Skip to main content
This page documents every parameter available in YC-Bench configuration files. All parameters are defined in default.toml, which serves as the canonical reference.
Location: src/yc_bench/config/presets/default.tomlEvery parameter is explicitly listed in the default preset so this file doubles as complete documentation.

​
Top-Level Fields

These fields identify and describe your experiment.
​
name
string
required
Experiment name, used in output filenames and logs.Default: "default"
​
description
string
Human-readable description of what this configuration tests.Default: "Hardened 3-year benchmark. Most tasks require prestige 3-5, 2-domain work, tight deadlines, costly cancellations."
​
extends
string
Name of another preset to inherit from. All presets except default use extends = "default" to inherit base parameters.Example: extends = "default"

​
Agent Configuration

The [agent] section configures the LLM agent behavior.
​
agent.model
string
required
LLM model identifier. Supports OpenRouter format.Default: "openrouter/z-ai/glm-5"Example: "openrouter/anthropic/claude-3.5-sonnet"
​
agent.temperature
float
Sampling temperature for model responses.Default: 0.0Range: 0.0 (deterministic) to 1.0 (creative)
​
agent.top_p
float
Nucleus sampling parameter.Default: 1.0Range: 0.0 to 1.0
​
agent.request_timeout_seconds
float
Timeout for each LLM API request.Default: 300.0 (5 minutes)
​
agent.retry_max_attempts
integer
Maximum retry attempts for failed API requests.Default: 3
​
agent.retry_backoff_seconds
float
Initial backoff delay between retries (exponential backoff).Default: 1.0
​
agent.history_keep_rounds
integer
Number of conversation rounds retained in context before each API call. Older rounds are dropped proactively. The scratchpad survives truncation.Default: 20
​
agent.system_prompt
string
Optional system prompt override. Omit to use the built-in prompt.Default: (built-in prompt used)

​
Loop Configuration

The [loop] section controls the agent interaction loop.
​
loop.auto_advance_after_turns
integer
Consecutive turns without sim resume before the loop forces a time-advance.Default: 10Purpose: Prevents agents from getting stuck in analysis loops without advancing simulation time.
​
loop.max_turns
integer
Hard cap on total turns. Comment out or set to a large number for unlimited.Default: (unlimited)Example: max_turns = 500

​
Simulation Parameters

The [sim] section defines the simulation timeline.
​
sim.start_date
string
required
Simulation start date in ISO 8601 format.Default: "2025-01-01"Format: YYYY-MM-DD
​
sim.horizon_years
integer
required
Simulation duration in years.Default: 3Tutorial/Easy/Medium/Hard/Nightmare: 1
​
sim.company_name
string
Name of the company in the simulation.Default: "BenchCo"

​
World Parameters

The [world] section contains scalar world-generation parameters.

​
Team & Resources

​
world.num_employees
integer
required
Number of employees in the company.Default: 10
​
world.initial_funds_cents
integer
required
Starting cash in cents.Default: 15_000_000 ($150,000)Tutorial: 25_000_000 ($250,000)Hard: 10_000_000 ($100,000)Nightmare: 8_000_000 ($80,000)
​
world.initial_prestige_level
float
Starting prestige level for all domains.Default: 1.0
​
world.work_hours_per_day
float
Business hours per day that employees work.Default: 9.0

​
Market

​
world.num_market_tasks
integer
Total number of tasks generated in the market.Default: 200
​
world.market_browse_default_limit
integer
Default number of tasks shown when browsing the market.Default: 50

​
Salary & Prestige Mechanics

​
world.salary_bump_pct
float
Salary increase percentage per completed task for each assigned employee. Compounds over time, accelerating payroll pressure.Default: 0.01 (1% per task)Tutorial: 0.0 (no compounding)Nightmare: 0.02 (2% per task - aggressive)
​
world.prestige_max
float
Maximum prestige level.Default: 10.0
​
world.prestige_min
float
Minimum prestige level.Default: 1.0
​
world.penalty_fail_multiplier
float
Prestige penalty multiplier when a task fails (misses deadline).Default: 1.4 (hardened)Tutorial: 0.3 (negligible)Nightmare: 2.0 (catastrophic)Interpretation: Higher values mean failing tasks costs MORE prestige. Canceling is worse than failing in default config.
​
world.penalty_cancel_multiplier
float
Prestige penalty multiplier when a task is canceled.Default: 2.0 (hardened)Tutorial: 0.5 (negligible)Nightmare: 2.5 (catastrophic)Interpretation: Accepting is a real commitment, not a free option.
​
world.reward_prestige_scale
float
Extra reward fraction per prestige level above 1.Default: 0.55Formula: reward_multiplier = 1 + scale × (prestige - 1)Example: At scale=0.55, a prestige-8 task pays 1 + 0.55×7 = 4.85× more than prestige-1.
​
world.prestige_decay_per_day
float
Daily prestige decay per domain. Domains not exercised lose prestige over time.Default: 0.005 (-0.15/month, ~1 level lost every 6 months)Purpose: Prevents single-domain hyper-specialization.
​
world.prestige_qty_scale
float
Required quantity scaling by prestige level.Default: 0.3Formula: qty_multiplier = 1 + scale × (prestige - 1)Example: At scale=0.3, prestige-5 tasks need 1 + 0.3×4 = 2.2× the work of prestige-1 tasks.

​
Deadlines

​
world.deadline_qty_per_day
float
Quantity units processed per business day when calculating deadlines.Default: 200.0Formula: deadline = max(deadline_min_biz_days, max_domain_qty / deadline_qty_per_day)Note: Domains are worked in parallel, so deadline scales with the heaviest domain, not the sum.
​
world.deadline_min_biz_days
integer
Minimum deadline in business days.Default: 7

​
Progress & Business Hours

​
world.task_progress_milestones
array
Progress fractions that trigger checkpoint events.Default: [0.25, 0.5, 0.75]
​
world.workday_start_hour
integer
Business day start hour (24-hour format).Default: 9
​
world.workday_end_hour
integer
Business day end hour (24-hour format).Default: 18

​
Distribution Types

YC-Bench uses statistical distributions to generate random quantities like task prestige, reward amounts, and work volumes. Each [world.dist.<name>] section specifies one distribution.

​
Available Distribution Types

Triangular distribution with a mode (peak). Most values cluster around the mode.Parameters:
  • type = "triangular"
  • low (float/int): Minimum value
  • high (float/int): Maximum value
  • mode (float/int): Most likely value (peak)
Example:
[world.dist.required_prestige]
type = "triangular"
low  = 1
high = 10
mode = 4    # Most tasks need prestige 3-5

​
World Distributions

These distributions define random task properties generated during world creation.

​
Required Prestige

​
world.dist.required_prestige
distribution
required
Prestige level required to accept a task (cast to int after sampling).Default (hardened):
type = "triangular"
low  = 1
high = 10
mode = 4    # Most tasks need prestige 3-5
Tutorial:
type = "constant"
value = 1    # ALL tasks accessible immediately

​
Reward Funds

​
world.dist.reward_funds_cents
distribution
required
Base reward paid on task completion, in cents. Scaled further by prestige.Default:
type = "triangular"
low  = 300_000      # $3,000
high = 4_000_000    # $40,000
mode = 1_400_000    # $14,000
Economics: Mode $14K means prestige-1 tasks burn cash, prestige-3 breaks even, prestige-4+ profits.

​
Domain Count

​
world.dist.domain_count
distribution
required
Number of domains each task requires work in (cast to int after sampling).Default (hardened):
type = "triangular"
low  = 1
high = 3
mode = 2    # Most tasks need 2 domains
Tutorial:
type = "constant"
value = 1    # Single-domain only

​
Required Quantity

​
world.dist.required_qty
distribution
required
Work units required per domain per task (cast to int after sampling).Default (hardened):
type = "triangular"
low  = 800          # No trivially small tasks
high = 4000
mode = 2000         # Substantial work required
Nightmare:
low  = 1200
high = 5000
mode = 2500         # Massive work volumes

​
Prestige Delta Reward

​
world.dist.reward_prestige_delta
distribution
required
Prestige delta awarded per domain on task success.Default:
type  = "beta"
alpha = 1.2
beta  = 2.8
scale = 0.35
low   = 0.0
high  = 0.35    # Mean ~0.1: climbing 1→5 takes ~40 tasks

​
Skill Boost

​
world.dist.skill_boost
distribution
required
Skill rate boost applied to each assigned employee on task success. Expressed as a fraction of current rate.Default:
type  = "normal"
mean  = 0.12    # Average 12% boost
stdev = 0.06
low   = 0.01
high  = 0.40

​
Salary Tiers

Three salary tier sections define employee distributions. All three share values must sum to exactly 1.0.

​
Junior Tier

​
world.salary_junior
object
required
Configuration for junior employees.Default:
[world.salary_junior]
name       = "junior"
share      = 0.50       # 50% of employees
min_cents  = 200_000    # $2,000/month
max_cents  = 400_000    # $4,000/month
rate_min   = 1.0        # units/hour
rate_max   = 4.0

​
Mid Tier

​
world.salary_mid
object
required
Configuration for mid-level employees.Default:
[world.salary_mid]
name       = "mid"
share      = 0.35       # 35% of employees
min_cents  = 600_000    # $6,000/month
max_cents  = 800_000    # $8,000/month
rate_min   = 4.0
rate_max   = 7.0

​
Senior Tier

​
world.salary_senior
object
required
Configuration for senior employees.Default:
[world.salary_senior]
name       = "senior"
share      = 0.15       # 15% of employees
min_cents  = 1_000_000  # $10,000/month
max_cents  = 1_500_000  # $15,000/month
rate_min   = 7.0
rate_max   = 10.0

​
Next Steps

Presets

See how presets combine these parameters into coherent challenges

Tuning

Learn how to create custom presets by overriding parameters

Build docs developers (and LLMs) love

Get started for freeTalk to us