edhpowerlevel can be predicted from the decklist alone, with no price or popularity inputs. On 3,039 decks the model never saw, it explains 89% of the variation (R² 0.894) and typically misses by about half a point on the 1 to 10 scale (mean absolute error 0.48). Bracket and win rate could not be predicted better than simple baselines.
Cutting a strong card moves the estimate the right way 97% of the time. Adding one does not yet clear the 90% bar. The best attempt reached 0.833 (55 of 66 swaps). You accepted the release on 2026-11-08 with that gap open.
The decklist predicts edhpowerlevel at R² 0.89
On the holdout, 3,039 decks no model saw during building, the locked model explains 89% of the spread in scores (R² 0.894).
| Model | Holdout R² | MAE (mean absolute error) |
|---|---|---|
| Locked model | 0.894 | 0.48 |
| Served card model | 0.857 | 0.42 |
| Rule baseline | 0.485 | 1.17 |
| Commander only | 0.219 | 1.43 |
| Guess the average | -0.001 | 1.67 |
The served card model adds a learned weight per card, so a single swap moves the estimate. It was scored after the holdout had been opened (post-hoc), so treat its 0.857 as a check on the locked result.
Within the training decks, the best tournament model scored 0.917 out-of-fold (each deck predicted by a model that never saw it). The best model using only readable card features scored 0.726.
Bracket and win rate were not predicted better than simple baselines. The bracket model gets the exact bracket 50% of the time, the same as the rule baseline (0.498 against 0.497). The win model does not beat the base win rate (log loss 0.698 against 0.696).
edhpowerlevel’s own site formula rates card price and EDHREC rank, so the target may carry both, and the R² measures agreement with that formula.
Deck text and similar decks account for most of the R²
| Family | R² lost if removed | R² lost if shuffled |
|---|---|---|
| Deck text and similar decks | +0.0925 | 0.688 |
| Mana curve | +0.0020 | 0.112 |
| Card power vs cost | +0.0016 | not run |
| Forge rules text | +0.0014 | not run |
| Interaction | +0.0009 | 0.027 |
| Bracket rules | +0.0006 | 0.007 |
| Card advantage | +0.0006 | 0.011 |
| Ramp | +0.0004 | 0.109 |
| Tutors and combos | +0.0003 | 0.028 |
| Synergy and structure | +0.0001 | 0.001 |
| Speed | 0.0000 | 0.007 |
| Card quality and LLM judges | -0.0001 | 0.003 |
| Resilience | -0.0004 | 0.016 |
Many readable features predict well alone (fast mana count 0.409, Game Changer count 0.425). The embedding overlaps them, so removing one family changes little.
As features, the LLM judges were weak: Sonnet’s power score explains 0.054 of the variance alone and Opus’s 0.063. Refitting without any judge columns gives holdout R² 0.8936, against 0.8944 with them.
Adds miss the 0.90 sign-rate bar that cuts pass
The model lowers its estimate on 97% of strong-card cuts (the sign rate), such as Sol Ring out for a Forest, against a pass bar of 0.90. Adds, such as Sensei’s Divining Top in for Swiftfoot Boots, rise less often.
Held-out-4 has 66 cases: 22 weak-to-strong swaps that keep the main card type (artifact for artifact, instant for instant), each applied to 3 real decks. The estimate rose in 55 of them.
Four attempts tried to fix adds, each scored on held-out swaps. Attempt 4 added card terms: each card’s power against its mana cost, scored from Forge rules text and printed cost only.
| Attempt | Held-out sign rate | Outcome |
|---|---|---|
| 1. Card prior | 0.652 | Failed the bar |
| 2. Role match | 1.000 | You rejected it: skewed by price, usage and card age |
| 3. Oracle power | 0.826 | Failed the bar and the litmus |
| 4. Rules power (card terms) | 0.833 | Failed the bar, passed the litmus |
The litmus: a card score must not track play count or release year (Spearman under 0.20 for both).
Cross-role swaps go wrong twice as often as same-role
You asked whether swaps across roles should rise at all. Swiftfoot Boots for Sensei’s Divining Top changes what the deck can do, so a lower estimate can be right. Ten of the 22 swaps change role (30 cases) and twelve keep it (36 cases).
| Swap type | Wrong sign | Share |
|---|---|---|
| Cross-role | 7 of 30 | 23% |
| Same-role | 4 of 36 | 11% |
For Boots to Top, the model charges for the lost protection: effects that grant abilities -0.027 and haste or evasion -0.022. The card terms recover the rise on 2 of 3 decks. You called this good news: “it implies that our deck architecture features are doing their job.”
Two swaps account for 6 of the 11 misses, wrong on every deck: Hero’s Downfall for Demonic Consultation, and Mentor of the Meek for Grand Abolisher. The text embedding (-0.057 to -0.047 and -0.047 to -0.039) and the per-card weights (-0.045 and -0.024) push them down. The readable features mostly point the right way, for example tutor count +0.029 on Consultation.
| Swap | Counted as | Wrong |
|---|---|---|
| Swiftfoot Boots to Sensei's Divining Top | Cross: protection to selection | 1 of 3 |
| Arcane Signet to Lion's Eye Diamond | Same: mana artifacts | 1 of 3 |
| Hero's Downfall to Demonic Consultation | Cross: removal to tutor | 3 of 3 |
| Explosive Vegetation to Finale of Devastation | Same: green sorceries | 1 of 3 |
| Feed the Swarm to Toxic Deluge | Same: black removal | 2 of 3 |
| Mentor of the Meek to Grand Abolisher | Cross: card draw to protection | 3 of 3 |
Readable features plus card terms reached 0.924 in an approximate split of each estimate, but that combination was picked after seeing these 66 cases.
The next round should test readable features with card terms
The run finished on 2026-11-08 after 73.5 hours, with all 13 subprojects landed. You accepted release R1 with its open problems, including the 0.833 add rate and the readable model trailing by 0.19 R².
Leads for the next round:
- Readable deck-architecture features plus the card power-cost terms, as the delta model for swaps.
- A fresh, frozen set of same-role swaps to test that combination honestly.
- Re-pin the Deckwright card store certification the Forge and power-cost features read from.
- Check the per-card weights against your litmus. They are learned from labels, may carry usage skew, and have not been measured.
The best readable feature alone scores R² 0.425
| # | Feature: what it measures (family) | Out-of-fold R² alone (average guess -0.002) |
|---|---|---|
| 1 | Deck text embedding: Average meaning of all card text, as numbers (deck text and similar decks) | 0.844 |
| 2 | Embedding strength probe: Strength read off the text embedding, fitted to labels (deck text and similar decks) | 0.765 |
| 3 | Bag of effect atoms: Count of every rules effect in the deck (deck text and similar decks) | 0.727 |
| 4 | Nearest-neighbour label mean: Average score of the most similar training decks (deck text and similar decks) | 0.726 |
| 5 | Game Changer count: Cards on the official Game Changers list (bracket rules) | 0.425 |
| 6 | Fast mana count: Ramp at 0-1 mana that nets mana at once (ramp) | 0.409 |
| 7 | Life costs: Costs paid in life (forge rules text) | 0.404 |
| 8 | Term sums: Raw power and cost terms per card, summed (card power vs cost) | 0.400 |
| 9 | Rule bracket floor: Lowest bracket the list qualifies for (bracket rules) | 0.393 |
| 10 | Ramp to curve ratio: Ramp count over average spell cost (ramp) | 0.369 |
| 11 | Tutors: Library searches for nonland cards (forge rules text) | 0.351 |
| 12 | Mana value bucket shares: Share of spells at each mana value (mana curve) | 0.310 |
| 13 | Mana per source: Mana made per source (forge rules text) | 0.307 |
| 14 | Efficiency by curve slot: Card power per cost at each mana value (card power vs cost) | 0.305 |
| 15 | Mana screw rate: Goldfish games missing a land drop by turn 4 (resilience) | 0.285 |
| 16 | Free interaction: Answers castable for free (interaction) | 0.281 |
| 17 | Role ramp efficiency: Power per cost of the ramp cards (card power vs cost) | 0.277 |
| 18 | Combo pieces reachable by tutors: Combo pieces the tutors can find (tutors and combos) | 0.267 |
| 19 | Cheap tutors: Tutors at 3 mana or less (tutors and combos) | 0.267 |
| 20 | Top-k efficiency: Power per cost of the best 5, 10, 20 cards (card power vs cost) | 0.262 |
| 21 | Best combo speed: Cheapest two-card combo in the deck (tutors and combos) | 0.260 |
| 22 | Land count: Lands in the deck (mana curve) | 0.254 |
| 23 | Ramp count: Cards tagged as ramp (ramp) | 0.237 |
| 24 | Commander identity: Which commander it is (deck text and similar decks) | 0.235 |
| 25 | Mana available on turn 4: Mana available on turns 3-4 (goldfish) (speed) | 0.232 |
| 26 | Tutor count: Cards that search for a nonland card (tutors and combos) | 0.226 |
| 27 | Efficiency sum: Total power per cost over all cards (card power vs cost) | 0.223 |
| 28 | Nearest-neighbour similarity: How close the nearest training decks are (deck text and similar decks) | 0.222 |
| 29 | Mana value spread: Spread of spell costs (mana curve) | 0.222 |
| 30 | Cheap ramp share: Share of ramp at 2 mana or less (ramp) | 0.220 |
| 31 | Mean nonland mana value: Average spell cost (mana curve) | 0.219 |
| 32 | Mana sources: Mana-making abilities (forge rules text) | 0.214 |
| 33 | Plan part odds: Odds each plan part shows up on time (resilience) | 0.205 |
| 34 | Stack interaction: Counterspells and stack answers (interaction) | 0.201 |
| 35 | Instant-speed interaction share: Share of answers at instant speed (interaction) | 0.189 |
| 36 | Combo assembled by turn 6: Odds a combo is assembled by turn 6 (tutors and combos) | 0.174 |
| 37 | Counterspells: Counterspells (forge rules text) | 0.170 |
| 38 | Mana colours: Colours a card can make (forge rules text) | 0.158 |
| 39 | Tipping-point mana value: Cost where most card value is reached (mana curve) | 0.149 |
| 40 | Cheap interaction: Answers at 2 mana or less (interaction) | 0.148 |
| 41 | Sacrifice costs: Costs paid by sacrificing (forge rules text) | 0.140 |
| 42 | Unrestricted tutors: Tutors that find any card (tutors and combos) | 0.135 |
| 43 | Alternative costs: Alternative ways to cast (forge rules text) | 0.134 |
| 44 | Cards seen by turn 6: Cards seen by turn 6 (goldfish) (card advantage) | 0.134 |
| 45 | Deckwright card value sum: Deckwright text-based card value, summed (card quality and LLM judges) | 0.132 |
| 46 | Two-card combo count: Two-card combos fully in the deck (tutors and combos) | 0.122 |
| 47 | Coloured source shortfall: Coloured sources short of the Karsten table (mana curve) | 0.114 |
| 48 | Type band deviation: Distance from the archetype's card-type mix (synergy and structure) | 0.102 |
| 49 | Efficiency mean: Average power per cost (card power vs cost) | 0.095 |
| 50 | Role draw efficiency: Power per cost of the draw cards (card power vs cost) | 0.089 |
| 51 | Graveyard hate and stax: Graveyard hate and static denial (interaction) | 0.084 |
| 52 | Plan redundancy: Copies per plan part against Deckwright floors (resilience) | 0.081 |
| 53 | Role interaction efficiency: Power per cost of the answers (card power vs cost) | 0.076 |
| 54 | Deckwright speed axis: Deckwright's goldfish speed score (speed) | 0.075 |
| 55 | Protection count: Hexproof, indestructible and similar (resilience) | 0.073 |
| 56 | Draw to ramp ratio: Card draw over ramp (card advantage) | 0.065 |
| 57 | Opus judge power: Opus reads the list and scores power (card quality and LLM judges) | 0.063 |
| 58 | Ritual count: One-shot mana bursts (ramp) | 0.061 |
| 59 | Source to pip ratio: Coloured sources per coloured pip (mana curve) | 0.059 |
| 60 | Effect ceiling usage: Distinct effects used, of 14 (synergy and structure) | 0.057 |
| 61 | Sonnet judge power: Sonnet reads the list and scores power (card quality and LLM judges) | 0.054 |
| 62 | Tapped land share: Share of lands entering tapped (mana curve) | 0.050 |
| 63 | Fast-tail turn to commander: Fast-tail turn the commander lands (speed) | 0.048 |
| 64 | Replacement effects: Replacement effects (forge rules text) | 0.046 |
| 65 | Activated abilities: Activated abilities (forge rules text) | 0.042 |
| 66 | Mana flood rate: Goldfish games flooding by turn 6 (resilience) | 0.042 |
| 67 | Land search share: Share of ramp that fetches lands (ramp) | 0.041 |
| 68 | Cheap draw count: Draw at 2 mana or less (card advantage) | 0.038 |
| 69 | Evasion keywords: Evasion keywords (forge rules text) | 0.037 |
| 70 | Mulligan rate: Goldfish openers the mulligan rule throws back (resilience) | 0.036 |
| 71 | Parse coverage: Share of card text the parser read (card quality and LLM judges) | 0.035 |
| 72 | Repeatable draw engines: Permanents that keep drawing (card advantage) | 0.033 |
| 73 | Role protection efficiency: Power per cost of protection cards (card power vs cost) | 0.033 |
| 74 | Static abilities: Static abilities (forge rules text) | 0.032 |
| 75 | Typal density: Share sharing the main creature type (synergy and structure) | 0.031 |
| 76 | Utility land count: Lands that do more than tap for mana (mana curve) | 0.031 |
| 77 | Counter placers: Counter placers (forge rules text) | 0.027 |
| 78 | Judge plan coherence: LLM rating of how well cards serve one plan (card quality and LLM judges) | 0.027 |
| 79 | Triggered abilities: Triggered abilities (forge rules text) | 0.026 |
| 80 | Efficiency coverage: Share of spells the power-cost score covers (card power vs cost) | 0.026 |
| 81 | Selection count: Scry, loot and filter effects (card advantage) | 0.026 |
| 82 | Copy effects: Copy effects (forge rules text) | 0.023 |
| 83 | Commander dependence: Speed lost without the commander (resilience) | 0.022 |
| 84 | Mass removal: Board wipes (forge rules text) | 0.022 |
| 85 | Colour count and pip intensity: Colours and coloured pips (mana curve) | 0.022 |
| 86 | Extra turn count: Extra-turn cards (bracket rules) | 0.021 |
| 87 | Theme density: Share on the main Deckwright theme (synergy and structure) | 0.020 |
| 88 | Cost sum: Total card cost (card power vs cost) | 0.019 |
| 89 | Haste and evasion count: Haste and evasion granters (speed) | 0.016 |
| 90 | Win clause count: Cards that win outright (tutors and combos) | 0.015 |
| 91 | Card selection: Scry, surveil and dig (forge rules text) | 0.015 |
| 92 | Targeted removal: Targeted removal (forge rules text) | 0.013 |
| 93 | Role wincon efficiency: Power per cost of win conditions (card power vs cost) | 0.012 |
| 94 | Cost reducers: Cost reducers (forge rules text) | 0.010 |
| 95 | Land search: Lands fetched (forge rules text) | 0.010 |
| 96 | Recursion: Graveyard recursion (forge rules text) | 0.009 |
| 97 | Token makers: Token makers (forge rules text) | 0.009 |
| 98 | Median turn to commander: Median turn the commander lands (speed) | 0.009 |
| 99 | Steal effects: Steal effects (forge rules text) | 0.009 |
| 100 | Protection keywords: Protection keywords (forge rules text) | 0.008 |
| 101 | Power sum: Total card power (card power vs cost) | 0.006 |
| 102 | Effect graph density: How connected enablers and payoffs are (synergy and structure) | 0.005 |
| 103 | Judge pairwise strength: LLM head-to-head deck comparisons (card quality and LLM judges) | 0.004 |
| 104 | Mana density: Lands plus ramp over deck size (ramp) | 0.004 |
| 105 | Deckwright top-10 card value: Value of the ten best cards (card quality and LLM judges) | 0.001 |
| 106 | Mass land denial count: Mass land denial cards (bracket rules) | 0.001 |
| 107 | Mass removal: Board wipes (interaction) | 0.001 |
| 108 | Targeted removal: Single-target removal (interaction) | 0.000 |
| 109 | Cards drawn: Cards drawn (forge rules text) | 0.000 |
| 110 | Recursion count: Recursion and reanimation (resilience) | 0.000 |
| 111 | Commander mana value: Commander's mana value (mana curve) | -0.001 |
| 112 | Sweeper to board ratio: Board wipes over own creatures (interaction) | -0.001 |
| 113 | Card advantage count: Cards that draw or make extra cards (card advantage) | -0.001 |
| 114 | Enabler-payoff pairs: Enabler and payoff pairs (synergy and structure) | -0.002 |
| 115 | Cost reducer count: Cards that reduce spell costs (ramp) | -0.002 |
Features not built or excluded
- Not built (6): Interaction held share (the simulator has no interaction cards or opponents to hold them for); Turn to engine (the simulator trace records no turn on which an engine card is cast); Combat kill turn (the goldfish simulator has no creatures in combat, so no kill turn); Damage by turn 6 (the goldfish simulator has no combat and deals no damage); Colour screw rate (the simulator’s card types carry no colours, so colour screw cannot occur); Deckwright power position (needs Deckwright simulation).
- Excluded by design (price, popularity, owner labels, game results) (12): Owner-declared bracket; Archidekt edhBracket; Playgroup power level; Competitiveness rating; Deck Elo; Commander win rate; Deck price; Card popularity rank; Deck popularity; EDHREC synergy; edhpowerlevel score as input; Salt score.
- Targets, not inputs (3): Win target; Bracket target; edhpowerlevel target.