A Thousand Players
Crowd run · Market Brawl
Recorded run · 1000 players, 2,831 rounds, Ultra report

A crowd of synthetic players breaks your game so you can fix it.

NVIDIA Nemotron models write the players, Market Brawl runs their matches headless, and Nemotron Ultra reads a summary of all 2,831 rounds in one call. Every finding below comes with a seed you can replay in the real game.

See the 4 bugs and their proofTry a live model call

1,000
players
2,831
rounds played
9.6
simulated hours
71%
quit (bored or frustrated)

The Crowd

Recorded rounds replaying side by side. Hover for the player, click to open the replay.

synthetic playerCPU rivalhit
List of all 64 rounds (text version)
  1. Aaliyah casual, r1, loss
  2. Cara cheeser, r1, win
  3. Ravi learner, r1, win
  4. Della turtle, r1, loss
  5. Kenji masher, r1, loss
  6. Finn rusher, r1, win
  7. Maya explorer, r1, win
  8. Pia speedrunner, r1, win
  9. Chloe casual, r1, loss
  10. Jude cheeser, r1, win
  11. Riley learner, r1, loss
  12. Tomasz turtle, r1, win
  13. Leila masher, r1, win
  14. Ravi rusher, r1, loss
  15. Rhea explorer, r1, win
  16. Mateo speedrunner, r1, loss
  17. Derek casual, r1, loss
  18. Sofia cheeser, r1, win
  19. Ben learner, r1, loss
  20. Ivan turtle, r1, loss
  21. Clara masher, r1, loss
  22. Lena rusher, r1, win
  23. Zara explorer, r1, loss
  24. Zara speedrunner, r1, loss
  25. Tom casual, r1, loss
  26. Floyd cheeser, r1, win
  27. Case learner, r1, loss
  28. Otto turtle, r1, loss
  29. Samir masher, r1, loss
  30. Ida rusher, r1, win
  31. Hana explorer, r1, loss
  32. Yara speedrunner, r1, win
  33. Nico casual, r1, loss
  34. Nia cheeser, r1, win
  35. Dana learner, r1, loss
  36. Ida turtle, r1, loss
  37. Harper masher, r1, win
  38. Dev rusher, r1, win
  39. Maya explorer, r1, win
  40. Arjun speedrunner, r1, win
  41. Zoe casual, r1, win
  42. Omar cheeser, r1, win
  43. Kian learner, r1, loss
  44. Zoe turtle, r1, loss
  45. Tara masher, r1, loss
  46. NurseUHI rusher, r1, loss
  47. Mark explorer, r1, loss
  48. Tess speedrunner, r1, win
  49. Sofia casual, r1, loss
  50. Lina cheeser, r1, win
  51. Mona learner, r1, loss
  52. Eunice turtle, r1, loss
  53. Gladys masher, r1, loss
  54. Helen rusher, r1, loss
  55. Steve explorer, r1, win
  56. Rhea speedrunner, r1, loss
  57. Nelly casual, r1, loss
  58. Betty cheeser, r1, win
  59. Sam learner, r1, loss
  60. Dorothy turtle, r1, loss
  61. Olivia masher, r1, loss
  62. Sam rusher, r1, win
  63. Dave explorer, r1, loss
  64. Eli speedrunner, r1, win

The evidence

Aggregates computed by code from every round, not by a model.

Where knockouts happen

0s5s10s15s20s26sleft wallcenterright wallx where the loser fell (player starts left, rival right)
Where and when rounds end in a knockout. Rows are round time, columns are arena position.
Data table
t from-10.5-8.8-7.0-5.3-3.5-1.80.01.83.55.37.08.8
0s000013000010
5s23244194423219165662752
10s1429417710424520519415376208
15s9822193858778172562724
20s511694810513910
26s411100001222

Who keeps playing

100%round 192%round 257%round 325%round 49%round 5
Share of players who started each round. Session ends: frustration 708, cap 42, satisfied 250.

Archetypes: win and quit

cheeser92%6%speedrunner65%28%rusher42%72%explorer32%60%masher27%83%learner15%86%turtle11%78%casual6%97%win ratequit rate (bored or frustrated)
Each archetype is a family of synthetic personas with its own decision table.
Data table
ArchetypePlayersRoundsWinQuitAvg s
cheeser10029992%6%17.4
speedrunner6018065%28%9.8
rusher16042442%72%10.2
explorer8025332%60%12.1
masher16040427%83%11.3
learner12036115%86%11.1
turtle12047011%78%14.4
casual2004406%97%10.9

How long rounds last

02294370660585751044212328142331624918+
Rounds by length in seconds. The last bar also counts every longer round, including timeouts.

What Ultra found

nvidia/Nemotron-3-Ultra-550b-a55b · read a summary of 2,831 rounds (220,334 prompt tokens) in one call

Ultra's summary, prose not checkedThe game has a severe difficulty cliff: casual, learner, and turtle players quit at 78-97% rates and win only 6-15% of rounds, while cheesers exploit a dominant SWEEP/GROUND strategy to win 92% with 6% quit. Round 2 is a massive frustration wall (31% overall quit rate, 50-75% for weak archetypes). A physics bug (!airborne) appears on turtle and explorer runs, and masher/learner show zero progression — win rates flatline across rounds. (Verification: 0 of 31 claims did not match the recomputed value; 13 cited seeds did not exist in the runs and were removed.)

Green claims were recomputed from the raw runs by code and match within tolerance. Amber claims were not checked or did not match.

highdifficultyUltra states 95% confidence

Casual/learner/turtle hit a difficulty wall at Round 2 and quit en masse

Round 2 quit rates are catastrophic for weak archetypes: casual 59%, learner 27%, turtle 5%, masher 50%, rusher 35%. Overall 31% of players quit right after Round 2. These archetypes also have win rates of 4-8% in Round 2 and 0-5% in Round 3, meaning they cannot progress. The session end reasons confirm this: 708 frustration quits vs only 250 satisfied.

Numbers checked by code (the title and prose are Ultra's and are not checked)
  • quit rate casual, round 2: 59% (51-66%), verified by code
  • quit rate learner, round 2: 27% (20-36%), verified by code
  • quit rate masher, round 2: 50% (42-58%), verified by code
  • win rate casual, round 2: 4% (1-8%), verified by code
Replay the evidence
▶ seed 792093 p0001 r2▶ seed 988036 p0033 r2▶ seed 918544 p0045 r1▶ seed 198252 p0051 r2
Suggested fix

Add a difficulty ramp: reduce rival aggression/HP in Rounds 1-2, or give players a defensive tool (parry/dodge) that works at low skill.

Related reading (Tavily web search results, not reviewed)
  • Core Mechanics of Fighting Games - DEV CommunityThe most common way of reducing HP is to connect an attack onto the opponent. The Player performs inputs which cause the Character to attack. If the Character's attack fails to connect with the opponent's Character, it is a miss and no HP is lost. If it does connect, the opposing Character loses som
  • Debating Difficulty curves in game design - Game WisdomIf you fail to understand the skill curve and pain points of your game and try to add in difficulty settings, what ends up happening is that any problems with your core gameplay loop will be harder to find on the lower settings and made more apparent on the higher ones. Not understanding why people
  • Difficulty curves: how to get the right balance1. Add difficulty to your game: either reduce the availability of resources or add complexity. You can reduce resources by raising the costs of in-game actions, making costs more frequent, or giving players fewer ways to gather resources. To increase complexity, you can add more enemies, give player
highexploitUltra states 98% confidence

Cheeser archetype exploits SWEEP/GROUND spam for 92% win rate and near-zero frustration

Cheesers (100 players, 299 rounds) win 92% overall with only 6% quit rate and 7% koShare. They land SWEEP (3538 total events) and GROUND (2526) at massive volume while taking minimal damage (meanHpLoss 48.1 vs 85.9 overall). Their strategy is repeatable: SWEEP x3-6 + GROUND x1-4 per round, often winning in 15-20s. No other archetype approaches this win rate.

Numbers checked by code (the title and prose are Ultra's and are not checked)
  • win rate cheeser: 92% (89-95%), verified by code
  • quit rate cheeser: 6% (3-12%), verified by code
  • mean hp loss cheeser: 48.1, verified by code
  • KO share cheeser: 7% (5-10%), verified by code
Replay the evidence
▶ seed 458977 p0002 r1▶ seed 150359 p0004 r1▶ seed 966421 p0006 r1
Suggested fix

Nerf SWEEP/GROUND frame advantage or add rival counter-play (armor/parry vs low attacks). Require move variety for full damage.

Related reading (Tavily web search results, not reviewed)
highbalanceUltra states 90% confidence

Masher and learner archetypes show zero learning — win rates flatline across all rounds

Masher win rates: R1 26%, R2 24%, R3 32%, R4 29%, R5 50% (only 5 rounds). Learner: R1 10%, R2 15%, R3 14%, R4 27%, R5 67% (but only 3 players reach R5). Both have 73-85% koShare and 83-86% quit rates. They land the same moves (JAB, ROUNDHOUSE_KICK) round after round with no adaptation. The 'learner' persona does not learn.

Numbers checked by code (the title and prose are Ultra's and are not checked)
  • win rate masher, round 1: 26% (19-33%), verified by code
  • win rate masher, round 3: 32% (22-45%), verified by code
  • win rate learner, round 1: 10% (5-16%), verified by code
  • win rate learner, round 3: 14% (7-21%), verified by code
Replay the evidence
▶ seed 218315 p0008 r1
Suggested fix

Give masher/learner a visible progression system: unlock moves/combos after X wins, or add a practice mode with move tutorials.

Related reading (Tavily web search results, not reviewed)
  • The Procession of Progression in Game DesignProgression through a video game can be divided into two categories: player and game. Player progression is the player learning the rules and mechanics while developing their skills. As an example: someone learning how to make use of the portal based movement in Portal. While game progression refers
highbugUltra states 85% confidence

Physics bug: !airborne invariant violation on turtle and explorer players

Multiple turtle and explorer rounds show '!airborne@<time>' events, marking engine invariant violations. Examples: p0009 R3 (!airborne@16.6), p0114 R3 (!airborne@18.0), p0254 R1 (!airborne@11.1), p0745 R3 (!airborne@13.2), p0355 R2 (!airborne@16.6). These correlate with high frustration and KO losses. The bug likely causes unfair knockdowns or state desync.

Numbers checked by code (the title and prose are Ultra's and are not checked)
  • quit rate turtle, round 3: 25% (18-33%), verified by code
  • KO share turtle: 89% (85-93%), verified by code
  • quit rate explorer, round 3: 35% (23-48%), verified by code
Replay the evidence
▶ seed 822230 p0009 r3▶ seed 278329 p0355 r2
Suggested fix

Fix the airborne state invariant — likely a collision/grounding check that fails on certain move sequences. Add automated regression test for !airborne.

Related reading (Tavily web search results, not reviewed)
  • What Do Bug Fixes in Video Game Update Notes Tell Us?likely to have a severe negative impact on users . [...] These results can help game developers identify which types of bugs to pay more attention to when testing and fixing bugs. [...] Developers can also use these results to help adjust practices related to game development process in order to bett
mediumbalanceUltra states 80% confidence

Rusher archetype is the only non-cheeser with viable win rate (42%) but still quits 72%

Rushers (160 players) win 42% overall — second best after cheesers — but 72% still quit (mostly frustration). They win early (R1 44%, R2 38%) but quit spikes at R2 (35%) and R3 (46%). Their style (BODY_JAB, ROUNDHOUSE_KICK, GROUND) works but feels unrewarding. They need a satisfaction hook: visible progress, combo variety, or less HP loss on wins (meanHpLoss 82.9).

Numbers checked by code (the title and prose are Ultra's and are not checked)
  • win rate rusher: 42% (36-47%), verified by code
  • quit rate rusher: 72% (64-78%), verified by code
  • quit rate rusher, round 2: 35% (27-43%), verified by code
  • mean hp loss rusher: 82.9, verified by code
Replay the evidence
▶ seed 365564 p0012 r1▶ seed 582506 p0054 r1▶ seed 320064 p0096 r1
Suggested fix

Add rusher-friendly rewards: meter gain on aggression, visible combo counter, or reduce HP loss on successful rushdowns.

mediumbalanceUltra states 30% confidence

Speedrunner and explorer have polarized outcomes — either dominate or quit fast

Speedrunners (60 players) win 65% but 28% quit; explorers (80 players) win 32% with 60% quit. Both have low avgDuration (9.8s, 12.1s) meaning rounds end fast — either quick KOs or quick deaths. Speedrunners use GROUND/ROUNDHOUSE_KICK/PUNCH combos; explorers mix JAB/SWEEP/ROUNDHOUSE_KICK. The variance suggests matchup RNG or seed sensitivity, not skill expression.

Numbers checked by code (the title and prose are Ultra's and are not checked)
  • win rate speedrunner: 65% (57-73%), verified by code
  • quit rate speedrunner: 28% (19-41%), verified by code
  • avg round speedrunner: 9.8 s, verified by code
  • win rate explorer: 32% (25-39%), verified by code
Replay the evidence
Suggested fix

Reduce round-ending volatility: increase HP pools, add comeback mechanics, or make KO require combo/confirmation not single big hit.

Related reading (Tavily web search results, not reviewed)
mediumbalanceUltra states 85% confidence

Turtle archetype is a trap — high patience but 11% win rate, 89% KO share, 78% quit

Turtles (120 players) have high patience (avg 0.8+) and frustration tolerance but get crushed: 11% win rate, 89% koShare, 96.6 meanHpLoss. They block/parry (blk 1-5, par 1-4 per round) but it doesn't convert to offense. They quit late (R3 25%, R4 52%, R5 46%) after prolonged suffering. The defensive playstyle is non-viable.

Numbers checked by code (the title and prose are Ultra's and are not checked)
  • win rate turtle: 11% (7-15%), verified by code
  • KO share turtle: 89% (85-93%), verified by code
  • quit rate turtle, round 4: 52% (42-63%), verified by code
  • mean hp loss turtle: 96.6, verified by code
Replay the evidence
▶ seed 469644 p0005 r1▶ seed 861899 p0007 r2▶ seed 239488 p0050 r3
Suggested fix

Give turtles a counter-attack mechanic: successful block/parry grants guaranteed punish window or meter. Make defense rewarding.

Related reading (Tavily web search results, not reviewed)
  • Street Fighter 6 Balance Patch: Key Character Changes and Their Impact – Fighter101Street Fighter 6 Balance Patch: Key Character Changes and Their Impact. While not every character received changes, significant adjustments were made for some, focusing on balancing offensive and defensive c
  • Patch Notes 2.53.1 – FOR HONORDev comment: Varangian had an issue where if you were close to a wall, a throw would wallsplat too late and Varangian was not able to get their correct Heavy Attack punish on the opponent. With this change, it should always be guaranteed. Jiang jun: Can now enter Sifu's Poise at the same timing a
  • Patch Notes v1.13Patch Notes v1.13. Fighters. **Developer Comments:** The goal here is to increase the risk of using Full Block Stance. It will also help people without Melee Attack put pressure on Warlord and Conqueror. **Developer Comments:** We wanted to give more incentive to use F
lowbalanceUltra states 70% confidence

Round 3+ survival bias inflates win rates — only strong players remain

Win rates appear to improve in later rounds (overall R4 40%, R5 45%) but this is survivorship bias: only 253 players reach R4, 92 reach R5. The weak archetypes (casual, learner, turtle) are nearly extinct by R4. Real win rates for the original population stay flat. This masks the difficulty cliff in aggregate metrics.

Numbers checked by code (the title and prose are Ultra's and are not checked)
  • win rate everyone, round 4: 40% (33-45%), verified by code
  • win rate everyone, round 5: 45% (35-54%), verified by code
  • quit rate everyone, round 3: 39% (35-43%), verified by code
  • quit rate everyone, round 4: 38% (33-44%), verified by code
Replay the evidence
▶ seed 272146 p0003 r5▶ seed 186153 p0037 r5▶ seed 451580 p0096 r5
Suggested fix

Report cohort-based win rates (of 1000 starters) not conditional-on-reaching-round. Add early-game telemetry to catch the cliff.

Who did what

A cheap swarm and one big reader. Each model has one job.

Nemotron Nano

nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B

Writes every persona's policy in bulk and proposes exploit strategies. Can also pick each action for a spotlight player (implemented, not in the main run).

calls
3
tokens in
1,849
tokens out
6,225
cost
not priced

Nemotron Super

nvidia/nemotron-3-super-120b-a12b

Turns English into things the engine runs: bug reports into oracles, patch notes into tuning.

calls
2
tokens in
2,579
tokens out
479
cost
not priced

Nemotron Super-VL

nvidia/Nemotron-3.5-Super-VL

Describes replay frames in the Eyes panel. A description aid: in a 6-round test it flagged 3 of 3 bug rounds and 0 of 3 clean ones, but the engine decides what is a bug.

Not part of the recorded run. It runs live in the app (the Eyes panel on a replay), and live calls are not logged here.

Nemotron Ultra

nvidia/Nemotron-3-Ultra-550b-a55b

Reads a summary of all 2,831 rounds in one long-context call (one of its 27 recorded calls) and writes the findings, citing seeds.

calls
27
tokens in
892,282
tokens out
16,225
cost estimate
~$0.94 (estimate)

Tavily

web search API

Finds related developer writeups: offline for each finding, and live from the Find precedent button on Repro.

Not part of the recorded run. It runs live in the app (Find precedent on Repro), and live calls are not logged here.

usdEstimate is filled only for Ultra, at an UNVERIFIED aggregator price of $1/M input and $3/M output tokens. Nano, Super and Super-VL report tokens only (usdEstimate 0) because no verified Token Factory price was available. Token counts come from the API responses. Prices are unverified, so no dollar figure is shown except where a row carries an estimate.

What this cannot tell you