DraftFM: A Foundation Model for Day-Zero Drafting in Magic: The Gathering
DraftFM is a discrete-choice policy that enables "day-zero" drafting predictions for unseen Magic: The Gathering expansions before any draft logs exist Cards are encoded as frozen 775-dimensional vectors from public card records, structured features, and fixed text embeddings—no card identities, set identities, or usage statistics are used A 1.6M-parameter network trained on 149M human picks from 29 expansions achieved 50.8%, 60.4%, and 56.7% top-1 agreement on three entirely withheld expansions
Analysis
TL;DR
- DraftFM is a discrete-choice policy that enables "day-zero" drafting predictions for unseen Magic: The Gathering expansions before any draft logs exist
- Cards are encoded as frozen 775-dimensional vectors from public card records, structured features, and fixed text embeddings—no card identities, set identities, or usage statistics are used
- A 1.6M-parameter network trained on 149M human picks from 29 expansions achieved 50.8%, 60.4%, and 56.7% top-1 agreement on three entirely withheld expansions (vs. ~7% uniform chance)
- The same architecture refitted on all 32 observed expansions produced a sealed card ranking for the unreleased set The Hobbit, published ~36 hours before it became draftable
- The sealed ranking agreed with six independent expert reviewers roughly as much as those reviewers agreed with each other
Why It Matters
This work demonstrates a practical application of transfer learning and zero-shot generalization in a complex, high-dimensional discrete-choice domain where training data for new environments does not yet exist. For AI practitioners, it shows how carefully designed feature representations can enable models to generalize to entirely unseen categories without retraining, which has broad implications for recommendation systems, market entry prediction, and any domain where new items appear before behavioral data is available.
Technical Details
- Architecture: A compact 1.6-million-parameter neural network functioning as a discrete-choice policy that scores exactly the cards in the current pack, conditioned on the drafted pool and draft state
- Feature Engineering: Every card is represented as a frozen 775-dimensional vector derived from public card records, structured features, and a fixed text embedding—ensuring no leakage of card identities, set identities, or usage statistics into the model
- Training Data: 149 million human picks aggregated from 29 Magic: The Gathering expansions
- Evaluation: Three expansions were withheld entirely during training; top-1 agreement rates of 50.8%, 60.4%, and 56.7% were achieved against human picks, compared to ~7% for uniform random selection
- Real-World Deployment: The model was refitted on all 32 observed expansions to produce a card ranking for The Hobbit set before its release, with cryptographic sealing of the prediction and validation against six independent expert reviewers
Industry Insight
- The "frozen feature + lightweight head" paradigm demonstrated here is directly transferable to recommendation systems launching new products, where historical interaction data is absent but rich metadata is available
- The cryptographic sealing and pre-release validation approach sets a strong precedent for trustworthy AI deployment in time-sensitive domains, showing how model predictions can be made tamper-evident and auditable
- The near-expert-level agreement suggests that well-regularized, metadata-driven models can approximate human expertise in complex strategic domains, opening possibilities for AI-assisted decision support in competitive gaming, forecasting, and resource allocation
Disclaimer: The above content is generated by AI and is for reference only.