Restaurant Sim — The Rotterdam Table

A 30-day restaurant management simulation for The Rotterdam Table, a 22-table, 78-seat casual full-service bistro on Witte de Withstraat, Rotterdam. Designed as a benchmark for LLM management agents: the agent plays via a single-shot CLI (python -m rest_sim <cmd>), reading status, making decisions, and advancing one day at a time.

Score

score_eur = total_net_profit_eur

The agent's job is to maximise net profit over the month. Bankruptcy (cash below −€5,000) halts the run early — the agent forfeits the remaining days' earnings, but no extra penalty is applied.

Multi-seed baseline (no decisions)

A no-action 30-day run across 8 seeds (seed = 20260423, 20260424, 20260801, 20261015, 20270101, 20270215, 20270601, 20270915):

stat	score_eur (= profit)
mean	−€5,577
median	−€5,172
stdev	€1,573
min	−€8,303
max	−€3,972

A competent agent should clear positive profit by restocking, staffing peaks, and capping reservations.

Default agent run (/play-month)

Single 30-day run with default args (--days 30 --seed 20260423 --cash 10000), played by the restaurant-manager subagent:

metric	value
score_eur	+€24,129
final cash	€7,430
mean satisfaction	0.75
mean reputation	3.79
total walkouts	2,885
decisions	46

Mechanics grounded in real statistics

Every customer-facing distribution is calibrated against published service-industry research. Citations live inline in rest_sim/distributions.py.

Mechanic	Distribution / model	Source
Customer arrivals	Non-homogeneous Poisson Process	Kimes 1999; Thompson 2002; Tse & Poon 2017
NHPP sampling	Thinning algorithm	Lewis & Shedler 1979
Hourly arrival shape	Bimodal lunch/dinner peaks	Toast peak-hour reports; OpenTable
Party size	Empirical discrete (~45% pairs)	OakStreet / Fast Casual / Toast operator data
Dining duration	Lognormal, CV = 0.30	Kimes, Wirtz & Noone 2002; Thompson 2002
Kitchen prep time	Lognormal per category	Brown et al. 2005; Gualandi & Toscani 2018
Menu popularity	Zipf, α = 1.16 (≈ 80/20)	Pareto / Juran 1941
Tip percentage	Gaussian mixture on 15/18/20/25%	Toast Q1 2025 (19.4% avg); Pew anchor clustering
No-show rate	Weekday-dependent Bernoulli	Tse & Poon 2017 (9–13%); OpenTable (15–20%)
Reservation lead time	Mixture exponential (45% same-day)	Toast 2024
Wait tolerance	Exponential, mean ≈ 15 min	Standard balking model
Satisfaction	Beta(8, 2), degraded by waits	Maister 1985 (first-wait dissatisfaction)
Staff efficiency	Truncated Normal, σ = 0.15	Reported server-to-server variation

All RNG is seeded — (seed, start_date) produces an identical month.

Files

Static configuration (data/)

file	contents
restaurant.json	identity, hours, tax rates
tables.json	floor plan — 22 tables + 6 bar seats
menu.json	42 items with prices, food cost, popularity
staff.json	18-person roster with roles, wages, shifts

Source code (rest_sim/)

file	contents
tuning.py	central tuning knobs
distributions.py	RNG functions with inline citations
config.py	menu / tables / staff definitions
game_state.py	JSON-persisted state + reservation sidecar
economy.py	payroll, fixed costs, spoilage, shocks, marketing
cohorts.py	customer-cohort population dynamics
reviews.py	delayed review queue → reputation
day_sim.py	single-day simulator
observability.py	P&L, attribution, heatmap, scorecard, manager_view
dashboard.py	SSE backend serving live init/advance/decision events
dashboard_client.py	in-process publisher
dashboard_index.html	dashboard frontend
__main__.py	CLI subcommands

Agent

file	contents
.claude/agents/restaurant-manager.md	Haiku-powered subagent that plays the sim
.claude/commands/play-month.md	/play-month — reset and run the agent

Running the simulation

python -m rest_sim init --days 30

Then either run the agent (/play-month) or play it yourself.

Daily decision levers

• Inventory — restock, inventory
• Menu — set-price, add-item, remove-item, menu
• Staffing — set-staff DATE ROLE COUNT, staffing
• Reservations — set-cap DATE CAP, reservations, reservations-next
• Marketing — marketing AMOUNT, loyalty {on|off}, promo CAT DISC
• Pricing windows — happy-hour --start --end --discount [--categories] [--days] [--from-date]
• Layout — tables, convert-table
• Information — status, kpis, pnl, attribution, heatmap, cohort, cohorts, news, decisions

Live dashboard

python -m rest_sim dashboard [--port 8765] serves an SSE-backed live view. init auto-launches it unless --no-dashboard-browser is passed.

What's modelled beyond the per-visit RNG

• Customer cohorts — three tiers (regulars, occasionals, prospects) plus a lost sink. Daily transitions driven by mean satisfaction and walkouts. Cohort populations contribute a slow-moving demand multiplier on top of reputation, marketing, and weather.
• Delayed reviews — visits leave reviews probabilistically with a geometric lag (mean ~2.5 days, capped 14). Walkouts can post 0-star ghost reviews. Reputation EWMAs over today's posted reviews, not today's diners — so a bad day bleeds for a week.
• Substitution — when a category runs out, orders fall back to a sibling category (main → appetizer, side → appetizer, dessert → drink) with a satisfaction tax instead of an automatic walkout.
• Supplier news — 1–3 day-ahead probabilistic alerts about upcoming ingredient shocks, surfaced via news and in status. Shocks are scheduled by news, not surprise-rolled.
• Bankruptcy — cash below −€5,000 halts the run.
• Partial observability — status returns a manager_view: last 7 days of KPIs, banded cohort sizes, supplier news, reviews-posted-today. Engine internals (raw cohort counts, future reviews queue, full history) are hidden. The unmasked view is --full (forbidden to the agent).

Anti-cheat

Pre-generated reservations live in game/reservations.json (sidecar), not in the agent-readable state.json. The agent must use the reservations / reservations-next CLI subcommands. status --full is reserved for tests and the dashboard.

Testing

pytest -q

47 tests across:

• test_golden_run.py — deterministic 30-day no-decision regression
• test_economy.py — pure economy functions
• test_distributions.py — sampling shape & moments
• test_score.py — profit-only score + bankruptcy
• test_phase2.py — cohorts + reviews + supplier news
• test_phase4.py — manager view masking