Prestige System - YC-Bench

Overview

Prestige is the core progression mechanic in YC-Bench. It represents your company’s reputation and technical capability in four AI research domains. Higher prestige unlocks access to more valuable tasks, but comes with increased work requirements and decay pressure.

Four Domains

Every company tracks prestige independently in four domains (from src/yc_bench/db/models/company.py:12):

class Domain(str, Enum):
    RESEARCH = "research"
    INFERENCE = "inference"
    DATA_ENVIRONMENT = "data_environment"
    TRAINING = "training"

Domain Semantics

Domain	Represents	Example Tasks
Research	Novel algorithm development, paper implementations	”Implement sparse attention mechanism”, “Reproduce GPT-4 architecture”
Inference	Model serving, optimization, deployment	”Optimize inference latency”, “Deploy model to production”
Data/Environment	Dataset curation, data pipelines, infrastructure	”Build training data pipeline”, “Set up distributed training cluster”
Training	Model training, hyperparameter tuning, stability	”Train 7B parameter model”, “Optimize training stability”

Domain names are for flavor — mechanically, they’re interchangeable. What matters is per-domain prestige tracking and multi-domain task requirements.

Prestige Range

Prestige levels range from 1.0 to 10.0 (enforced by database constraints in src/yc_bench/db/models/company.py:38):

1.0: Starting prestige (all domains)
3.0–5.0: Typical operating range for mid-game
7.0–10.0: Elite tier (highest-paying tasks, hardest to maintain)

Prestige Distribution

┌─────────┬──────────────────────────────────────┬──────────────┐
│ Level   │ Market Access                        │ Difficulty   │
├─────────┼──────────────────────────────────────┼──────────────┤
│ 1.0-2.0 │ Entry-level tasks (low pay)          │ Easy         │
│ 3.0-5.0 │ Majority of market (profitable)      │ Moderate     │
│ 6.0-8.0 │ High-value tasks (requires focus)    │ Hard         │
│ 9.0-10.0│ Elite tasks (razor-thin margins)     │ Very Hard    │
└─────────┴──────────────────────────────────────┴──────────────┘

Per-Domain Prestige Gating

This is the key mechanic that prevents single-domain hyper-specialization.

Acceptance Rule (from `src/yc_bench/cli/task_commands.py:68`)

When accepting a task:

for req in task.requirements:  # Each domain the task requires
    domain_prestige = company.prestige[req.domain]
    if task.required_prestige > domain_prestige:
        raise PrestigeError(
            f"Company prestige in {req.domain} ({domain_prestige:.1f}) "
            f"does not meet task requirement ({task.required_prestige})"
        )

Example

Task “Train large language model”:

Required prestige: 5
Required domains: [training, data_environment]

Acceptance check:

if company.prestige.training >= 5 AND company.prestige.data_environment >= 5:
    accept_task()
else:
    reject("Insufficient prestige")

You cannot accept this task if:

Training prestige = 6.0, Data prestige = 4.5 ❌
Training prestige = 5.0, Data prestige = 5.0 ✅

Most tasks in medium and hard presets require 2 domains. Agents must maintain broad prestige across all four domains to access the full market.

Prestige Levels (1.0 to 10.0)

Prestige is stored as a Decimal(6, 3) — three decimal places of precision. Example: 5.125.

Starting Prestige

All domains start at 1.0 (from default.toml:57):

initial_prestige_level = 1.0

Prestige Floor and Ceiling

From default.toml:68:

prestige_min = 1.0
prestige_max = 10.0

Prestige cannot fall below 1.0 or exceed 10.0.

Prestige Rewards

Successful task completion awards prestige to all domains required by the task.

Reward Formula (from `src/yc_bench/core/handlers/task_complete.py:72`)

if task.success:
    for domain in task.requirements:
        prestige_delta = task.reward_prestige_delta  # 0.0–0.35 (typical ~0.1)
        company.prestige[domain] = min(
            prestige_max,
            company.prestige[domain] + prestige_delta
        )

Prestige Delta Distribution (from `default.toml:148`)

[world.dist.reward_prestige_delta]
type = "beta"
alpha = 1.2
beta = 2.8
low = 0.0
high = 0.35

Typical deltas: 0.05–0.15 per domain per task. Prestige climb example:

Start: 1.0
Complete 5 tasks (avg +0.1 each): → 1.5
Complete 10 more tasks: → 2.5
Complete 20 more tasks: → 4.5

Climbing from prestige 1.0 → 5.0 in one domain takes ~40 successful tasks on average (assuming reward_prestige_delta ≈ 0.1).

Prestige Decay Mechanics

Prestige decays continuously at a rate of prestige_decay_per_day (default: 0.005/day).

Decay Formula (from `src/yc_bench/core/engine.py:108`)

def apply_prestige_decay(db, company_id, days_elapsed):
    decay_per_day = 0.005  # default
    total_decay = decay_per_day * days_elapsed
    floor = 1.0
    
    for domain in [research, inference, data_environment, training]:
        company.prestige[domain] = max(
            floor,
            company.prestige[domain] - total_decay
        )

Decay Rates

From default.toml:83:

prestige_decay_per_day = 0.005

Monthly decay: 0.005 × 30 = 0.15 Annual decay: 0.005 × 365 = 1.825

Decay Examples

Starting Prestige	Days Idle	Final Prestige
5.0	30 (1 month)	4.85
5.0	180 (6 months)	4.10
5.0	365 (1 year)	3.175
2.0	200	1.0 (floored)

A domain left untouched for 6 months loses ~1.0 prestige level. Agents must stay active across all domains to maintain market access.

Prestige-Scaled Rewards

Higher-prestige tasks pay proportionally more, scaled by reward_prestige_scale.

Reward Scaling Formula (from `src/yc_bench/services/generate_tasks.py`)

base_reward = sample_from_distribution(reward_funds_cents)  # e.g., $14,000
scale = world_config.reward_prestige_scale  # 0.55 (default)
prestige = task.required_prestige  # 1–10

final_reward = base_reward * (1 + scale * (prestige - 1))

Scaling Factor (from `default.toml:78`)

reward_prestige_scale = 0.55

Reward Multiplier by Prestige

Prestige	Multiplier	Example ($14K base)
1	1.00×	$14,000
3	2.10×	$29,400
5	3.20×	$44,800
7	4.30×	$60,200
10	5.95×	$83,300

At reward_prestige_scale = 0.55, a prestige-10 task pays ~6× more than a prestige-1 task. This creates strong incentive to climb prestige despite decay pressure.

Prestige-Scaled Work Volume

Higher-prestige tasks require proportionally more work, scaled by prestige_qty_scale.

Work Scaling Formula

base_qty = sample_from_distribution(required_qty)  # e.g., 2000 units
scale = world_config.prestige_qty_scale  # 0.3 (default)
prestige = task.required_prestige  # 1–10

final_qty = base_qty * (1 + scale * (prestige - 1))

Scaling Factor (from `default.toml:88`)

prestige_qty_scale = 0.3

Work Multiplier by Prestige

Prestige	Multiplier	Example (2000 base)
1	1.00×	2,000 units
3	1.60×	3,200 units
5	2.20×	4,400 units
7	2.80×	5,600 units
10	3.70×	7,400 units

Higher prestige pays more but demands proportionally more capacity. A prestige-7 task pays 4.3× more but requires 2.8× more work. Net margin improves, but throughput pressure increases.

Prestige Penalties

Failing or cancelling a task applies a prestige penalty to all required domains.

Failure Penalty (from `src/yc_bench/core/handlers/task_complete.py:114`)

if not task.success:  # Completed after deadline
    penalty = penalty_fail_multiplier * task.reward_prestige_delta
    for domain in task.requirements:
        company.prestige[domain] = max(
            prestige_min,
            company.prestige[domain] - penalty
        )

Penalty multiplier (from default.toml:74):

penalty_fail_multiplier = 1.4  # 1.4× the reward delta

Example:

Task reward: +0.10 prestige per domain
Failure penalty: -0.14 prestige per domain

Cancellation Penalty (from `src/yc_bench/cli/task_commands.py:398`)

penalty = penalty_cancel_multiplier * task.reward_prestige_delta
for domain in task.requirements:
    company.prestige[domain] = max(
        prestige_min,
        company.prestige[domain] - penalty
    )

Penalty multiplier (from default.toml:74):

penalty_cancel_multiplier = 2.0  # 2× the reward delta

Example:

Task reward: +0.10 prestige per domain
Cancel penalty: -0.20 prestige per domain

Cancelling is worse than failing (2.0× vs. 1.4× penalty). Accepting a task is a real commitment — agents cannot freely “option” tasks without consequences.

Prestige Strategy Implications

Broad vs. Narrow Specialization

Narrow strategy (focus on 1–2 domains):

✅ Climb prestige faster in chosen domains
✅ Access high-paying tasks in those domains
❌ Locked out of multi-domain tasks (most of the market)
❌ Unused domains decay → further market restriction

Broad strategy (maintain all 4 domains):

✅ Access to full market (including 2–3 domain tasks)
✅ Resilient to decay (constant activity in all domains)
❌ Slower prestige climb (rewards split across domains)
❌ Higher employee coordination complexity

In medium and hard presets, most tasks require 2 domains. Agents that specialize too narrowly will starve as their unused domains decay below market requirements.

Decay Mitigation

To maintain prestige level P in a domain:

monthly_tasks_needed = (monthly_decay) / (avg_reward_prestige_delta)
                     = 0.15 / 0.10
                     = 1.5 tasks/month

To climb prestige while offsetting decay:

net_gain = (tasks_per_month × avg_delta) - monthly_decay
         = (4 × 0.10) - 0.15
         = +0.25/month

Completing ~2 tasks per domain per month maintains prestige. Completing 4+ tasks per domain per month enables prestige growth.

Observing Prestige

Agents can query current prestige via:

yc-bench company status

Example output:

{
  "company_id": "...",
  "name": "BenchCo",
  "funds_cents": 15000000,
  "prestige": {
    "research": 3.125,
    "inference": 2.850,
    "data_environment": 4.200,
    "training": 3.675
  },
  "sim_time": "2025-06-15T14:30:00Z"
}

Next Steps

Task Management

Learn how prestige affects task acceptance and deadline calculations.

Employee System

Understand how employees contribute to tasks across domains.

Scoring

Learn how prestige levels factor into final benchmark scores.

Configuration

Tune prestige decay, rewards, and penalties in your presets.

Get Started

Core Concepts

Configuration

Development

​Overview

​Four Domains

​Domain Semantics

​Prestige Range

​Prestige Distribution

​Per-Domain Prestige Gating

​Acceptance Rule (from src/yc_bench/cli/task_commands.py:68)

​Example

​Prestige Levels (1.0 to 10.0)

​Starting Prestige

​Prestige Floor and Ceiling

​Prestige Rewards

​Reward Formula (from src/yc_bench/core/handlers/task_complete.py:72)

​Prestige Delta Distribution (from default.toml:148)

​Prestige Decay Mechanics

​Decay Formula (from src/yc_bench/core/engine.py:108)

​Decay Rates

​Decay Examples

​Prestige-Scaled Rewards

​Reward Scaling Formula (from src/yc_bench/services/generate_tasks.py)

​Scaling Factor (from default.toml:78)

​Reward Multiplier by Prestige

​Prestige-Scaled Work Volume

​Work Scaling Formula

​Scaling Factor (from default.toml:88)

​Work Multiplier by Prestige

​Prestige Penalties

​Failure Penalty (from src/yc_bench/core/handlers/task_complete.py:114)

​Cancellation Penalty (from src/yc_bench/cli/task_commands.py:398)

​Prestige Strategy Implications

​Broad vs. Narrow Specialization

​Decay Mitigation

​Observing Prestige

​Next Steps

Task Management

Employee System

Scoring

Configuration

Build docs developers (and LLMs) love

Overview

Four Domains

Domain Semantics

Prestige Range

Prestige Distribution

Per-Domain Prestige Gating

Acceptance Rule (from `src/yc_bench/cli/task_commands.py:68`)

Example

Prestige Levels (1.0 to 10.0)

Starting Prestige

Prestige Floor and Ceiling

Prestige Rewards

Reward Formula (from `src/yc_bench/core/handlers/task_complete.py:72`)

Prestige Delta Distribution (from `default.toml:148`)

Prestige Decay Mechanics

Decay Formula (from `src/yc_bench/core/engine.py:108`)

Decay Rates

Decay Examples

Prestige-Scaled Rewards

Reward Scaling Formula (from `src/yc_bench/services/generate_tasks.py`)

Scaling Factor (from `default.toml:78`)

Reward Multiplier by Prestige

Prestige-Scaled Work Volume

Work Scaling Formula

Scaling Factor (from `default.toml:88`)

Work Multiplier by Prestige

Prestige Penalties

Failure Penalty (from `src/yc_bench/core/handlers/task_complete.py:114`)

Cancellation Penalty (from `src/yc_bench/cli/task_commands.py:398`)

Prestige Strategy Implications

Broad vs. Narrow Specialization

Decay Mitigation

Observing Prestige

Next Steps