Grim.Cards Simulation Study — Edition 2026-07-08

Data snapshot: 2026-07-08 · Dataset version: 2.4 · Published: 2026-07-08 Permanent URL: grim.cards/case-study/2026-07-08 License: Creative Commons Attribution 4.0 International (CC BY 4.0) Cohort: Real, human-submitted Commander and Standard decks — automated Crucible decks and system sample decks excluded throughout.

Table of Contents

1. Executive Summary
2. Methodology & Provenance
3. Dataset Overview
4. The Gauntlet: Matchup Results
5. Top Commanders
6. Top Cards by Usage
7. Win-Rate Distribution
8. Deck Iteration & Retests
9. Monthly Trends
10. Color-Identity Breakdown
11. Construction & Win-Rate Correlations
12. Card-Category Insights (heuristic)
13. Most Impactful Cards: Decision Impact (counterfactual proxy)
14. Card Performance Roll-Up: Board Impact (measured proxy)
15. Tempo & Game-Length Signals
16. Matchup Structure: Deck vs. Opponent
17. Per-Deck Appendix (Anonymized)
18. Power Score Distribution (secondary, internal composite only)
19. Key Findings & What We Cannot Conclude Yet
20. Future-Comparison Baseline Table
21. Limitations
22. License & Citation

1. Executive Summary

Between 2026-05-14 and 2026-07-08, Grim.Cards ran 4,972 AI-versus-AI simulated games across 344 completed deck tests submitted by 119 unique players. The cohort covers 253 player-submitted decks — 218 in Commander, 35 in Standard — tested against a fixed gauntlet of competitive meta opponents.

Win rate is defined throughout as wins ÷ total games played, with draws included in the denominator. Every statistic in this report is split by format; the two formats are never pooled into a single aggregate rate.

Headline numbers:

The sub-50% headline rates are structurally expected: a one-vs-the-field gauntlet means the field, taken together, wins more than any single challenger. The figures are not a verdict on deck quality; they are the measured outcomes of these simulations on these decklists against these opponents.

What they reveal is more interesting than the headline alone. In Commander, the same pool of 218 decks wins 53.5% against one opponent and only 31.2% against another — a 22.3-point swing produced entirely by the choice of meta opponent, not a change in the decks themselves. In Standard, that swing reaches 47.6 points. The opponent facing a deck matters enormously in this data; the section on matchup structure quantifies precisely how much.

2. Methodology & Provenance

Simulation method

Every game is played AI versus AI on a custom build of the open-source Forge engine. Player-submitted decks are piloted by Forge's AI against a fixed gauntlet of reference meta decks. No human plays any game; results describe how the engine pilots these decklists, not human player skill.

Win-rate definition

Win rate = wins ÷ total games played (draws included in the denominator). This definition is applied consistently to every figure in this report unless explicitly labelled otherwise. For a deck with W wins, L losses, and D draws: win rate = W ÷ (W + L + D).

Cohort definition

• Included: Real, human-submitted decks (user_id ≠ '__grinder__', is_sample = false), with at least one completed simulation.
• Excluded: Automated Crucible reference decks; system sample decks; any simulations without a completed result.
• Format split: Commander and Standard are reported separately in every section. The format field is obtained by joining simulation records to deck metadata; the simulation table itself carries no format column.

Minimum cohort threshold

No rate or breakdown is published when the underlying cohort is fewer than 10 unique decks. Cohorts below this floor are suppressed or merged into "Other" and noted.

Retest identification

A deck is classified as "retested" when it has more than one completed simulation on record. The delta in win rate is measured from the deck's first completed simulation to its most recent completed simulation.

Category classification (heuristic)

Functional card categories (tutor/search effects, sacrifice outlets, discard effects, reanimation effects) are assigned by a keyword heuristic applied to oracle text and pre-existing card preference flags. This is not exhaustive; edge cases may be mislabelled. All category-level findings carry a heuristic label.

Decision impact (counterfactual proxy)

Where reported, decision impact is derived from Grim.Cards' counterfactual engine: for each recorded decision involving a card, the engine computes the delta between the line actually played and the engine's own next-best alternative. A positive decision-impact value means the played line outperformed the alternative; a negative value means the alternative would have scored better on average. This is a play-quality proxy on replayed decisions — not a damage count, kill count, or causal win contribution.

Sign convention: raw internal values follow the negative-is-better-for-the-alternative convention. In this report, improvement is shown as a positive number and worse-than-alternative is shown as a negative number, consistently with the Grim.Cards counterfactual edge-sign convention.

Card board impact (measured proxy)

Card performance figures pool per-deck board-quality deltas across every deck in the cohort running a given card. A positive value means the board state measurably improved around turns the card was seen; a negative value means it declined. This is a board-state proxy, not damage, kills, or a causal win claim. Only cards appearing in at least 10 distinct decks with at least 25 recorded observations are ranked.

Correlation disclaimer

Every construction, color, category, and card-level breakdown in this report is correlational only. The dataset is a self-selected, non-random sample. No finding implies a causal relationship, and nothing in this report constitutes strategy advice.

3. Dataset Overview

Data window: 2026-05-14 to 2026-07-08

Commander dominates both by volume (86% of decks, 87% of games) and by player count. The Standard cohort, at 35 decks and 633 games, clears every minimum threshold but should be read as a smaller, less stable sample. Standard win-rate figures are real and reported, but they carry wider uncertainty than Commander figures derived from nearly seven times as many games.

4. The Gauntlet: Matchup Results

The Gauntlet is the core of Grim.Cards' simulation: each deck is played against a fixed set of meta opponents, and the results are the primary objective outcomes of the study. Every matchup cell covers all decks in that format, so the cohort size is stable across opponents within a format.

4.1 Commander Gauntlet (n = 218 decks per opponent; 855 games per opponent)

The field is not uniform. Player Commander decks win the majority of their games against Breya Artifact Combo (53.5%) and Derevi Bant Control (52.2%), flip to a slight deficit against Aesi Landfall (47.5%), then fall sharply against Atraxa Superfriends (33.7%) and Edgar Markov Vampires (31.2%). The spread from best to worst matchup is 22.3 percentage points — all measured on an identical 218-deck sample.

This spread is the most important structural fact in the Commander data. It means a deck's measured win rate in any single matchup tells you nearly as much about that particular opponent as it does about the deck.

4.2 Standard Gauntlet (n = 35 decks per opponent; 126 games per opponent)

In Standard, the opponent effect is even more pronounced. The same 35-deck pool wins 61.9% against Temur Harmonizer Combo but only 14.3% against Mono Red Aggro — a 47.6-point swing. The Mono Red Aggro matchup is the most lopsided result in the dataset: 18 wins against 108 losses across 126 games. The Standard cohort is small (35 decks), so per-matchup estimates carry more uncertainty, but the directional spread is unambiguous.

5. Top Commanders

The following commanders meet the minimum cohort threshold of 10 decks. Only Meren of Clan Nel Toth (11 decks) clears this floor in the current dataset. All other commanders listed in the underlying data fall below 10 unique decks, and their win rates are therefore suppressed per the study's privacy and reliability threshold.

Qualifying Commanders (n ≥ 10 decks)

Meren of Clan Nel Toth is the only commander in this snapshot with enough distinct decks to report a stable win rate. At 40.6% across 165 games in 11 decks, the figure sits 3.2 points below the Commander format average of 43.8%. This is a descriptive observation on a small cohort; it reflects these 11 submitted decks' performance against this gauntlet, not a universal statement about the commander.

Commanders with fewer than 10 decks are present in the dataset but are suppressed here. The next most-tested commanders by deck count include Ureni of the Unwritten (7 decks), The Ur-Dragon (5 decks), The First Sliver (4 decks), and several others — each below the publication threshold for win-rate reporting. As the dataset grows, more commanders will cross into reportable territory in future editions.

6. Top Cards by Usage

Basic lands are excluded throughout. Win rates are suppressed for any card appearing in fewer than 10 distinct decks.

6.1 Commander — Most-Played Cards (n ≥ 10 decks)

The breadth of adoption for Sol Ring (191 of 218 decks, 87.6% penetration) and Command Tower (168 decks, 77.1%) reflects the Commander format's structural staple economy — these cards are nearly ubiquitous. Their containing-deck win rates (44.2% and 43.2% respectively) are essentially identical to the format average of 43.8%, which is exactly what one would expect from cards present in nearly every deck: they track the field.

More interesting are the cards where containing-deck win rate diverges from the mean. Kodama's Reach decks win at 53.9% (35 decks, 621 games), Path of Ancestry at 51.2% (60 decks, 1,165 games), and Cultivate at 49.6% (48 decks, 842 games) — all meaningfully above average. At the other end, Reliquary Tower decks win at 39.0% (70 decks, 1,378 games) and Terramorphic Expanse at 39.2% (32 decks, 531 games). These are correlations between card inclusion and deck-level win rates, not causal claims.

6.2 Standard — Most-Played Cards (n ≥ 10 decks)

Only Lightning Bolt clears the 10-deck floor in Standard. At 44.0% across 150 games in 10 decks, its containing-deck win rate is 6.7 points above the Standard format average of 37.3%. All other Standard cards in the dataset fall below 10 distinct decks and are suppressed. The Standard card list reflects a small, concentrated sample — most cards that appear do so in fewer than 10 decks.

7. Win-Rate Distribution

Win rates are reported at the deck level, one win rate per deck across all of that deck's games in the relevant format.

7.1 Commander (n = 218 decks)

Summary statistics (Commander, n = 218): Mean 44.3% · Median 43.3% · Min 6.7% · Max 86.7%

The Commander distribution is left-skewed in a specific way: the modal bracket sits at 30–40% (57 decks), but a long right tail — 35 decks between 50–60%, 23 between 60–70%, 13 between 70–80%, and 6 above 80% — pulls the mean (44.3%) above the median (43.3%). The majority of decks cluster in the 30–50% range; genuinely dominant decks (>70% win rate) are a small minority at 19 total.

The 9 decks below 10% win rate represent the other tail: these are the hardest-tested decks in the dataset against this particular gauntlet, not necessarily the "worst" decks in any absolute sense.

7.2 Standard (n = 35 decks)

Summary statistics (Standard, n = 35): Mean 35.8% · Median 33.3% · Min 0% · Max 80%

The Standard distribution is similarly concentrated at 30–40% (12 decks), with a thin right tail and a meaningful left tail (4 decks at 0–10%). The gap between mean (35.8%) and median (33.3%) again reflects skew from a small number of high-performing decks. With 35 decks total, each bracket is thin and individual bracket counts should be interpreted cautiously.

8. Deck Iteration & Retests

Retests are Commander decks that have been submitted for simulation more than once; this section tracks the change in win rate from a deck's first completed simulation to its most recent.

Standard retest data does not meet the minimum cohort threshold and is not reported.

Commander Retests (n = 53 retested decks)

Of the 53 Commander decks that were submitted for at least two completed simulations, 23 showed a higher win rate on their latest test and 21 showed a lower win rate — nearly a coin flip between improved and declined. The average change across all 53 retested decks is +1.6 percentage points, a modest positive figure.

Several important caveats apply. Different simulations of the same deck can produce different results by design (the engine's game-by-game variation is intentional). More practically, monthly cohorts differ — the decks retested in later months are not always the same population as those first tested in May. The +1.6-point average is a descriptive observation on this sample of 53 decks, not a claim that iteration reliably improves win rates across all decks.

9. Monthly Trends

Monthly figures represent completed simulations run within each calendar month. Because different decks enter the dataset each month, monthly win-rate shifts reflect who submitted that month, not longitudinal improvement of any single deck.

Commander win rate has moved from 40.4% in May to 43.9% in June to 47.4% in July (partial month, 64 simulations). The 7.0-point rise from May to July is real in these numbers but must be read carefully: the July figures cover only the first eight days of the month (the snapshot date is 2026-07-08), and each monthly cohort is a different mix of decks. Standard only has sufficient data from June (24 simulations, 20 decks, 37.5%), so no Standard trend can be reported.

The volume growth from May (74 Commander sims) to June (163 Commander sims) is the clearest structural story: the platform roughly doubled in simulation volume month-over-month, providing a firmer statistical base in June and July than in the inaugural month.

10. Color-Identity Breakdown

Colors overlap within decks — a five-color deck counts toward all five color buckets — so these figures are not independent effects. The win rate for "decks including Red" is the aggregate win rate of all decks in the color-identity group, not of Red cards specifically.

10.1 Commander Color Presence (n = 218 decks)

Red-containing Commander decks (n=101) post the highest group win rate at 47.0%, 3.2 points above the format average. Green-containing decks (n=124) follow at 45.7%. Blue, Black, and White cluster tightly at 42.9%, 42.8%, and 42.6% respectively. The spread across all five colors is only 4.4 percentage points — a narrower gap than the matchup spread reported in Section 4. Color overlap confounds any interpretation; a Red/Green/Blue deck contributes to three color buckets simultaneously.

10.2 Standard Color Presence (n = 35 decks)

In Standard, Red-containing decks (n=18) again lead at 37.5% — essentially at the format average of 37.3% — while Black-containing decks (n=12) trail at 27.2%. Standard cohort sizes per color are thin (10–18 decks), so color-level estimates carry significant uncertainty. The observed spread (10.3 points from Red to Black) may reflect deck archetype composition as much as color choice.

11. Construction & Win-Rate Correlations

All construction findings are correlational. Deck features (land ratios, color counts, type mixes) are measured from deck submissions; win rates are measured from simulations. No causal direction can be inferred from these associations.

11.1 Color Count vs. Win Rate (Commander)

(4-color decks suppressed: below 10-deck floor.)

Mono-color Commander decks (n=32) post the highest average win rate at 51.7%, while two-color decks (n=68) post the lowest at 39.2%. Three-color (n=87) and five-color (n=27) decks land between these extremes. This is a correlation; the decks in each color-count group differ in archetype, commander, and construction in ways that likely explain much of this spread.

(Standard: only 2-color decks — n=21, 35.7% — clear the threshold. No other Standard color-count bracket reaches 10 decks.)

11.2 Land Ratio vs. Win Rate

Commander:

(Bands represent coarsened land percentage groups.)

Commander decks in the ~40% land band (n=54) win at 49.5% — 7 points above the ~35% band (n=132, 42.5%) and 6.2 points above the ~30% band (n=25, 43.3%). The ~35% band is the modal group, containing the plurality of submitted decks.

Standard:

Standard decks in the ~40% land band (n=18) win at 43.7%, 10.4 points above the ~35% band (n=10, 33.3%). Given the small cohort sizes, this difference should be treated as a directional observation.

11.3 Card-Type Mix by Win-Rate Bracket (Commander)

High-win-rate Commander decks (n=58, avg 68.7%) average 36.4% lands — 1.0 point above the low bracket's 35.4% (n=78). Their average mana value (3.45) is 0.37 higher than low-bracket decks (3.08), and their average creature percentage (29.7%) is 1.8 points higher, while their average spell percentage (16.7%) is 2.3 points lower than low-bracket decks (19.0%). These are descriptive correlations across the submitted deck sample; they describe what decks in each bracket look like, not what causes high win rates.

Overall Commander type mix (n = 218 decks): Lands 35.9% · Creatures 29.0% · Instants 9.4% · Sorceries 8.0% · Combined instants/sorceries 17.6% · Artifacts 9.2% · Enchantments 7.7% · Planeswalkers 0.6%. (Instant/sorcery split available from 218 decks; note that unclassifiable spells exist only in the combined figure.)

Overall Standard type mix (n = 35 decks): Lands 36.5% · Creatures 28.7% · Instants 14.1% · Sorceries 10.3% · Combined instants/sorceries 24.7% · Artifacts 5.0% · Enchantments 4.8% · Planeswalkers 0.5%.

(Standard win-bracket breakdown: high-win-rate bracket does not clear the 10-deck floor. Mid bracket: n=12, 45.8% avg win rate, 37.7% lands, avg MV 2.06. Low bracket: n=19, 22.3% avg win rate, 35.9% lands, avg MV 2.99.)

12. Card-Category Insights (heuristic)

⚠ Heuristic label: Functional card categories are assigned by keyword matching against oracle text and pre-existing card tags. Category membership may misclassify edge cases. All figures in this section are correlational — they describe the win rates of decks containing at least one card in each category, not the causal effect of including those cards.

12.1 Commander (n = 218 decks)

(Sacrifice outlets: "without" group has only 5 decks — below the 10-deck floor — so the contrast and delta are suppressed.)

Three of the four categories show the same directional pattern in Commander: decks without the category win at a higher rate than decks with it. The discard-effect contrast is the largest of the three reportable pairs: 43.6% with (n=188) vs. 48.8% without (n=30), a −5.2-point gap. Reanimation follows at −4.5 points (n=138 vs. n=80), and tutor/search at −3.7 points (n=206 vs. n=12).

This pattern most likely reflects deck-composition confounding: decks built around discard, reanimation, and tutoring strategies tend to be built differently in other ways that affect their matchup against this specific gauntlet. It does not mean these card types are harmful to include.

12.2 Standard (n = 35 decks)

In Standard, the same directional pattern appears with larger gaps. The reanimation contrast is the sharpest: decks with reanimation effects win at 24.7% (n=10) versus 40.3% without (n=25), a −15.6-point difference. The tutor gap is −12.9 points (n=15 vs. n=20). Standard cohort sizes are small enough that these gaps could shift substantially with additional data. All figures are heuristic-categorized, correlational, and not a basis for deck-construction conclusions.

13. Most Impactful Cards: Decision Impact (counterfactual proxy)

⚠ Counterfactual proxy label: Decision-impact figures are derived from Grim.Cards' counterfactual engine, which replays recorded game decisions and compares the line actually taken against the engine's own next-best alternative. A positive value means the played line outperformed the alternative on average; a negative value means the alternative would have scored better. This is a play-quality proxy on a small number of recorded decisions — not a damage count, kill count, or win attribution. Correlation-not-causation applies.

Sign convention: improvement over the alternative = positive; worse than the alternative = negative.

Commander (minimum: 10 decks, 25+ observations)

Lightning Greaves is the only Commander card meeting both the deck-count floor (10) and the observation floor (25 observations) in the current dataset. Its average decision impact of −242.77 points indicates that, across the 13 recorded decisions involving it, the engine's alternative line would on average have scored better by 242.77 counterfactual points. This is a play-quality signal on a sparse sample of 13 decisions across 10 decks — it measures how the engine navigated decisions in spots involving this card, not the card's intrinsic quality.

(Standard: no cards meet both floors in the current Standard dataset. Section omitted.)

14. Card Performance Roll-Up: Board Impact (measured proxy)

⚠ Measured proxy label: Card performance figures pool per-deck board-quality deltas across all qualifying decks in the cohort that run a given card. A positive figure means the board state measurably improved on turns the card appeared; a negative figure means it declined. This is a board-state proxy — not damage, kills, or a causal win contribution. Rankings are weighted by breadth: a card appearing in more distinct decks with more total observations is more trustworthy. Card-type and color buckets pool very different individual cards; aggregate figures describe the bucket's central tendency, not any individual card. Multicolor cards count toward each of their colors in the color breakdown. Only cards in ≥10 decks with ≥25 observations are ranked.

14.1 Commander — Top Board-Impact Cards (n = 66 qualifying cards)

Atarka, World Render leads the board-impact rankings at +20.84 points per observation across 51 observations in 10 decks, closely followed by Terror of the Peaks (+16.84, 11 decks, 50 obs). Both are creature-type cards with large board-state effects; their high figures reflect the board-quality change measured on turns they were in play. Swords to Plowshares stands out as the highest-impact non-creature card and the one with the broadest evidence base: +6.39 across 132 observations in 35 distinct decks, making it the most reliably measured single card in the top list.

The Dragon-tribal cluster — Atarka, Terror, Lathliss, Miirym, Dragon's Hoard, Dragon Tempest — dominates the upper end of the ranking. This likely reflects the submitted deck composition: Commander is heavily represented by Dragon-tribal builds, and those strategies share board-quality patterns that the metric captures.

14.2 Commander — Bottom Board-Impact Cards

Damnation posts the lowest board impact at −9.89 across 36 observations in 10 decks. A negative board-impact figure for a board wipe like Damnation reflects the nature of the metric: destroying all creatures, including the player's own, represents a net board-state decline in the turn it resolves, even if the play is tactically correct. This is an important caveat for interpreting the bottom rankings — many of the cards here (Damnation, Blasphemous Act, Carrion Feeder, Ashnod's Altar) are designed to create temporary board-state sacrifices in exchange for longer-term strategic advantage. A low board-impact score does not mean these cards are poorly designed or strategically counterproductive; it means the board measurably got worse on the turns they were recorded.

Smothering Tithe (−5.57, 17 decks, 86 obs) and Rhystic Study (−4.22, 21 decks, 77 obs) — two of Commander's most commonly cited advantage engines — show negative board impact, possibly because their effect on board quality is diffuse and delayed rather than immediate.

(Standard: no cards meet the qualification floor of 10 decks and 25 observations in the current Standard dataset. Section omitted.)

14.3 Commander — Board Impact by Card Type

Creatures carry the highest pooled board impact (+2.51 per observation across 27,067 observations — by far the largest evidence base). Instants are modestly positive (+0.58). Sorceries, enchantments, and artifacts are all negative, with artifacts at the bottom (−1.89 across 4,570 observations). These are aggregate figures pooling very different individual cards within each type; the creature figure, for instance, is pulled upward by the dragon-tribal cards that dominate the top-performers list. The type breakdown describes central tendencies, not universal rules.

14.4 Commander — Board Impact by Color Identity

Red cards post the highest pooled board impact (+5.12 across 8,360 observations in at least 26 decks), driven in part by the dragon-tribal cards that dominate the top of the individual rankings. Colorless cards — primarily mana rocks and utility artifacts — post the only negative average (−1.14), consistent with the artifact type's negative aggregate in the type breakdown. Multicolor cards count toward each of their component colors, so these figures overlap.

15. Tempo & Game-Length Signals

Game length is measured as the mean of per-match average, shortest, and longest game turn counts across all decks in the format.

Commander games in this dataset run approximately 10.4 turns on average, with typical games completing between 8 and 12 turns. Standard games are slightly shorter at 9.5 turns average, with a tighter range of 7.3–11.7 turns. These are rough tempo markers for how long the Forge engine takes to close matchups against the gauntlet opponents on these decklists; they reflect engine behavior and not human play speed or style.

The difference between formats — 0.9 turns shorter in Standard — is directionally consistent with Standard's faster mana curve (average mana value 2.06–2.99 in Standard brackets vs. 3.08–3.45 in Commander brackets) and the higher proportion of instants/sorceries in Standard's card-type mix (24.7% combined vs. 17.6% in Commander).

16. Matchup Structure: Deck vs. Opponent

This section examines how much of the observed variation in matchup win rates is attributable to the deck itself versus the specific Gauntlet opponent — or some combination of the two. All analyses are descriptive decompositions of simulated results. No causal claim is made.

Commander only. Standard matchup-structure data does not meet the minimum thresholds for this analysis.

16.1 Power Score Band × Opponent Win-Rate Grid

Decks with a recorded Power Score were divided into three bands. Each cell reports the win rate of that band against a specific opponent. Minimum per-cell floor: 5 games.

Band definitions (Commander, n = 49 decks across all three bands):

Per-cell win rates:

The gradient across Power Score bands is consistent and steep in every matchup. High-PS decks win at 54.8%–76.3% depending on opponent; low-PS decks win at 14.9%–33.3%. The within-band spread (the range across opponents within a single PS band) is also substantial: for high-PS decks, the spread from best matchup (76.3% vs. Breya) to worst (54.8% vs. Edgar) is 21.5 points; for low-PS decks, the spread from best (33.3% vs. Derevi) to worst (14.9% vs. Atraxa) is 18.4 points. Both the deck's Power Score and the opponent's identity independently associate with outcomes in this data.

16.2 Matchup-Inversion Rate

Of the comparable deck pairs (pairs where both decks faced the same two meta opponents with sufficient games per cell), 37.4% reversed their preference order between the two opponents — that is, deck A beat opponent X more than deck B did, but deck B beat opponent Y more than deck A did.

An inversion rate of 37.4% means that more than one in three comparable deck pairs disagree on which of two opponents is easier — the preferred opponent for one deck is the harder one for the other. This is a measured property of the data's structure: the Gauntlet opponent meaningfully reshuffles deck outcomes, not just shifts them uniformly up or down.

16.3 Games-Weighted Variance Decomposition

The total variance in per-(deck, opponent) win-rate cells was decomposed into a deck main effect and an opponent main effect.

Coverage: 49 decks, 16 opponents, 226 cells.

The deck itself accounts for the largest share of win-rate variation (57.9%). The opponent accounts for 14.7% — meaningful but secondary. The residual (27.4%) captures interaction effects and noise: instances where a specific deck–opponent combination produces an outcome not predictable from either the deck or the opponent alone.

This decomposition is descriptive, not a significance test. It does not establish that deck quality "causes" wins; it describes how much of the observed variance tracks the deck versus the opponent in this simulation dataset.

16.4 Per-Opponent Spearman Correlation: Power Score vs. Win Rate

Across all five Commander meta opponents, the Spearman rank correlation between a deck's Power Score and its win rate against that opponent ranges from 0.72 to 0.87 — a consistently strong positive association. The highest correlation (0.87) appears against Atraxa Superfriends, the second-worst matchup by overall win rate; the lowest (0.72) against Derevi Bant Control, the second-best matchup.

These are rank correlations — they measure whether higher-PS decks tend to win more against each opponent — not effect-size estimates. Power Score is an internal composite (see Section 18); the correlations describe how well it ranks decks within this dataset's simulation outcomes, not whether it measures any objective external construct.

17. Per-Deck Appendix (Anonymized)

Privacy: Pseudonyms (C-001 through C-049) are randomly reassigned every edition and carry no link to any user account or deck identity. Published values are deliberately coarsened: Power Score is rounded to the nearest 5, win rates and matchup spreads to whole percentage points, and game counts are banded. Only derived statistics are shown — never deck names, decklists, card lists, user identifiers, or account links.

Coverage: 49 Commander decks with ≥20 games. Standard: below the appendix floor; omitted.

Summary (n = 49 Commander decks):

17.1 OLS Residual Outliers

Among the 49 published decks, OLS regression of win rate on (coarsened) Power Score identifies two notable outliers — the widest overperformance and underperformance relative to Power Score expectation:

C-035 wins 80% despite a Power Score of 60 — 19 points above the OLS expectation for a deck at that score. C-032 wins only 28% despite a Power Score of 35 — 9 points below expectation. Both are descriptive observations from the regression fit on this 49-deck anonymized sample; they highlight that Power Score does not fully explain win-rate variation, which is consistent with the 27.4% residual observed in the variance decomposition.

17.2 Full Anonymized Per-Deck Table (Commander, n = 49)

Spread (—) = fewer than 3 qualifying opponent matchup cells; spread not computed. PS = Power Score rounded to nearest 5.

18. Power Score Distribution (secondary, internal composite only)

Important caveat: Power Score is an internal Grim.Cards composite indicator, not an objective or universal measure of deck strength. It is one input derived from simulated game features combined with other constructed factors. It is not a win-rate measurement. It should not be used to rank, compare, or judge decks in isolation. It appears here as a secondary, context-only section at the end of this report.

The following distribution covers completed user simulations with a recorded grade.

The grade distribution is right-skewed toward lower grades, with F (93 decks) and D (81 decks) accounting for the large majority. The S grade has only 1 deck — below the 10-deck reporting floor for any further breakdown of that grade. Grade distributions reflect the internal composite's calibration against this gauntlet's specific opponents and are not transferable to other contexts.

19. Key Findings & What We Cannot Conclude Yet

What the data shows

1. Opponent identity is a primary driver of matchup outcomes. The same 218-deck Commander pool wins 53.5% against Breya Artifact Combo and only 31.2% against Edgar Markov Vampires — a 22.3-point spread produced entirely by changing the opponent, not the decks. In Standard, the spread reaches 47.6 points (61.9% vs. Temur Harmonizer Combo; 14.3% vs. Mono Red Aggro). Any single-matchup win rate, read in isolation, reflects both the deck and the opponent.

2. Deck identity explains more variance than opponent identity — but not all of it. The games-weighted variance decomposition (n=49 Commander decks, 16 opponents) attributes 57.9% of matchup win-rate variance to the deck main effect and 14.7% to the opponent main effect, with 27.4% residual. The deck matters more than the opponent in aggregate, but the opponent effect is real and non-trivial.

3. Win rates span an 80-point range in Commander. Across 218 Commander decks, measured win rates range from 6.7% to 86.7% (median 43.3%). The distribution is left-skewed: the modal bracket is 30–40% (57 decks), while 19 decks exceed 70%. In Standard (35 decks), the range is 0%–80% (median 33.3%).

4. The most popular card is not the highest-performing one. Sol Ring (191 Commander decks, 87.6% penetration) posts a containing-deck win rate of 44.2% — essentially at the format mean. Kodama's Reach (35 decks, 53.9%) and Path of Ancestry (60 decks, 51.2%) are both less popular and associated with higher win rates. Popularity and performance diverge across the most-played card list.

5. Dragon-tribal cards dominate the board-impact rankings. Among the 66 Commander cards meeting the qualification floor (≥10 decks, ≥25 observations), the top five by board impact are all associated with Dragon-tribal strategies: Atarka, World Render (+20.84 pts/obs, 10 decks), Terror of the Peaks (+16.84, 11 decks), Lathliss, Dragon Queen (+9.06, 13 decks), Miirym, Sentinel Wyrm (+6.67, 10 decks), and Dragon's Hoard (+3.02, 11 decks). This reflects the composition of submitted decks as much as any property of the cards themselves.

6. Swords to Plowshares is the most broadly evidenced high-impact card. With +6.39 board-impact points per observation across 132 observations in 35 distinct decks, it carries the largest evidence base of any card in the top-performers list — more decks and more observations than any competitor in the positive tier.

7. Cards designed to sacrifice board position show negative board-impact scores. Damnation (−9.89, 10 decks, 36 obs), Blasphemous Act (−5.03, 14 decks, 39 obs), Ashnod's Altar (−6.61, 18 decks, 49 obs), and Carrion Feeder (−5.14, 10 decks, 37 obs) all appear in the bottom tier. The metric captures immediate board-state change; cards designed to sacrifice the board for strategic gain will mechanically register negative values in the turns they resolve. These low scores describe the metric's behavior on these card types — not a strategic verdict on the cards.

8. Retested decks show a slight average improvement, but the split is nearly even. Among 53 retested Commander decks, 23 improved and 21 declined (9 flat), with an average change of +1.6 percentage points. The near-even split makes the average directional but not a reliable predictor for any individual deck.

9. Power Score correlates strongly with win rate across all Commander Gauntlet opponents. Per-opponent Spearman ρ values range from 0.72 (Derevi Bant Control) to 0.87 (Atraxa Superfriends) across n=43 decks. The association is consistent and strong in rank terms, though Power Score is an internal composite and the residual (27.4% of variance unexplained by deck or opponent) shows it does not fully determine outcomes.

10. The matchup-inversion rate is 37.4%. More than one in three comparable Commander deck pairs (2,110 of 5,637 pairs across 10 opponent pairings) reverse their preference order between two meta opponents. The Gauntlet is not a uniform difficulty scale; the specific opponent genuinely reshuffles deck orderings.

What we cannot conclude yet

No causal claims on construction. Every correlation between land count, color, card type, or card category and win rate is a descriptive association in a self-selected, non-random sample. Decks in each construction group differ from one another in dozens of ways beyond the measured variable.

No conclusions about Magic as a game. These results describe how Forge's AI pilots these specific decklists against these specific gauntlet opponents. They do not generalize to human play, other metagames, or other simulation environments.

Commander format trends are not yet stabilized. The monthly Commander win rate moved from 40.4% (May, 74 sims) to 43.9% (June, 163 sims) to 47.4% (July, 64 sims — partial month). Three data points with different deck compositions cannot establish a trend. Future editions will clarify whether the July uptick persists.

Most commanders cannot yet be reported. Only Meren of Clan Nel Toth (11 decks) clears the 10-deck floor. All other commanders have too few distinct submitted decks for a publishable win-rate estimate. This will improve as the dataset grows.

Standard results are directional, not conclusive. At 35 decks and 633 games, the Standard cohort clears every minimum threshold but is five to six times smaller than the Commander sample. All Standard figures carry materially wider uncertainty.

Functional category effects may be confounded. The negative correlations between discard, reanimation, and tutor effects and win rate in both formats (e.g., −15.6 points for Standard reanimation, n=10) most plausibly reflect the construction profiles of decks that include these categories, not the categories themselves. Category membership is also assigned by heuristic and may mislabel edge cases.

Decision-impact data is sparse. Only one Commander card (Lightning Greaves, 10 decks, 13 observations) meets the qualification floor for the counterfactual decision-impact section. The 13-decision sample is too thin for robust conclusions; this section will become more informative as more simulations accumulate.

20. Future-Comparison Baseline Table

The following table records the key metrics from this edition as a baseline for comparison in future editions. Future editions should reference this table to identify what changed.

21. Limitations

The following limitations apply to every finding in this report. They are not qualifications that weaken specific conclusions; they are structural properties of the dataset that any reader should understand before drawing inferences.

1. Simulated, AI-piloted games only. All games are played by Forge's AI engine against the Gauntlet opponents. Results describe how the AI pilots these decklists — not how a human player would perform with the same cards. Differences in AI decision-making from human strategy may produce systematic biases in measured outcomes.

2. Correlational only. No construction breakdown, card-level figure, category finding, or color analysis in this report establishes a causal relationship. The dataset is observational and self-selected.

3. Non-random, self-selected sample. Decks are player-submitted, not randomly drawn from any population of Magic decks. The sample reflects the interests and play styles of Grim.Cards users during this specific period. Findings do not generalize to the broader Commander or Standard populations.

4. Power Score is an internal composite. Power Score is an internal Grim.Cards metric, not an objective external measure. Its use in Section 16 and Section 18 is descriptive — it describes how the internal composite correlates with simulated outcomes, not how "good" any deck is in an absolute sense.

5. Decision-impact is a play-quality proxy, not a win metric. Counterfactual decision-impact figures measure how the engine's chosen line compared to its own next-best alternative at recorded decision points. They are not damage counts, kill counts, or win contributions. They reflect AI decision behavior at specific moments.

6. Board-impact is a board-state proxy, not a win metric. Card performance figures measure board-quality change around turns a card is seen — not a causal contribution to winning. Cards that deliberately worsen the immediate board state (wipes, sacrifice effects) will mechanically register negative scores regardless of their strategic value.

7. Functional category classification is heuristic. Tutor, sacrifice, discard, and reanimation category assignments are derived from oracle-text keyword matching and pre-existing card flags. Edge cases, non-standard phrasings, and hybrid effects may be misclassified.

8. Standard cohort is small. With 35 decks and 633 games, the Standard sample is one-sixth the size of the Commander sample. Standard figures are real and reported but carry substantially wider uncertainty.

9. Monthly trends reflect deck-mix changes, not deck improvement. Different decks enter the dataset each month. Monthly win-rate changes describe who submitted that month, not longitudinal progress of any individual deck or player.

10. Instant/sorcery split note. The type-mix figures for instants and sorceries separately are sourced from the split-type fields available for all 218 Commander decks and all 35 Standard decks. The combined instant/sorcery figure and the separate instant and sorcery figures may not sum precisely due to unclassifiable spells that exist only in the combined bucket.

22. License & Citation

License

This dataset and report are published under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

You are free to share and adapt the material for any purpose, including commercially, provided you give appropriate credit, provide a link to the license, and indicate if changes were made.

Full license text: https://creativecommons.org/licenses/by/4.0/

How to Cite This Edition

Grim.Cards. Grim.Cards Simulation Study — Edition 2026-07-08 (Dataset version 2.4). Grim.Cards, 2026-07-08. Available at: https://grim.cards/case-study/2026-07-08. License: CC BY 4.0.

BibTeX:

@misc{grimcards2026casestudy,
  author       = {{Grim.Cards}},
  title        = {Grim.Cards Simulation Study --- Edition 2026-07-08},
  year         = {2026},
  month        = {07},
  note         = {Dataset version 2.4. Data window: 2026-05-14 to 2026-07-08.
                  Cohort: real, human-submitted Commander and Standard decks;
                  automated Crucible and sample decks excluded.
                  License: CC BY 4.0.},
  url          = {https://grim.cards/case-study/2026-07-08},
  howpublished = {\url{https://grim.cards/case-study/2026-07-08}}
}

Provenance Summary

Machine-Readable Downloads

The following artifacts accompany this edition and are available as committed static files:

The HTML artifact embeds a valid schema.org Dataset and Report JSON-LD node (including DataDownload distribution entries, variableMeasured list, temporalCoverage, datePublished, creator, publisher, license, and an atomic statements array of qualified claims suitable for knowledge-graph ingestion). All figures in the JSON-LD are identical to the figures in this prose report; the two are generated from the same computed dataset and cannot disagree.

This edition was generated from production data on 2026-07-08. To generate a new edition, run scripts/generate-case-study.js against the production read replica. Prior editions are retained at their permanent dated URLs and in dated artifact files; re-running never overwrites them. See case-studies/README.md for full instructions.