataset and 40 tokens. Additional portions of the dataset can be unlocked progressively: 30% → 45% → 70% → 100%. Duration: 75-90 mins

• Each team starts with 30% dataset access and 40 tokens.
• Teams may submit models for private score checks at a cost of 1.5 tokens per attempt.
• Private score checks do not appear on the public leaderboard.
• Dataset access can be unlocked progressively if model performance is X% more than reference model threshold for the available data.
• Pushing score to public leaderboard costs 3 tokens for the first push, 6 tokens for the second, and 7 tokens for every subsequent push.
• A team must push at least once to view the live public leaderboard.
• Teams that do not manually push by the end of the round will have their best private score automatically posted, but they don’t get advantage during Round 2.
• A sidequest is activated every 20 minutes that earns the team 5 tokens if completed successfully.
• Tokens earned through sidequests and bonuses can be used throughout the event.
• The round ends when the Round 1 time limit expires. The Round1 leaderboard is frozen.

A manually pushed score takes precedence over an auto-pushed score of the same value. Remaining token balance may be used as the secondary ranking factor.

Only the top-ranked teams from Round 1 qualify for Round 2. Each qualifying team is assigned another team to sabotage. The method of sabotage can be chosen by the qualifying teams in order of their position on the public leaderboard. Teams must adapt their models to the resulting sabotaged dataset and compete on a fresh Round 2 leaderboard. Duration: 75-90 mins

• Only teams qualifying from Round 1 may participate.
• Qualifying teams are ranked from 1 to N.
• Each team is assigned another team to sabotage through a randomly selected rotation offset, i.e., Team i sabotages Team (i+k) mod n.
• Every team receives exactly one sabotage target.
• Teams that manually pushed at least once in Round 1 receive priority when choosing their sabotage type.
• Sabotage types are selected from a shared pool and are removed as teams choose them.
• Teams that did not manually push in Round 1 are assigned a remaining sabotage type automatically.
• The sabotage may be of any form including deleted columns, swapped categorical columns, hidden transformations, row-level feature swaps, or injected code/text modifications.
• The % dataset collected during Round 1 by each team is the base on which the sabotage is performed in Round 2.
• The sabotaged dataset becomes the official dataset for Round 2.
• Teams cannot purchase additional data during Round 2.
• Private score checks cost 1.5 tokens.
• Leaderboard pushes cost 3 tokens for the first push, 6 tokens for the second, and 7 tokens for every subsequent push.
• Correctly identifying the sabotage done on their dataset awards the team 5 tokens once.
• Round 2 uses a separate leaderboard from Round 1.

Manually pushed scores take precedence over auto-pushed scores of the same value. Remaining token balance is used as a secondary ranking factor.

The highest-ranked teams from Round 2 advance to the final round. This is a fully physical explanation and defence round, where teams present their modelling approach, explain the decisions they made throughout the competition, discuss their trade-offs, and defend how they adapted to the adversarial changes introduced during Round 2. Duration: 45-60 mins

• Only teams qualifying from Round 2 may participate.
• The round is conducted entirely in person.
• Teams must explain their modelling approach and key decisions.
• Teams must be able to justify their modelling choices and trade-offs.
• Teams must explain how they adapted to the Round 2 sabotage.
• No tokens are used during this round.
• No further model submissions are made during this round.
• Final evaluation is based on the judges' assessment.

Judges’ Decision is final.