Recorded synthetic players suggest where to draft onboarding aids. Paired simulations measure tuning changes, not whether the text teaches.
Round 1 only, 200 personas; the quit/frustration model is NOT re-simulated; an aid that raises casual win rate may still not teach anything; real new players are not involved. One seed, not replicated: differences of a few wins out of 200 are within noise.
Win rate is wins per recorded round. Quit rate is sessions ending from synthetic frustration or boredom per player. Round 2 duration is the median among players who reached it.
Measured proposals
Recorded offline model proposal. Ranked by casual win-rate change. Proposed text and rationale are not verified as teaching effective skills.
Master Your First Combo
Player-facing text: Tap J four times quickly: jab → cross → hook → uppercut.
Proposed rationale (casual): Casual players struggle with combo execution (only 5.9% win rate). Lowering jab startup and recovery makes chaining easier while increasing damage and hit stop improves feedback on successful hits, encouraging practice of the basic 4-punch chain described in controls.
Overrides: anim.JAB.damage: 15, anim.JAB.startup: 0.05, anim.JAB.active: 0.12, anim.JAB.recovery: 0.22, anim.JAB.hitStop: 4, anim.JAB.reach: 1.2, anim.JAB.drive: 1.5
200 personas, 1 round each, seed 1; wall time 12.86 s. Bug rounds: 1 → 0.
| Archetype | Before | After | Change |
|---|
| casual | 12.8% (5/39) | 23.1% (9/39) | 10.3 percentage points |
|---|
| cheeser | 91.3% (21/23) | 91.3% (21/23) | 0.0 percentage points |
|---|
| explorer | 53.8% (7/13) | 46.2% (6/13) | -7.7 percentage points |
|---|
| learner | 8.7% (2/23) | 8.7% (2/23) | 0.0 percentage points |
|---|
| masher | 25.0% (9/36) | 30.6% (11/36) | 5.6 percentage points |
|---|
| rusher | 42.4% (14/33) | 48.5% (16/33) | 6.1 percentage points |
|---|
| speedrunner | 70.0% (7/10) | 70.0% (7/10) | 0.0 percentage points |
|---|
| turtle | 4.3% (1/23) | 13.0% (3/23) | 8.7 percentage points |
|---|
Replication (seeds 1–8)
Honest caveat: still round 1 only, quit model not re-simulated, no real new players.
| Mean paired difference | 95% Interval (t-based) | Seeds (improved / tied / worsened) | Verdict |
|---|
| +20.50 wins | [+15.35, +25.65] | +8 / =0 / -0 (8 improved, 0 tied, 0 worsened) | clear gain |
Practice Sidestepping to Create Openings
Player-facing text: Press Q or V to sidestep deeper or nearer to the screen and find your opening.
Proposed rationale (casual): Casual players have low round2Count (159) suggesting few survive past first round. Increasing jab reach slightly helps compensate for positioning errors when learning sidesteps, making it easier to hit evasive rivals and encouraging use of the Q/V sidestep mechanic mentioned in controls.
Overrides: anim.JAB.reach: 1.3
200 personas, 1 round each, seed 1; wall time 12.94 s. Bug rounds: 1 → 1.
| Archetype | Before | After | Change |
|---|
| casual | 12.8% (5/39) | 20.5% (8/39) | 7.7 percentage points |
|---|
| cheeser | 91.3% (21/23) | 91.3% (21/23) | 0.0 percentage points |
|---|
| explorer | 53.8% (7/13) | 38.5% (5/13) | -15.4 percentage points |
|---|
| learner | 8.7% (2/23) | 13.0% (3/23) | 4.3 percentage points |
|---|
| masher | 25.0% (9/36) | 25.0% (9/36) | 0.0 percentage points |
|---|
| rusher | 42.4% (14/33) | 39.4% (13/33) | -3.0 percentage points |
|---|
| speedrunner | 70.0% (7/10) | 80.0% (8/10) | 10.0 percentage points |
|---|
| turtle | 4.3% (1/23) | 8.7% (2/23) | 4.3 percentage points |
|---|
Replication (seeds 1–8)
Honest caveat: still round 1 only, quit model not re-simulated, no real new players.
| Mean paired difference | 95% Interval (t-based) | Seeds (improved / tied / worsened) | Verdict |
|---|
| +3.13 wins | [+0.15, +6.10] | +6 / =0 / -2 (6 improved, 0 tied, 2 worsened) | not distinguishable from noise |
Learn to Guard Effectively
Player-facing text: Hold away from your rival to guard and reduce incoming damage by 80%.
Proposed rationale (casual): Casual players have a 97% quit rate, often due to feeling overwhelmed. Reducing guard damage multiplier further (from 0.2 to 0.15) increases protection during blocking, giving new players more safety to learn timing without excessive punishment, addressing their low win rate.
Overrides: rules.guardDamageMultiplier: 0.15
200 personas, 1 round each, seed 1; wall time 13.04 s. Bug rounds: 1 → 1.
| Archetype | Before | After | Change |
|---|
| casual | 12.8% (5/39) | 12.8% (5/39) | 0.0 percentage points |
|---|
| cheeser | 91.3% (21/23) | 91.3% (21/23) | 0.0 percentage points |
|---|
| explorer | 53.8% (7/13) | 53.8% (7/13) | 0.0 percentage points |
|---|
| learner | 8.7% (2/23) | 8.7% (2/23) | 0.0 percentage points |
|---|
| masher | 25.0% (9/36) | 19.4% (7/36) | -5.6 percentage points |
|---|
| rusher | 42.4% (14/33) | 33.3% (11/33) | -9.1 percentage points |
|---|
| speedrunner | 70.0% (7/10) | 60.0% (6/10) | -10.0 percentage points |
|---|
| turtle | 4.3% (1/23) | 4.3% (1/23) | 0.0 percentage points |
|---|
Replication (seeds 1–8)
Honest caveat: still round 1 only, quit model not re-simulated, no real new players.
| Mean paired difference | 95% Interval (t-based) | Seeds (improved / tied / worsened) | Verdict |
|---|
| -2.13 wins | [-3.70, -0.55] | +0 / =1 / -7 (0 improved, 1 tied, 7 worsened) | clear loss |
Replication across seeds 1–8
Replication across 8 seeds (wall time: 359.22 s).
| Aid | Mean paired difference | 95% Interval (t-based) | Seeds (improved / tied / worsened) | Verdict |
|---|
| Master Your First Combo | +20.50 wins | [+15.35, +25.65] | +8 / =0 / -0 | clear gain |
|---|
| Practice Sidestepping to Create Openings | +3.13 wins | [+0.15, +6.10] | +6 / =0 / -2 | not distinguishable from noise |
|---|
| Learn to Guard Effectively | -2.13 wins | [-3.70, -0.55] | +0 / =1 / -7 | clear loss |
|---|
Honest caveat: still round 1 only, quit model not re-simulated, no real new players.