Deckwright · Strength research · Results

Deck strength from the decklist alone

Run finished 2026-11-08Release R1 acceptedmtg-strength trunk 7a4746a

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).

Holdout R² by modelLocked: all features + similar decks0.894Served card model (+ card weights)0.857 post-hocRule baseline (bracket rules)0.485Commander only0.219Guess the average0.000
ModelHoldout R²MAE (mean absolute error)
Locked model0.8940.48
Served card model0.8570.42
Rule baseline0.4851.17
Commander only0.2191.43
Guess the average-0.0011.67

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

FamilyR² lost if removedR² lost if shuffled
Deck text and similar decks+0.09250.688
Mana curve+0.00200.112
Card power vs cost+0.0016not run
Forge rules text+0.0014not run
Interaction+0.00090.027
Bracket rules+0.00060.007
Card advantage+0.00060.011
Ramp+0.00040.109
Tutors and combos+0.00030.028
Synergy and structure+0.00010.001
Speed0.00000.007
Card quality and LLM judges-0.00010.003
Resilience-0.00040.016

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.

Sign rate by swap set: only cuts passpass bar 0.90Cut a strong card (216 swaps)0.966Add, tuning set (76)0.895Add, first held-out set (69)0.768Add, held-out-4, card terms (66)0.833

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.

AttemptHeld-out sign rateOutcome
1. Card prior0.652Failed the bar
2. Role match1.000You rejected it: skewed by price, usage and card age
3. Oracle power0.826Failed the bar and the litmus
4. Rules power (card terms)0.833Failed 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 typeWrong signShare
Cross-role7 of 3023%
Same-role4 of 3611%

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.

SwapCounted asWrong
Swiftfoot Boots to Sensei's Divining TopCross: protection to selection1 of 3
Arcane Signet to Lion's Eye DiamondSame: mana artifacts1 of 3
Hero's Downfall to Demonic ConsultationCross: removal to tutor3 of 3
Explosive Vegetation to Finale of DevastationSame: green sorceries1 of 3
Feed the Swarm to Toxic DelugeSame: black removal2 of 3
Mentor of the Meek to Grand AbolisherCross: card draw to protection3 of 3
Held-out-4 sign rate by piece of the modelpass bar 0.90Embedding and card weights only0.652 (43/66)Readable deck features only0.667 (44/66)Full served card model0.758 (50/66)Full + card terms (attempt 4)0.833 (55/66)Readable features + card terms0.924 (61/66)

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:

The best readable feature alone scores R² 0.425

Top 15 features by R² aloneDeck text embedding0.844Embedding strength probe0.765Bag of effect atoms0.727Nearest-neighbour label mean0.726Game Changer count0.425Fast mana count0.409Life costs0.404Term sums0.400Rule bracket floor0.393Ramp to curve ratio0.369Tutors0.351Mana value bucket shares0.310Mana per source0.307Efficiency by curve slot0.305Mana screw rate0.285
#Feature: what it measures (family)Out-of-fold R² alone (average guess -0.002)
1Deck text embedding: Average meaning of all card text, as numbers (deck text and similar decks)0.844
2Embedding strength probe: Strength read off the text embedding, fitted to labels (deck text and similar decks)0.765
3Bag of effect atoms: Count of every rules effect in the deck (deck text and similar decks)0.727
4Nearest-neighbour label mean: Average score of the most similar training decks (deck text and similar decks)0.726
5Game Changer count: Cards on the official Game Changers list (bracket rules)0.425
6Fast mana count: Ramp at 0-1 mana that nets mana at once (ramp)0.409
7Life costs: Costs paid in life (forge rules text)0.404
8Term sums: Raw power and cost terms per card, summed (card power vs cost)0.400
9Rule bracket floor: Lowest bracket the list qualifies for (bracket rules)0.393
10Ramp to curve ratio: Ramp count over average spell cost (ramp)0.369
11Tutors: Library searches for nonland cards (forge rules text)0.351
12Mana value bucket shares: Share of spells at each mana value (mana curve)0.310
13Mana per source: Mana made per source (forge rules text)0.307
14Efficiency by curve slot: Card power per cost at each mana value (card power vs cost)0.305
15Mana screw rate: Goldfish games missing a land drop by turn 4 (resilience)0.285
16Free interaction: Answers castable for free (interaction)0.281
17Role ramp efficiency: Power per cost of the ramp cards (card power vs cost)0.277
18Combo pieces reachable by tutors: Combo pieces the tutors can find (tutors and combos)0.267
19Cheap tutors: Tutors at 3 mana or less (tutors and combos)0.267
20Top-k efficiency: Power per cost of the best 5, 10, 20 cards (card power vs cost)0.262
21Best combo speed: Cheapest two-card combo in the deck (tutors and combos)0.260
22Land count: Lands in the deck (mana curve)0.254
23Ramp count: Cards tagged as ramp (ramp)0.237
24Commander identity: Which commander it is (deck text and similar decks)0.235
25Mana available on turn 4: Mana available on turns 3-4 (goldfish) (speed)0.232
26Tutor count: Cards that search for a nonland card (tutors and combos)0.226
27Efficiency sum: Total power per cost over all cards (card power vs cost)0.223
28Nearest-neighbour similarity: How close the nearest training decks are (deck text and similar decks)0.222
29Mana value spread: Spread of spell costs (mana curve)0.222
30Cheap ramp share: Share of ramp at 2 mana or less (ramp)0.220
31Mean nonland mana value: Average spell cost (mana curve)0.219
32Mana sources: Mana-making abilities (forge rules text)0.214
33Plan part odds: Odds each plan part shows up on time (resilience)0.205
34Stack interaction: Counterspells and stack answers (interaction)0.201
35Instant-speed interaction share: Share of answers at instant speed (interaction)0.189
36Combo assembled by turn 6: Odds a combo is assembled by turn 6 (tutors and combos)0.174
37Counterspells: Counterspells (forge rules text)0.170
38Mana colours: Colours a card can make (forge rules text)0.158
39Tipping-point mana value: Cost where most card value is reached (mana curve)0.149
40Cheap interaction: Answers at 2 mana or less (interaction)0.148
41Sacrifice costs: Costs paid by sacrificing (forge rules text)0.140
42Unrestricted tutors: Tutors that find any card (tutors and combos)0.135
43Alternative costs: Alternative ways to cast (forge rules text)0.134
44Cards seen by turn 6: Cards seen by turn 6 (goldfish) (card advantage)0.134
45Deckwright card value sum: Deckwright text-based card value, summed (card quality and LLM judges)0.132
46Two-card combo count: Two-card combos fully in the deck (tutors and combos)0.122
47Coloured source shortfall: Coloured sources short of the Karsten table (mana curve)0.114
48Type band deviation: Distance from the archetype's card-type mix (synergy and structure)0.102
49Efficiency mean: Average power per cost (card power vs cost)0.095
50Role draw efficiency: Power per cost of the draw cards (card power vs cost)0.089
51Graveyard hate and stax: Graveyard hate and static denial (interaction)0.084
52Plan redundancy: Copies per plan part against Deckwright floors (resilience)0.081
53Role interaction efficiency: Power per cost of the answers (card power vs cost)0.076
54Deckwright speed axis: Deckwright's goldfish speed score (speed)0.075
55Protection count: Hexproof, indestructible and similar (resilience)0.073
56Draw to ramp ratio: Card draw over ramp (card advantage)0.065
57Opus judge power: Opus reads the list and scores power (card quality and LLM judges)0.063
58Ritual count: One-shot mana bursts (ramp)0.061
59Source to pip ratio: Coloured sources per coloured pip (mana curve)0.059
60Effect ceiling usage: Distinct effects used, of 14 (synergy and structure)0.057
61Sonnet judge power: Sonnet reads the list and scores power (card quality and LLM judges)0.054
62Tapped land share: Share of lands entering tapped (mana curve)0.050
63Fast-tail turn to commander: Fast-tail turn the commander lands (speed)0.048
64Replacement effects: Replacement effects (forge rules text)0.046
65Activated abilities: Activated abilities (forge rules text)0.042
66Mana flood rate: Goldfish games flooding by turn 6 (resilience)0.042
67Land search share: Share of ramp that fetches lands (ramp)0.041
68Cheap draw count: Draw at 2 mana or less (card advantage)0.038
69Evasion keywords: Evasion keywords (forge rules text)0.037
70Mulligan rate: Goldfish openers the mulligan rule throws back (resilience)0.036
71Parse coverage: Share of card text the parser read (card quality and LLM judges)0.035
72Repeatable draw engines: Permanents that keep drawing (card advantage)0.033
73Role protection efficiency: Power per cost of protection cards (card power vs cost)0.033
74Static abilities: Static abilities (forge rules text)0.032
75Typal density: Share sharing the main creature type (synergy and structure)0.031
76Utility land count: Lands that do more than tap for mana (mana curve)0.031
77Counter placers: Counter placers (forge rules text)0.027
78Judge plan coherence: LLM rating of how well cards serve one plan (card quality and LLM judges)0.027
79Triggered abilities: Triggered abilities (forge rules text)0.026
80Efficiency coverage: Share of spells the power-cost score covers (card power vs cost)0.026
81Selection count: Scry, loot and filter effects (card advantage)0.026
82Copy effects: Copy effects (forge rules text)0.023
83Commander dependence: Speed lost without the commander (resilience)0.022
84Mass removal: Board wipes (forge rules text)0.022
85Colour count and pip intensity: Colours and coloured pips (mana curve)0.022
86Extra turn count: Extra-turn cards (bracket rules)0.021
87Theme density: Share on the main Deckwright theme (synergy and structure)0.020
88Cost sum: Total card cost (card power vs cost)0.019
89Haste and evasion count: Haste and evasion granters (speed)0.016
90Win clause count: Cards that win outright (tutors and combos)0.015
91Card selection: Scry, surveil and dig (forge rules text)0.015
92Targeted removal: Targeted removal (forge rules text)0.013
93Role wincon efficiency: Power per cost of win conditions (card power vs cost)0.012
94Cost reducers: Cost reducers (forge rules text)0.010
95Land search: Lands fetched (forge rules text)0.010
96Recursion: Graveyard recursion (forge rules text)0.009
97Token makers: Token makers (forge rules text)0.009
98Median turn to commander: Median turn the commander lands (speed)0.009
99Steal effects: Steal effects (forge rules text)0.009
100Protection keywords: Protection keywords (forge rules text)0.008
101Power sum: Total card power (card power vs cost)0.006
102Effect graph density: How connected enablers and payoffs are (synergy and structure)0.005
103Judge pairwise strength: LLM head-to-head deck comparisons (card quality and LLM judges)0.004
104Mana density: Lands plus ramp over deck size (ramp)0.004
105Deckwright top-10 card value: Value of the ten best cards (card quality and LLM judges)0.001
106Mass land denial count: Mass land denial cards (bracket rules)0.001
107Mass removal: Board wipes (interaction)0.001
108Targeted removal: Single-target removal (interaction)0.000
109Cards drawn: Cards drawn (forge rules text)0.000
110Recursion count: Recursion and reanimation (resilience)0.000
111Commander mana value: Commander's mana value (mana curve)-0.001
112Sweeper to board ratio: Board wipes over own creatures (interaction)-0.001
113Card advantage count: Cards that draw or make extra cards (card advantage)-0.001
114Enabler-payoff pairs: Enabler and payoff pairs (synergy and structure)-0.002
115Cost reducer count: Cards that reduce spell costs (ramp)-0.002

Features not built or excluded