Models propose · trusted gates admit · receipts decide
Proving Grounds
A model-coached deck league built to show its work.
Twenty deck candidates enter a seeded round robin. Qwen coaches can rank and tune candidates, but trusted code owns legality, scheduling, evaluation, and every replayable receipt. This first run is visibly marked sample / proxy. Below it, pinned Card-Forge has now executed 45 of 45 real-rules Brawl matches, each reproduced twice — the full round robin. Those receipts are scored separately and are never mixed into the synthetic standings. This sample stays sample permanently, by design: its twenty fixture participants aren’t the ten real decks Forge plays, so no volume of real receipts could ever honestly cover this specific standings table — relabeling it would be exactly the fabrication this league refuses to do. The real, receipted evidence lives in the sections below instead.
Seats
20
Matches
190
Coaches
3
Evidence
proxy
Round robin
League standings
What this run proves
The scheduler, seeds, deck hashes, standings, model identities, and replay receipts are real and repeatable. This table's outcomes remain a feature-proxy fixture—not Magic rules execution and not an Arena win-rate forecast. The first two-deck rules-engine receipt appears below, but it does not promote these synthetic standings. A real leaderboard stays locked until its own league receipts complete.
First rules-engine proof
Ugin defeats Light-Paws, 2–1
Card-Forge executed the same seeded best-of-three twice. Both runs reproduced the same winners and turn counts with exit code 0 and no tracked import or game exceptions.
Game 1
Light-Paws
turn 6
Game 2
Ugin
turn 11
Game 3
Ugin
turn 11
Trust receipt
- Evidence
- rules engine verified
- Repeat
- 2 / 2 exact
- Score
- 2–1
- Compatible decks
- 10/10
- Seed
- 20260726
Rules-engine standings · complete round robin
45 of 45 queued matches executed
Real Card-Forge games between pinned Competitive Brawl lists, each played twice on a fixed seed and promoted only when both runs normalized identically. This table is scored from those receipts alone and never mixes with the synthetic standings above. Every deck has now played all 9 of its opponents.
Promoted
45
Held
0
Unattributed
0
Decks ranked
10
All 45 planned matches have promoted receipts and every deck played every other once. This is still one seeded best-of-three per pairing under a pinned engine — a single experiment, not a metagame win rate. Only completed, zero-error, hash-checked rules-engine receipts may enter standings. Every score in this shard links to a receipt drill-down (click the Match W–L column) with the full six-seed combined history.
Six-seed combined · receipted
270 of 270 matches receipted across 6 seeds
The shard above is one seeded round robin. This sums every deck’s receipted games across all 6 independently-seeded round robins run so far (20260727, 20270801, 20270901, 20271001, 20271101, 20271201) — 270 real Card-Forge matches, each reproduced twice before counting. Coverage is even and complete, so this table is the only one on this page whose status reads receiptedrather than provisional or sample. Click any score to open that deck’s full receipt drill-down.
Forge is not MTG Arena and may implement cards or AI differently. This proves reproducible pinned-engine matchups repeated across six seeds, not a live metagame win rate. Deck AI quality, not just deck strength, influences these results. Only the ten Forge-compatible admitted decks are exercised here; the synthetic 20-deck sample table above uses an unrelated fixture roster and is never mixed with this data.
Why these numbers are different
Three kinds of evidence on this page, never mixed
- Engine-receipted (the standings above)
- Real Card-Forge rules-engine matches we ran ourselves, twice each, on a pinned engine jar and pinned decklists. Every score traces to a receipt hash you can open — that’s what the drill-down links go to. Small, honest sample; zero guessing.
- Crowd telemetry (what we don’t have)
- Sites like untapped.gg compute win rates from aggregated real-player match telemetry at a scale we can’t reproduce — thousands of live Arena games a day. We have zero player-match telemetry and won’t publish a win-rate number at that scale we can’t back with a receipt, so none of the numbers above claim to be that.
- Externals (cite-only)
- untapped.gg, MTGGoldfish, EDHREC, and similar trackers are linked by name wherever they have real telemetry we don’t — never scraped, never re-hosted as our own numbers. See Us vs. untapped.gg for the full comparison.
Reproducibility · 6 seed sets
24 of 45 pairings gave the same winner every time
The standings above say what was played. This says what reproduced. The same 45 pairings were replayed under 6 different seed sets — 270 promoted matches in all — and a pairing counts as reproducible only when every seed set credited the same deck. Split pairings are the interesting result, not a defect, so they are named below with the shape of the disagreement.
Agreement
53%
Reproducible
24
Seed-sensitive
21
Coin flips
3
Seed-sensitive pairings — the same two decks, different winners
shapes: 6 ×24 · 5-1 ×10 · 4-2 ×8 · 3-3 ×3
- 3-3 blue control? 3 · Dragonclaw Ranked Brawl 3
- 3-3 Golos Comp. Brawl 3 · If You're Not First You're Last 3
- 3-3 Miiriym Brawl 3 · Nicol Bolas, Dragon-God — BrawlREC 3
- 4-2 If You're Not First You're Last 4 · Dragonclaw Ranked Brawl 2
- 4-2 Nicol Bolas, Dragon-God — BrawlREC 4 · Dragonclaw Ranked Brawl 2
- 4-2 Ugin, Eye of the Storms 4 · blue control? 2
- 4-2 blue control? 4 · If You're Not First You're Last 2
- 4-2 Golos Comp. Brawl 4 · Comp. Brawl Breach 2
- 4-2 blue control? 4 · Miiriym Brawl 2
- 4-2 Miiriym Brawl 4 · Shiko Blink Brawl 2
- 4-2 Tsagan Competitive Brawl 4 · Miiriym Brawl 2
- 5-1 Miiriym Brawl 5 · Dragonclaw Ranked Brawl 1
- 5-1 Shiko Blink Brawl 5 · Dragonclaw Ranked Brawl 1
- 5-1 Ugin, Eye of the Storms 5 · If You're Not First You're Last 1
- 5-1 Miiriym Brawl 5 · Ugin, Eye of the Storms 1
- 5-1 Shiko Blink Brawl 5 · Ugin, Eye of the Storms 1
- 5-1 Ugin, Eye of the Storms 5 · Nicol Bolas, Dragon-God — BrawlREC 1
- 5-1 Nicol Bolas, Dragon-God — BrawlREC 5 · If You're Not First You're Last 1
- 5-1 Shiko Blink Brawl 5 · blue control? 1
- 5-1 Tsagan Competitive Brawl 5 · blue control? 1
- 5-1 Nicol Bolas, Dragon-God — BrawlREC 5 · blue control? 1
6 seed sets is a tiny sample. It supports 'this pairing is seed-sensitive' and never 'this deck is stronger in the metagame'. Agreement measures reproducibility under a pinned engine and pinned decks, not deck strength. Forge is not MTG Arena; deck AI quality is part of every result.
Planar Cube · Sealed rules-engine proof
45 of 45 Cube matches promoted
Ten two-color shells built from the live Cube pool played one complete Forge Sealed round robin. Every match ran twice with the same normalized result before promotion. This verifies the export, card compatibility, and match pipeline; it does not simulate drafting decisions.
Promoted
45
Held
0
Unattributed
0
Decks
10
One seeded Forge round robin is an engine experiment, not an Arena Cube tier list. Deck AI quality and the fixed guild-shell construction both influence these records.
Six-seed combined · Cube Sealed · receipted
270 of 270 Cube matches receipted across 6 seeds
The Cube shard above is one seeded Sealed round robin. This sums every guild shell’s receipted games across all 6 independently-seeded round robins run so far (20260730, 20260910, 20261765, 20261810, 20261855, 20261900) — 270 real Card-Forge Sealed matches, each reproduced twice before counting. Coverage is even and complete, so this table reads receiptedrather than provisional or sample. Click any score to open that shell’s full receipt drill-down.
Forge is not MTG Arena and may implement cards or AI differently. This proves reproducible pinned-engine Sealed matchups repeated across six seeds, not a live metagame win rate. Deck AI quality, not just deck strength, influences these results. Only the ten Forge-compatible admitted Cube guild shells are exercised here; the synthetic 20-deck sample table on this page uses an unrelated fixture roster and is never mixed with this data.
Live source intake
Brawl admission queue
These are real decklists from Archidekt's Competitive Brawl format—not the synthetic league above. Structure and frozen Oracle legality are checked; all ten verified lists resolve in the pinned Forge archive, and 10 have now passed repeatable game execution.
Sourced
18
Structure valid
12
Structure rejected
6
Legality verified
10
Archidekt Competitive Brawl deckFormat=26
Archidekt Competitive Brawl deckFormat=26
Match-proven decks 10/18. Frozen Oracle audit dc7ca6bb0221 verified 10 real Competitive Brawl lists; 8 are rejected or held. Forge resolves 10/10 verified lists without substitutions. Only decks appearing in a promoted receipt have rules-engine evidence; no proxy result can promote the rest.
Read the full MTG Meta Hub methodology & attribution
Magic: The Gathering is © Wizards of the Coast. BuildKit is unofficial Fan Content permitted under the Wizards Fan Content Policy. Not approved or endorsed by Wizards.