VERSION 2.0 · JULY 2026
Twenty-eight decisions that turn a founder simulation into a persistent economy: the world model, the Spark currency, the daily work, the rulebook, and the observation layer that makes any of it worth playing.
What CLIFF is, what it refuses to be, and the four decisions that everything downstream inherits. Settle these before a line of engine code is written, because none of them can be retrofitted.
The first draft described a facilitated session: four to twenty players, twelve fixed rounds, equal starting balances, no carryover, and a single winner by weighted score. That is a workshop product. It is sellable, but it is not what you are building.
A persistent world and a bounded session are not two settings of the same engine. They are two engines. Build persistent first. The session format survives as a mode: the same engine, a fixed seed, a ninety-minute clock, and a facilitator view.
A weighted composite score is the most tempting mechanic in this genre and the most corrosive. Publish the weights and players optimise the weights. Hide them and players feel cheated. CLIFF has no winner. It has outcomes, and the founder chooses which one to take.
Exit is the benchmark, not the trophy. When a founder exits, the valuation and the terms are written permanently into their record and into the ecosystem comparables that price every other company in that sector.
CLIFF does not tell anyone whether they are a founder. It tells them what they did under scarcity, and it links every claim to the decision that produced it. A verdict is a claim about a person that a few hours of play cannot support. An evidence-linked observation is defensible, useful, and true.
Weight a player's first run highest. Once word travels that CLIFF is watching, later runs measure learning rather than instinct.
The currency was called FounderCoin in one section of the first draft, CHALK in another, and FDC in the tables. Three names for one unit is what makes an economy feel unfinished. Settle it once and never write it another way.
Five jurisdictions, one migration route, one daily clock, and five investors who are wrong in interesting ways. The world is where the lessons live, and it must be small enough to keep accurate.
The first draft listed eight locations including Berlin, São Paulo, and Nairobi. Cut to five. Every jurisdiction is a maintenance liability whose rules change annually. The five that remain are the actual capital migration corridor a founder in your target market walks: India → Gulf → Singapore → London → Delaware.
| JURISDICTION | SIGNATURE LEVER | SIGNATURE COST |
|---|---|---|
| 🇺🇸 Delaware, US | Deepest capital pool, investor default | Highest burn, franchise and legal friction |
| 🇮🇳 India | Cheapest runway extension, deep talent | Slower closes, harder path to foreign capital |
| 🇬🇧 United Kingdom | SEIS and EIS relief pulls angels in early | Shallow growth-stage market above Series A |
| 🇸🇬 Singapore | Fastest clean incorporation, Asian holdco | Small domestic market, forces early expansion |
| 🇦🇪 UAE | Qualifying free zone income taxed at zero | Conditional: mainland revenue is non-qualifying, breaching de minimis forfeits status |
Companies incorporated in India, Singapore, or the UAE routinely redomicile into Delaware when they raise from American funds, and the cost in tax, time, and legal fees is discovered late and paid in full. Build the wall and let players hit it.
Build a Market Open instead of a login reward. Once per day at a fixed local time, everything resolves at once. Comparables recalculate. Investor agents post mandates. Acquisition offers appear carrying an expiry. The day's burn is charged. Pending events resolve.
Five investor agents, each with a stated thesis, a cheque range, a geography, and a documented weakness. The weakness matters more than the thesis. An investor who is reliably right teaches players to obey, and obedience is neither fun nor informative.
Six rules that decide whether a persistent economy survives past month two. If you build nothing else from this document exactly as specified, build this part exactly as specified.
One hard invariant governs the entire economy: a transfer between two players must never change the total supply of Spark. Only the world mints, through revenue, investment, and jurisdiction grants. Only the world burns, through operating cost, transaction friction, and penalties.
| CODE | DIRECTION | TARGET SHARE | TRIGGER |
|---|---|---|---|
MINT_REVENUE |
MINT | 60% | Market Open settlement, per active customer cohort |
MINT_ROUND |
MINT | 35% | Term sheet accepted; equity issued in the same transaction |
MINT_GRANT |
MINT | 5% | Jurisdiction relief claimed and conditions verified |
BURN_OPEX |
BURN | 55–65% | Daily settlement: headcount × jurisdiction cost index |
BURN_FRICTION |
BURN | 3% of P2P | 3% of every player-to-player transaction value |
BURN_PENALTY |
BURN | — | Compliance failure, breached covenant, forfeited status |
XFER_P2P |
NEUTRAL | 0% | Acquisition, swap, secondary. Net supply change must be zero. |
Sinks must be relentless. The primary one is daily operating burn scaled by headcount and jurisdiction. It should absorb between fifty-five and sixty-five percent of everything the faucets create. The second sink is a three percent friction fee burned on every player-to-player transaction.
If a language model decides valuations, players will find the phrasing that always produces a large number, publish it, and the economy ends that week. Every number in CLIFF is computed. The model supplies voice, posture, and reasoning, never price.
V = run_rate × sector_multiple × market_index × jurisdiction_factor × (0.5 + pmf/100)sector_multiple = clamp( decayed_median(last_20_closed_deals[sector]), floor, cap )
Every persistent economy inflates. The question is whether the correction arrives as a designed mechanism or as an emergency change that players experience as betrayal. Measure net supply change on a seven-day rolling window. Players will experience this as a market cycle rather than a correction.
Split it. Spark is spendable and transferable. Standing is neither. Standing is earned only from outcomes, decays with inactivity, and can never be bought with Spark at any price or through any intermediary.
Run the world in ninety-day seasons. At the close, every surviving company is marked to market and force-exited at its standing valuation. The founder banks a permanent record and their Standing carries forward. The economy does not.
Where the intelligence sits, what it costs, and the four rules that keep a language model out of the places it will break. This is the part your developers build from.
The dividing line runs through the whole product and it never moves. Code produces every number. The model produces every sentence. If a player can talk an agent into a valuation, a cheque size, or a term, the benchmark is dead on arrival and the economy follows it.
Set the target before the first sprint and hold the team to it: zero model calls on a typical day. A founder who logs in, reads the Market Open, adjusts headcount, and declines an acquisition offer should cost nothing but compute you already own.
| INTERACTION | LAYER | MODEL TIER | BUDGET |
|---|---|---|---|
| Market Open, burn, events | Deterministic | None | 0 tokens |
| Valuation and term generation | Deterministic | None | 0 tokens |
| Due diligence questions | Authored bank | None | 0 tokens · selected algorithmically |
| Free-text answer scoring | Model | Small | 5 short calls per raise, rubric-bounded |
| Negotiation responses | Templated | Small | Flavour only; concession already computed |
| Final investor decision | Model | Frontier | One call. The sentence the player screenshots. |
| Advisor counsel | Model | Tier by fee | Player-funded. Cost scales with the advisor they hired. |
CLIFF is also the most legible demonstration of model orchestration that Fahrenheit Research can build. An investor will not read an architecture diagram. They will play a game in which the quality of the intelligence they can afford visibly changes what happens to them.
The most damaging rule in the first draft was the equity dilution floor, which prevented any founder from falling below twenty-five percent ownership. Replace prohibition with consequence. Being fired from your own company is the most memorable thing that can happen in this game.
What a founder actually does between Market Opens. This is where the hours go, and it is the part of every business simulation that quietly decides whether anyone returns.
CLIFF has something no previous simulation had access to: a room full of other founders who all hold budgets. So make them the market. Your customers are other players. Not a demand curve. Companies with names, run by people, paying you out of their own treasury at every Market Open, free to leave at the next one.
A real launch produces two different outcomes: you get seen, and you get paid. Collapse them into one mechanic and you have built an upvote button, which players will brigade for their friends within a week because backing a bad product costs nothing.
| PROPERTY | VOUCH | ADOPT |
|---|---|---|
| Cost to give | Free, capped at two per Market Open | Recurring Spark, charged every cycle |
| Motive | Reputational | Self-interested: a real computed benefit |
| Reversible | Never. Permanent and attributed. | Cancellable at any Market Open |
| Determines | Visibility on the Floor | Revenue, and therefore valuation |
| Abuse defence | Weight decays when your vouches do not convert | Backing a bad product means paying for a bad product |
Replace budget sliders with a Workbench. A product is assembled from named components, each with a cost, a build time, and a quality that depends on who built it. A product made of named parts can be described in one sentence, which is what you need on the Floor.
Every operating decision in CLIFF trades between exactly three quantities, and no allocation holds all three. That constraint is the game, and it is why the loop can carry twenty minutes a day without a chart anywhere in sight.
A flat daily burn only threatens a player at the beginning. Once revenue clears it, the pressure disappears and the mid-game goes slack. The fix is to make cost scale with success rather than with time.
| COST LINE | SCALES WITH | DESIGN INTENT |
|---|---|---|
| Payroll | Headcount × jurisdiction index | Makes where you incorporated a permanent operating fact |
| Infrastructure | Customers served | Growth without pricing power becomes fatal. The mid-game teeth. |
| Support | Customers served | Forces the hire-or-absorb decision at exactly the wrong moment |
| Your own stack | Products you have adopted | Every cancellation is another founder's churn notification |
| Compliance | Jurisdiction, periodic | Arrives as a lump, always at the worst time, because filings do |
Everything elegant in this economy depends on other players. Launch with a thin cohort spread across sectors and the median of two deals is not a market, it is noise. So the world has resident companies, and they never go away.
They are deterministic state machines with behaviour profiles and seeded noise. A language model touches one only when a real player interacts with it, which makes the cost track real player count rather than world size.
Eight rules a founder holds in their head, and the far larger set the engine enforces quietly beneath them. The distinction between those two lists is the whole of the teaching design.
A founder should be able to start playing after reading eight sentences. Everything else the engine enforces is discovered by running into it, which is both better teaching and the only onboarding that fits inside the window you actually have.
Your first draft opened with a fifteen-minute guided tutorial. The window in which a new player decides whether this product is for them is five to fifteen minutes, so that tutorial is competing directly with your own game and it loses.
| THE LESSON | THE MOMENT IT ARRIVES | COST OF LEARNING IT |
|---|---|---|
| Revenue is not free | First infrastructure bill after ten adoptions | A few weeks of runway |
| Reliability is bought | Public outage on the Floor | Several named customers, at once |
| Price is a decision | First raise in price, first churn | The customers who never valued you |
| Terms outrank valuation | Exit, when the preference stack resolves | Most of the proceeds. Taught once, remembered permanently. |
| Ownership is control | The Market Open after crossing below half | A vetoed decision you wanted |
| Concentration is fragility | Largest customer asks for a discount | Either the customer or the price |
Beneath the eight there is a much larger ruleset that the engine enforces continuously and never explains up front. It is documented here because your developers need it, not because a player ever reads it. Every one of these is discoverable in play, deterministic, and identical for every player.
| DOMAIN | RULE |
|---|---|
| Supply | Only the world mints. A player-to-player transfer never changes total supply. |
| Friction | Three percent of gross value burns on every player-to-player transaction. |
| Capital | Three investors per raise, in sequence. Three negotiation rounds each, two terms per round. |
| Pricing | Every term-sheet number is computed. Negotiation moves terms, never the mathematics. |
| Control | Below half you lose the vote. Below a quarter the board may remove you at any Market Open. |
| Diligence | Revenue is repriced for concentration, tenure, and reciprocity before a round is offered. |
| Jurisdiction | Flips are uncapped and priced. Relief accrued to date is forfeited on the flip. |
| Insolvency | Zero Spark with no standing offer triggers wind-down, which scores as a result. |
| Season | At close, every surviving company is marked to market. Standing carries. Spark does not. |
The instinct to grant a daily login bonus is strong and it is wrong here. Grant a first-look window instead. A founder present at Market Open sees new ships, new investor mandates, and new acquisition offers before they reach the rest of the cohort.
Build the loop before the world. One jurisdiction, one investor agent, no compliance module, a deterministic valuation core, and a Market Open that fires on a timer.
| WEEKS | BUILD | QUESTION IT ANSWERS |
|---|---|---|
| 1–2 | State vector, ledger, mint and burn codes, seeded generator | Does the supply assertion hold under synthetic load? |
| 3–4 | Market Open, burn settlement, one jurisdiction (India) | Does a founder come back the next day without a reward? |
| 5–6 | One investor agent, computed term sheet, three-round negotiation | Can a player talk the number up? If yes, stop and fix. |
| 7–8 | Observation ledger, evidence-linked post-run report | Does the report tell a founder something they did not know? |