sundae_bar

Review the active Challenge brief and download the public suite bundle (suite.yaml, dataset.jsonl, rubric.txt)

Access sundae_bar Lab challenge page

Develop skill instructions and procedures that teach an AI agent how to solve the Challenge. Submit through sundae_bar Lab while the Challenge is open

Create SKILL.md file

Submit SKILL.md for evaluation; subject to daily limits (rolling 24-hour window) and per-Challenge limits. Monitor remaining submissions shown on Lab

Submit skill and iterate

Download grader rationale and inspect scored results JSON to understand how score was computed across multiple scenarios and graders

Review evaluation results

Obtain suite.yaml, dataset.jsonl, and rubric.txt from the Lab challenge page

Download Challenge suite bundle

Configure the fixed evaluation harness with load_skill tool to inject miner submissions. Use models specified in suite.yaml for that Challenge

Set up evaluation harness

Set up all configured graders (skill_use, scenario_quality, rubric model-judge, novelty_check) with published weights from suite.yaml

Implement graders

Process queued submissions against all dataset scenarios, apply pre-gates (skill_alignment, dataset_derived), compute weighted scores per scenario, and average across scenarios

Run evaluation queue

Submit setWeights transaction to SN121 based on leaderboard ranking (highest score, then earliest qualifying submission). Current #1 miner receives 40% of emissions

Post weights on-chain

Use the same weights from sundae_bar's extensive scoring rather than re-running full evaluation. Does not change incentive mechanism; amplifies the published vote(optional)

Optional: Child-key to sundae_bar validator