Methodology

Source data

Pack One uses public 17Lands Premier Draft data to construct historical replay seats. The site stores a fixed sample of replay drafts for each supported set. Card images are loaded from Scryfall. Pack One is not affiliated with 17Lands, Scryfall, Wizards of the Coast, or TCGplayer.

Verified trophy puzzles

Draft Run covers the loaded expansion catalog, and Powered Cube has a separate trophy-only ten-decision flow. Every source records seven wins and at least 100 prior games. Modern sources meet their set’s elite win-rate cutoff, at least 60%. STX, MID, and VOW use their existing Diamond/Mythic cohort from the earliest game’s rank because those archives do not publish win-rate buckets.

The official archives are re-read to verify trophy outcomes, skill evidence, complete candidate lists, historical choices, and every earlier-card pool against the held-out model artifacts. Incomplete or inconsistent trajectories are excluded. The source archive hashes and per-set coverage are recorded with the corpus. Partial credit uses the existing broader elite model with each source draft’s fold held out.

The strong-player cohort

The model is trained on an experienced, high-win-rate slice of the available draft data rather than every drafter equally. The exact cutoff and training sample are recorded in each set manifest and surfaced on the set archive pages. This makes the target explicitly “what this stronger cohort tended to support,” not “what all players picked.”

Pool-conditioned support

A first-pick model can rely heavily on global card preference because the pool is empty. Later in the pack, however, context matters. Pack One's current model adjusts card support using shrinkage-weighted card and pool co-pick information, then normalizes support among the cards available at that decision. Draft Run uses the original drafter’s historical pool for each independent question. Full Pack reconditions support on the cards you chose while keeping available packs fixed; it does not simulate the other seats.

Evaluation and holdout

Training and evaluation are separated by draft identifier using a five-fold holdout. That reduces leakage from seeing other picks from the same draft during training. Pack One also evaluates whether the model separates historical strong-player choices from deliberately poor or random baselines. Those checks are useful evidence that the score has discrimination, but they do not turn the model into a universal definition of optimal drafting.

What the probabilities are not

The displayed percentages are not calibrated probabilities that a card is correct. They are normalized model support within the current pack. A 40% card and a 35% card represent a much closer decision than a 70% card and a 10% card, but neither percentage should be interpreted as expected win rate or chance to trophy.

Model humility is part of the product. Pack One should be most interesting where good drafters can inspect a disagreement, understand the model's confidence, and decide whether the model saw something they missed—or whether they still prefer their own line.

Versioning

Every set manifest includes the model version, source date, cohort definition, training counts, and replay count. When the scoring or model changes materially, Pack One treats that as a versioned product change rather than silently rewriting what old scores meant.