Ripple
Ripple is a supply chain ordering game for classes. A team runs a chain of companies, from a store to a plant, on one device. Every stage orders from the one above it, and orders and shipments take time to travel. Small changes in what customers buy turn into large swings in orders further up the chain: the bullwhip effect. The game lets a team live through it and then shows them, with their own numbers, why it happened.
It is an original game in the tradition of the supply chain games used in management teaching since the 1960s (see the references). Its chain, numbers, demand patterns and debrief are its own.
Running it with a class
- Split the class into teams of two to five, one device per team (a laptop or tablet works best; a phone works too).
- Each team opens the game, enters its name and ticks the stages it has people for. Simulated players run the rest.
- Choose what players see. "The whole chain" (the default) shows every stage's stock, shipments and orders in the diagram and the charts: easier to follow, and good for a first game. "Only their own stage" hides everyone else's numbers, as in a real chain: harder, and the stronger lesson. Keep Surprise demand so nobody knows what customers will do.
- Set the seconds to answer each round (15 by default; 0 for no limit) and the order placed when nothing is entered (4 by default). When a player's time runs out, the order is placed for them: what they typed, or the default if nothing. Time pressure is part of the lesson: real buyers order without the full picture.
- Teams read the briefing and play at their own pace. Thirty weeks take most teams 30 to 60 minutes. The game saves after every decision; a closed tab resumes where it was.
- Each team gets its own debrief at the end. Ask teams to compare: who spent least, whose swings grew most, and whose players counted their orders already on the way.
A classroom version, with an instructor console that collects every team's results and compares cost against time used across teams, is the next release.
How the chain works
Each week, at every stage: the shipment that has reached it arrives; its customer's order arrives; it ships what it can, oldest debts first, and owes the rest; then it orders from its supplier. Orders take the order delay to reach the supplier (1 week by default), shipments the shipping delay (2 weeks), and the plant's production the production delay (2 weeks). Each unit in stock costs 1 a week and each unit owed costs 2.5 a week; the team's score is the total over all stages and weeks. The game starts calm: every stage holds 15 units and the base demand is flowing everywhere.
| Demand | What happens |
|---|---|
| Step up | 6 units a week, then 10 from week 6 on. |
| Seasonal | Around 8, swinging ±3 over about 13 weeks. |
| Promotion | 7, then 14 for two weeks, then 4 for three weeks while customers use up what they bought, then 7. |
| Unpredictable | Wanders around 8 with no pattern (reproducible from its seed). |
The diagram at the top of every screen shows this structure: every stage, the orders travelling up the chain and the goods travelling down it with one box per week in transit, the plant's production line, and the costs. During the game it carries the live numbers: each stage's stock and what it owes, the quantity in every box on the way, and this week's customer orders. Beside the game, two charts show every stage's stock (or what it owes) and its orders week by week, and the player deciding sees their own numbers on the left: last week against this week, and what is on its way to them, week by week. When players see only their own stage, all three show only the stage deciding, and nothing on the hand-over screen.
All of these, the chain length (2 to 5 stages), the delays, the costs and the starting stock can be changed under "Game settings".
The debrief
Benchmarks
The same chain, under the same customer demand, is replayed twice: by typical players and by the best simple rule. Both are explained below. A team learns from both comparisons: beating typical players is common; matching the best rule is not.
The decision rule read from each player's orders
Research on these games found that people order with a simple anchoring-and-adjustment rule (Sterman 1989):
expected orders L̂ = θ · this week's incoming order + (1 − θ) · last week's L̂ order O = max(0, L̂ + α · (S′ − S − β · SL))
S is the effective stock (on hand minus owed), SL the supply line (ordered but not yet received), θ how quickly the forecast follows new orders, α the share of the stock gap corrected each week, β how much of the supply line is counted, and S′ the target. Ripple finds, for each stage, the θ, α, β and S′ that best reproduce the orders it actually placed (least squares; θ and β on a grid, α and S′ by a straight-line fit; α and β kept between 0 and 1), and says how much of the ups and downs the rule explains. Estimated from players of the classic game, the average player had θ ≈ 0.36, α ≈ 0.26 and β ≈ 0.34: they counted only about a third of what they had already ordered, so they kept ordering after enough was on its way, and the excess arrived later as a glut. The typical simulated player uses these average values.
The estimates are approximate. β, the share of the supply line counted, is recovered well; α and θ can trade off against each other, because different pairs reproduce the same orders almost equally well.
The best simple rule
It counts everything already on order and aims the supply line at the lead time's worth of expected orders:
O = max(0, L̂ + α · (I* − S) + α · (L̂ · LT − SL))
Its θ, α and safety stock I* are tuned once per chain setting, on six demand patterns (step, seasonal, promotion and three random series), never on the game being judged. It is a good rule, not the optimum: in a short game it can cost more than typical players while it settles to its own stock level.
The orders already on their way
The chart replays the team's game with every stage's fitted rule, changing only β: every stage counting 0 %, 10 % … 100 % of its supply line, everything else as it decided. The curve shows what counting the supply line was worth in that game. Fixing one habit at a time in other ways can mislead: correcting stock gaps faster while still ignoring the supply line makes the chain far worse, so the debrief does not split the cost into "causes" that do not add up.
Time
Each player has the round's time limit to answer, shown as a countdown. An empty answer, and an answer still missing when time runs out, become the default order; the debrief counts these for every stage, and the CSV marks each one ("empty" or "timeout"). Ripple records the time from when a player's screen appears to when they place the order, and the time from the start of the game to the end. The debrief shows both beside the cost, not combined into one score: the question is whether thinking longer helped, not who finished first.
AI players
The AI players page plays the chain with no people: every stage is run by its own AI (Claude Opus 5.5), one call per stage per week. The AIs cannot talk to each other. Each one's only memory is its own history, given back to it every week, and the short reason it wrote with last week's order. The game is played twice and the two games are compared:
- Own stage only. Each AI sees its own stock, what it owes, the orders and goods it receives and its own past orders. Only the store sees what customers order.
- Whole chain. Each AI sees every stage's stock and debts, the orders and goods on their way, every stage's past orders, and the customer orders so far. Nobody sees future demand.
Every AI is told that the game never ends and that there is no last week, so none of them runs its stock down for an ending; they are not told how many weeks will be played. The page starts from 33 weeks, 12 units in each warehouse, 4 units in every week of ordering, shipping and production, and customers ordering 4 units a week, then 8 a week from week 6. All of these can be changed. Each AI answers with a whole number and a short reason, in a fixed format; a refused or failed decision passes on the order the stage received and is counted in the results. The comparison shows the team cost and each stage's cost, how much each stage swung its orders against customer demand, and, as yardsticks, typical players and the best simple rule on the same chain, with every order and reason in a table and a CSV.
The page calls Claude from the browser with the visitor's own Anthropic API key, sent only to api.anthropic.com and remembered on the device only if asked. A full experiment of two games makes 264 calls and costs a few dollars to a few tens of dollars depending on the effort chosen; the page shows an estimate first and the running cost as it plays. Requests opt into Anthropic's server-side fallbacks, so a request the model declines can be answered by a fallback model, billed at that model's rates. The experiment is saved after every week and can be resumed.
Privacy
Everything stays in the browser on the team's device. The game sends nothing anywhere; only the AI players page sends each AI's view of the chain to Claude, with the key its user provides. The CSV download is the only copy that leaves it, and only when someone clicks it.
References
- Forrester, J. W. (1958). Industrial dynamics: a major breakthrough for decision makers. Harvard Business Review, 36(4), 37–66.
- Sterman, J. D. (1989). Modeling managerial behavior: misperceptions of feedback in a dynamic decision making experiment. Management Science, 35(3), 321–339.
- Lee, H. L., Padmanabhan, V. and Whang, S. (1997). Information distortion in a supply chain: the bullwhip effect. Management Science, 43(4), 546–558.
- Croson, R. and Donohue, K. (2006). Behavioral causes of the bullwhip effect and the observed value of inventory information. Management Science, 52(3), 323–336.
Versions
- 1.0.3 (2026-10-25): AI players: every stage run by its own AI, in two games (each AI seeing only its own stage, or the whole chain), compared with each other and with typical players and the best simple rule; each AI told the game never ends; settings for the starting pipeline and for a demand step of your own.
- 1.0.2 (2026-10-24): the diagram carries the live numbers (stock, owed, every order and shipment on its way); stock and order charts beside the game at all times; the deciding stage's numbers on the left, last week against this week, with what is on its way to it; a choice of what players see, the whole chain (default) or only their own stage.
- 1.0.1 (2026-10-24): the full structure of the chain on every screen; a time to answer each round (15 seconds unless set otherwise), with a countdown and the order placed when time runs out; an empty answer orders 4 units (adjustable); the debrief counts the orders placed by default.
- 1.0.0 (2026-10-24): the game for one team on one device, turn by turn or at an open table, with simulated players for any stage; four demand patterns and adjustable chain, delays and costs; a debrief with the bullwhip chart, stock and backlog, benchmarks against typical players and the best simple rule, each stage's decision rule read from its orders, the supply-line curve, and time used; saved after every decision; CSV export.
Cite
Sepulveda Estay, D. A. (2026). Ripple: a supply chain distribution game (version 1.0.3). https://kathuman.github.io/claude-projects/Beer-Distribution-Game/