ReverseMounts – Wynncraft Mount Breeding Algorithm

This repository contains a reverse-engineered statistical model of the Wynncraft mount (horse/wyvern) breeding system. It includes the raw breeding data, a validation test suite, and a Monte Carlo simulator that can predict offspring outcomes from parent stats.

---

Table of Contents

• Background
• Dataset
• Reverse-Engineered Algorithm
• Repository Files
• Running the Test Suite
• Using the Simulator
  • Compare Mode
  • Predict Mode
• Key Findings
• Model Limitations

---

Background

In Wynncraft, mounts (horses, wyverns, adasaurs) have 8 stats. Each stat has three values:

Field	Meaning
val	Current trained level
lim	Feeding cap (you can't feed past this)
max	Absolute ceiling (can only be raised by breeding again)

When you breed two mounts, both parents are consumed and you get one offspring. The game uses RNG, so the exact outcome is probabilistic—but the general formulas are discoverable from enough data.

This repo analyzes 35 real breeding submissions (horses and wyverns) and extracts the underlying mechanics.

---

Dataset

• File: wynnbreeder_export (1).json
• Submissions: 39 total (32 complete with offspring, 7 pending)
• Mount types: Horse, Wyvern

Each submission is a JSON object with three mounts:

{
  "id": 1,
  "status": "complete",
  "mounts": [
    { "role": "offspring", "type": "horse", "potential": 425, ... },
    { "role": "parent_a",  "type": "horse", "potential": 362, ... },
    { "role": "parent_b",  "type": "horse", "potential": 359, ... }
  ]
}

Each mount has these fields:

potential, color, name,
energy_value, energy_max,
speed_val, speed_lim, speed_max,
accel_val, accel_lim, accel_max,
altitude_val, altitude_lim, altitude_max,
energy_stat_val, energy_stat_lim, energy_stat_max,
handling_val, handling_lim, handling_max,
toughness_val, toughness_lim, toughness_max,
boost_val, boost_lim, boost_max,
training_val, training_lim, training_max

---

Reverse-Engineered Algorithm

After analyzing all submissions, these are the inferred formulas:

1. Potential (Hard Rule)

potential = Σ stat_max   (sum of all 8 stat max values)

Verified 100% across the entire dataset. No exceptions.

2. Offspring val (Current Level)

val = max(parent_a.val, parent_b.val) + random_variance

• Typical variance: +0 to +6
• Extreme range: -13 to +6
• When parents are identical, offspring val can still exceed parent val by 1–6 points.
• Negative variance happens when parents have very different levels or very high stats.

3. Offspring lim (Level Limit)

lim = avg(parent_a.lim, parent_b.lim) + random_bonus

• Typical bonus: +5 to +15
• Extreme range: -1 to +20
• Parent max values do NOT affect offspring limit. Only parent limits matter.

4. Offspring max (Absolute Ceiling)

max = avg(parent_a.lim, parent_b.lim) + random_bonus
max = max(max, lim)   // max can never be below lim

• Typical bonus: +15 to +45
• Extreme range: +5 to +69
• Again, parent maxes don't transfer—only parent limits drive the offspring max.
• Rarely, max == lim for all stats (observed in 3 low-potential breeds from identical fresh parents).

5. Energy

energy_max   = max(parent_a.energy_max, parent_b.energy_max) + [0, +5]
energy_value ≈ avg(parent_a.energy_value, parent_b.energy_value) ± variance

6. Color

Offspring colors are usually a combination of parent colors (e.g. Bay-Dawn × Gray-Reddish → Bay-Reddish). There is a small chance for mutation producing a color not present on either parent.

7. Randomness

The exact mechanics involve per-stat RNG. Identical parents can produce wildly different offspring. Verified: four breeds from identical 30/30/30 horses produced potentials of 309, 329, 333, and 460.

---

Repository Files

File	Description
wynnbreeder_export (1).json	Raw breeding data (35 complete submissions)
breeding_model.py	Statistical model with prediction ranges and constraint checkers
test_breeding.py	pytest suite (229 tests) validating every submission
simulator.py	Monte Carlo simulator with compare and predict modes
README.md	This file
.gitignore	Excludes venv/ and __pycache__/

---

Running the Test Suite

# Create virtual environment (first time only)
python -m venv venv
source venv/bin/activate

# Install pytest
pip install pytest

# Run all tests
python -m pytest test_breeding.py -v

Test Categories

• Hard constraints (must pass for 100% of data):
  • potential == sum(max_values)
  • max >= lim for all stats
  • 0 <= val <= max for all stats
• Soft constraints (statistical ranges, derived from observed data):
  • val within predicted range
  • lim within predicted range
  • max within predicted range
  • energy fields within predicted range
• Edge-case tests:
  • Identical parents produce different offspring
  • Fresh mounts have potential 240
  • Parent max does not directly transfer
  • Color inheritance heuristic

Current status: 229/229 tests passing.

---

Using the Simulator

The simulator (simulator.py) uses Gaussian distributions fitted to the observed data to generate offspring. It supports two modes:

Compare Mode

Validates the model against the real dataset. For each submission, it runs N simulations and reports whether the actual offspring falls inside the predicted min/max range.

python simulator.py compare --runs 1000

Example output summary:

============================================================
Overall stat hit rate : 825/832 = 99.2%
Potential hit rate    : 29/32 = 90.6%
============================================================

The misses are mostly the rare lim=max edge cases (submissions 8–10) where RNG produced abnormally low potential, plus a couple of energy outliers.

Predict Mode

Takes two parent stat blocks (as JSON) and runs thousands of Monte Carlo simulations to show predicted offspring ranges.

python simulator.py predict \
  --parent-a '{"speed_val":30,"speed_lim":30,"speed_max":30,"accel_val":30,"accel_lim":30,"accel_max":30,"altitude_val":30,"altitude_lim":30,"altitude_max":30,"energy_stat_val":30,"energy_stat_lim":30,"energy_stat_max":30,"handling_val":30,"handling_lim":30,"handling_max":30,"toughness_val":30,"toughness_lim":30,"toughness_max":30,"boost_val":30,"boost_lim":30,"boost_max":30,"training_val":30,"training_lim":30,"training_max":30,"energy_value":191,"energy_max":259}' \
  --parent-b '{"speed_val":30,"speed_lim":30,"speed_max":30,"accel_val":30,"accel_lim":30,"accel_max":30,"altitude_val":30,"altitude_lim":30,"altitude_max":30,"energy_stat_val":30,"energy_stat_lim":30,"energy_stat_max":30,"handling_val":30,"handling_lim":30,"handling_max":30,"toughness_val":30,"toughness_lim":30,"toughness_max":30,"boost_val":30,"boost_lim":30,"boost_max":30,"training_val":30,"training_lim":30,"training_max":30,"energy_value":191,"energy_max":259}' \
  --runs 10000

Example output:

Predicted offspring after 10000 simulations:
Stat            Min    Max   Mean     SD
----------------------------------------
potential       370    583    472   29.0
speed_val        24     36     30    1.7
speed_lim        29     50     39    3.3
speed_max        35     98     59   10.4
...
energy_max      259    269    261    1.8
energy_value     97    211    183   18.7

This tells you that breeding two identical fresh 30/30/30 horses will most likely produce an offspring with:

• Potential: 370–583 (mean ~472)
• Per-stat lim: ~39 (range 29–50)
• Per-stat max: ~59 (range 35–98)

---

Key Findings

1. Potential is deterministic: sum of all 8 max values. Always.
2. Parent maxes don't matter: Only parent limits drive offspring limits and maxes.
3. Feed both parents equally: The average of parent limits is what counts. A high-limit parent bred with a low-limit parent will drag the average down.
4. Train one parent, not both: Offspring val is based on the higher parent's val, so you only need to train one mount.
5. RNG is significant: Even with perfect identical parents, you can get a 309-potential dud or a 460-potential stud.
6. lim=max edge case exists: Rarely, offspring max equals lim for all stats, creating a "dead-end" breed that can't grow further through feeding.

---

Model Limitations

• The exact RNG distributions are approximations based on 35 submissions (~280 stat observations). More data would tighten the ranges.
• Color mutation mechanics are not fully modeled—only a heuristic inheritance check is implemented.
• The simulator does not account for mount type differences (horse vs wyvern vs adasaur) because the dataset only contains horses and wyverns, and they appear to follow the same formulas.
• Adasaur mounts are not represented in the dataset at all.