ProjekinAja. Founder's Playbook Stage 3 · Revenue Architecture Lesson 3.8
● Stage 3 · Lesson 3.8

Financial Projections — Building a Model That Is Wrong in the Right Ways

A financial model is not a forecast — it is a machine for testing assumptions; here is how to build the minimum viable model that makes you a better decision-maker.

⏱ ~22 min read 📊 Financial modeling 🔢 Driver-based P&L

🎯 By the end you'll understand

Why every financial model is wrong — and why being wrong in the right ways is actually useful. How to build a driver-based model from the 8–12 inputs that control all your outputs, so that changing one assumption propagates correctly through the full P&L. What the three-scenario structure does (and what it does not do). The 10 most common projection errors that investors use to identify founders who have reverse-engineered a predetermined conclusion. And how to read your own model critically — not as a forecast, but as a sensitivity instrument for discovering which decisions actually matter.

KEY INSIGHTA model is a machine for testing assumptions, not a prediction of the future

The purpose of a financial model is not to produce the right number — it is to reveal which inputs you are most exposed to, so you know where to focus your energy and where you can afford to be wrong.

The most common mistake in early-stage financial modeling is treating a spreadsheet as a forecast. No pre-seed or seed-stage model has ever been correct. The market does not care about your row 47. What a well-built model actually does is force you to articulate your assumptions explicitly, make the dependencies between them visible, and then reveal which assumptions dominate the outcome. That is genuinely useful — not because you will be right, but because you will know where being wrong is expensive.

🔑 The right framing for a startup financial model

A financial model is a structured set of assumptions connected by arithmetic. The assumptions will be wrong. The arithmetic cannot be wrong. If the arithmetic is wired correctly — if revenue flows from customers, if customers flow from acquisition, if burn flows from headcount and OpEx — then the model is useful even though every number in it is an estimate. The mistake is optimizing for impressive outputs rather than honest inputs. Investors who have reviewed hundreds of models are reading the assumptions, not the revenue line. They are asking: "Does this founder know which numbers are the load-bearing ones?"

CORE CONCEPTDriver-based modeling: the 8–12 inputs that control everything

Most early-stage models are built in the wrong direction: founders start with a revenue target ("we want to be at $5M ARR by year 3"), back into customer counts, and then populate the cost side to achieve a desired burn rate. This produces a model that is internally consistent but epistemically empty. The revenue target was arbitrary, the customer counts are derived from it, and the costs are whatever is left over. Nothing in that model tells you anything about how your business actually works.

Driver-based modeling works in the opposite direction. You start by identifying the handful of inputs that drive everything else — typically 8 to 12 for an early-stage B2B SaaS company — and build the model upward from those. Every output becomes a function of the drivers. Change one driver, and every downstream calculation updates.

Driver category Typical inputs What it drives
Growth inputs New customers per month, ACV New ARR added each month; ARR waterfall
Retention inputs Monthly churn rate, expansion rate Net Revenue Retention; ARR at end of period
Margin inputs Gross margin % Gross profit; contribution margin; LTV calculation
Cost structure OpEx as % of revenue, headcount, fully-loaded cost per head EBITDA; cash burn; runway to zero

For a manufacturing SaaS company like Forge, the eight drivers that govern the 24-month P&L are: new customers per month, ACV, monthly churn rate, expansion rate, gross margin, OpEx as a percentage of revenue, headcount, and average fully-loaded cost per head. Every other number in the model is arithmetic on top of those eight.

💡 Why 8–12 drivers, not 2 or 50?

Two drivers (revenue growth rate, burn rate) is too abstract — it does not show you which levers to pull. Fifty drivers is a bookkeeping exercise disguised as a model — there is no insight in a model where every line item is its own assumption. Eight to twelve forces discipline: you have to choose which assumptions are load-bearing and which are derived. The process of choosing is the insight. Most founders discover, in building a driver-based model for the first time, that 3 or 4 of their 8 drivers account for 80%+ of variance in the output. Those 3 or 4 are where your attention belongs.

DRIVERS (8 inputs) New customers ACV Churn rate Expansion Gross margin OpEx % Headcount Fully-loaded $/head ARR WATERFALL & REVENUE New ARR · Churned ARR · Expansion · Net ARR Monthly MRR → Annual ARR run-rate P&L OUTPUTS Gross Profit Operating Expense EBITDA / Burn Runway to Zero
Driver-based architecture: 8 inputs at the top flow into an ARR waterfall, then into the P&L outputs. Changing any single driver propagates correctly through every downstream calculation — which is what makes the model useful as a sensitivity instrument.

The key structural principle: the model has inputs and it has arithmetic. The inputs are in clearly marked cells (or, in this lesson, sliders). The arithmetic is fixed formulas connecting inputs to outputs. The moment you start treating an output cell as an input — manually overriding month 18 revenue to hit a round number — you have broken the model. The value of a driver-based model is precisely that you cannot cheat: you have to change the drivers, not the results.

HOW IT WORKSThe three-scenario structure — what it does and what it does not do

Every serious financial model runs three scenarios: bear, base, and bull. This structure is so standard that skipping it signals inexperience. But the three-scenario structure is also widely misunderstood. Here is what it does and does not accomplish.

✓

What three scenarios do: reveal the range of plausible outcomes

A well-built three-scenario model shows investors and founders the spread between a difficult but plausible downside and an ambitious but achievable upside. The bear case is not the apocalypse — it is the outcome if the harder assumptions prove out (slower sales cycles, higher churn, lower ACV, harder hiring). The bull case is not the dream — it is the outcome if the favorable assumptions prove out (faster adoption, better retention, strong expansion). The base case is your best honest estimate, not the middle of the range.

✗

What three scenarios do not do: reduce your exposure to tail risks

The scenario structure addresses uncertainty within the model's framework. It does not protect against structural shifts — a competitor closing a massive round and offering the product for free, a regulatory change affecting the customer's ability to deploy hardware, a macro event that freezes enterprise capex. These risks are not in any scenario because they are outside the model's assumption space. Running three scenarios gives you a false sense that you have addressed all major risks. The honest use of scenarios is narrow: "conditional on our business model working at all, this is the range of financial outcomes."

⚠

The structural requirement: scenarios must differ in assumptions, not just outputs

The most common scenario-building mistake is to run the base case, then manually set the bear case to "base × 0.7" and the bull case to "base × 1.3". This produces scenarios that look different but share identical structural assumptions — they are just scaled copies. A real bear case has different churn, different conversion rates, different sales cycle lengths. If you could produce your bear case by multiplying the base case by a constant, you have not actually modeled a different state of the world. You have just moved the decimal point.

📐 How to set scenario driver values

For each of your 8–12 drivers, define three values: bear (what if this is harder than expected?), base (honest best estimate), and bull (what if this works better than expected?). The bear and bull values should be individually defensible — not just ±30% but specific reasoning. "Bear churn is 3% monthly because early B2B SaaS in industrial sectors often has 25–35% annual churn until you have a formal customer success motion" is a scenario assumption. "Bear churn is 3% because that's what happens in the bad scenario" is not.

SEE ITDriver-Based Projection Builder — Forge's 24-month model

⚙ Interactive · Driver-Based P&L + ARR Waterfall + Runway

Set Forge's 8 model drivers below. The widget builds a 24-month P&L and ARR waterfall, identifies the top 3 drivers by sensitivity, and runs all three scenarios simultaneously. The cash runway indicator shows months to zero. Try the worst-case combination deliberately — identify the failure mode before investors do.

Forge context: Two pilots converted to annual contracts at $42K ACV each. Starting ARR: $84K. Business model: $30K annual SaaS + $12K one-time hardware per site (modeled as ACV). Gross margin currently 61% with a path to 75%+ at 20 customers. Raise: $2.2M seed with approximately $1.8M deployable after fees.

1.0 customers/mo
$42,000 $/yr per customer
1.50% % of ARR/month
0.50% % of existing ARR/mo
61% blended
120% % of MRR
5 FTEs
$130,000 $/yr per FTE
Month 24 ARR
—
Month 24 Gross Profit
—
Monthly Burn (Mo 24)
—
Cash Runway
—
Top 3 drivers by sensitivity (±20% change impact on Month 24 ARR):
Month Bear ARR Base ARR Bull ARR
Cash runway (based on cumulative burn vs. $1.8M deployable) —
0 mo 12 mo 24 mo 36 mo
📊 Model interpretation

Adjust the drivers above to see how the model responds.

A few things worth noticing as you use the widget. First: the sensitivity ranking rarely matches your intuition. Most founders expect ACV to dominate — and it does matter — but churn rate and new customer pace often carry comparable weight over 24 months because they compound. At month 1, a 1% monthly churn difference is tiny. At month 24, that same difference has eroded or preserved 12 months of accumulated ARR. Second: the bear and bull cases are not mirror images of the base. A 30% reduction in new customer pace is more damaging than a 30% improvement is beneficial, because the downside is bounded by zero but the upside is not — except burn continues regardless. Third: the runway indicator is the most important output in the model for a seed-stage company. Investors are reading this number first. It tells them whether you need another raise before you have proof, and that shapes every conversation about valuation.

IN PRACTICEThe ten most common projection errors

Investors who have reviewed hundreds of pitch decks can identify a model built to support a predetermined conclusion in under two minutes. Here are the ten patterns they look for — and the reasoning behind each:

1

Perfectly smooth growth curves

Real SaaS ARR curves are lumpy. Deals close in bursts around contract renewals, quarter-ends, and budget cycles. A model that shows exactly 2 new customers every month for 24 months does not reflect how B2B sales actually works. It reflects a founder who chose a monthly increment and copied it across 24 columns. Investors see this immediately and conclude the model was not built from driver reasoning — it was drawn.

2

Breakeven exactly when the raise runs out

The most-cited red flag in early-stage modeling. If the model happens to reach breakeven in the exact month that the current raise is depleted, the model was reverse-engineered. A real model produces a breakeven date as an output; a predetermined conclusion produces a model where all inputs are chosen to hit a desired output. This tells an investor that the founder does not understand their own unit economics — they just know what conclusion they need to reach.

3

Churn approaching zero by year two

Early-stage SaaS companies rarely see monthly churn below 1–1.5% before they have a mature customer success function. A model that starts at 3% monthly churn and declines to 0.2% by month 24 without a specific mechanism — new CS hire, product stickiness improvement, contract restructuring — is an assumption unsupported by evidence. Investors will ask what causes the churn improvement. If the answer is "we'll be better at customer success by then," that is not a driver — it is a wish.

4

Sales velocity that outpaces headcount

If your model adds 3 new customers per month in month 18 with the same 2-person sales team you have today, you are implicitly assuming productivity scales without investment. In B2B sales, productivity per rep either stays roughly constant or declines as you move to less warm prospects. More customers per month means more reps, more support capacity, and more CS — none of which appear in a model that holds headcount constant while growth accelerates.

5

Gross margin improving faster than the business explains

Gross margin improvement is real and predictable for hardware+software businesses — software margin improves as customer count grows without proportional COGS growth, while hardware margin improves with volume. But the rate of improvement must be grounded in specific mechanisms: "at 20 customers, hardware COGS drops from $9K to $6K per unit due to volume pricing from the sensor supplier" is a driver. "Gross margin improves from 61% to 80% by year 2" without a mechanism is a wish.

6

No cash buffer — the model assumes spending exactly what's raised

If your model deploys 100% of the raise and reaches breakeven with zero buffer, one delayed deal or one unexpected cost pushes the company into crisis. Experienced investors use this as a signal of planning inexperience. A reasonable model plans to reach a significant milestone — ideally a Series A raise point — with 3–6 months of runway remaining, not 0.

7

COGS scaling linearly with revenue for a software business

One of the defining characteristics of SaaS economics is that COGS scales sublinearly — adding a new customer costs less than the average customer cost, which is why gross margins expand over time. A model where COGS is a fixed percentage of revenue forever is modeling a services business, not a software business. If your product genuinely has linear COGS (e.g., heavy hardware or significant per-customer deployment labor), that is fine — but say so explicitly, because investors will otherwise wonder if you understand the difference.

8

Missing the S&M investment required to hit the growth number

If the model projects adding 3 new customers per month by month 12 but the sales and marketing budget stays flat, the customers are appearing for free. Even in a product-led or community-driven model, more customers require more pipeline, more sales support, or more demand generation investment. A model that grows revenue without growing S&M is implicitly assuming customer acquisition cost approaches zero. CAC does not trend to zero. It trends up as you exhaust warm markets and move into cold outreach.

9

One-line "miscellaneous" operating expenses

Professional services, software subscriptions, office, travel, legal, compliance, hardware inventory — these accumulate quickly in an industrial SaaS business and none of them are trivial. A model that has three cost lines (COGS, headcount, "other") and then a suspiciously tidy EBITDA is hiding the complexity of running a real company. Investors will ask what is in "other" and whether it scales with the business.

10

The bear case is still good

A bear case where you still hit $3M ARR and the business is obviously viable is not a bear case — it is a slightly less optimistic base case. A real bear case involves material pain: slower growth than expected, higher churn than hoped, extended cash burn, the possibility of running out of runway before the next raise. If your bear case would still allow a comfortable Series A at a strong valuation, you have not modeled the downside seriously. The question the bear case must answer is: "If things go badly, does the company survive?"

🔑 The meta-error that causes all ten

Every one of these errors has the same root cause: the founder built the model to support a conclusion rather than to test assumptions. When you start with "we need to show $5M ARR by month 30 to raise a Series A," you will unconsciously choose drivers that produce that output. The test: if you changed your ACV from $42K to $28K and your model did not break, your drivers are not actually driving your outputs. The model should feel uncomfortable when you enter realistic bear case numbers. If it does not feel uncomfortable, the model is not honest.

FORGEForge's first financial model — building from the Stage 3 milestone

Forge enters this exercise with a real foundation to model from, which is unusual and valuable. Most founders model from imagination. Forge has pilot data, signed contracts, and measured unit economics from the Stage 3 work.

A

Starting state (verified, not assumed)

Two customers at $42K ACV each = $84K starting ARR. Business model locked: $30K annual SaaS platform fee + $12K one-time hardware fee per site (hardware recognized in year 1 ACV for simplicity). Gross margin: 61% blended (software component at ~80% GM, hardware at ~25% GM). CAC: $18K measured. LTV at current churn assumptions: approximately $165K. Payback: ~8.5 months on blended first-year economics (CAC $18K ÷ monthly gross profit ~$2.1K). These are not model assumptions — they are measurements. The model's job is to project what happens if the same unit economics hold (base) or change (bear/bull).

B

The driver decisions that matter most for Forge

Working through sensitivity analysis on Forge's 8 drivers produces a clear hierarchy. New customers per month is the dominant driver in the first 12 months because the customer base is so small that each new customer represents a large percentage increase in ARR. Churn rate becomes the dominant driver in months 12–24 because the accumulated customer base is now large enough that monthly attrition compounds against a meaningful base. ACV is a one-time lever — you set it at contract signature, and expansion (NRR > 100%) is the mechanism for growing it. This ordering has a practical implication: in months 1–12, Forge should optimize sales motion. In months 12–24, Forge should invest in customer success to hold the churn line.

C

Honest bear case reasoning

The Forge bear case is not "what if things go slightly worse." It is: what if the sales cycle in industrial manufacturing proves to be 9–12 months rather than 4–6 months (plausible — procurement, legal, IT security review, site safety approval all add time in heavy industry)? What if Forge can only close 0.5 new customers per month for the first year rather than 1? And what if early churn is 2.5% monthly (real: early customers in industrial SaaS often churn when the champion leaves, when integration proves harder than expected, or when the ROI case does not survive the first annual review). Under these assumptions — 0.5 customers/month, 2.5% monthly churn, otherwise unchanged — Forge reaches approximately $430K ARR at month 24 and burns through the $1.8M deployable before reaching that milestone. That is the risk the seed capital is actually taking. A model that hides this is a disservice to both founders and investors.

D

Gross margin path and what unlocks 75%+

The path from 61% to 75%+ gross margin has specific milestones, not a smooth curve. The software component is already at ~80%. The hardware component (currently 25% GM on a $12K per-site cost) improves with volume: at 20 sites, sensor module COGS should drop from approximately $9K to $6K per unit with volume pricing, moving hardware GM from ~25% to ~50%. The blended margin at 20 customers would be approximately 73%, approaching the 75% target. This is a concrete, verifiable assumption. The model should show a flat gross margin until the volume pricing threshold is reached, then a step up — not a straight-line improvement from 61% to 75% over 24 months.

🔑 The most useful thing a model does for Forge right now

Forge's financial model reveals that the company has two critical dependencies that are not visible in the headline ARR number: (1) the sales velocity assumption (0.5 vs 1.0 vs 2.0 new customers/month) creates a 3–4x range in month-24 ARR, which is far wider than the sensitivity to ACV or gross margin; and (2) the runway math is sensitive enough that a 6-month delay in hitting the sales pace assumption triggers a bridge financing need. Both of these are actionable intelligence. Knowing that sales velocity is the dominant driver, Forge should invest disproportionately in the sales motion — pilot-to-contract conversion, reference customer development, sales enablement — rather than in product features that improve the bull case but do not change the bear case calculus.

WEAKNESSWhere this bites

✗ The model built to support a predetermined conclusion

The failure mode the brief flags is real and identifiable. A model reverse-engineered from a desired conclusion has three reliable signatures: perfectly smooth growth curves (real B2B sales are lumpy), breakeven that lands exactly when the current raise runs out (the model was designed around that constraint), and a bear case that is still comfortable (the "downside" was set at "slightly less good than base"). Experienced investors identify these patterns in 90 seconds. The tell is not the numbers — it is the shape of the numbers. Smooth curves and round-number milestones signal that the model was drawn, not computed. The antidote is to build from honest drivers and then live with the uncomfortable outputs, rather than choosing drivers to produce comfortable outputs.

✗ The model that is never updated

A financial model is not a one-time artifact for a fundraise — it is an instrument for ongoing decision-making. Yet most founders update their model once (for the pitch deck), then ignore it for 12 months. By then, the model's assumptions are stale, the actuals have diverged materially, and the model has no relationship to the real business. The practice is: every month, enter actuals into the model alongside projections. The variance between actual and projected is management information. If you are consistently missing the new customer assumption by 30%, that is a signal to either fix the sales process or update the model — not to ignore the gap. A model that is never compared to actuals is not a management tool. It is a document.

✗ Conflating the model's precision with the estimate's accuracy

A model that outputs "$4,872,316 ARR at month 24" feels more authoritative than one that outputs "approximately $4.9M ARR." It is not. The precision of the arithmetic is a property of the spreadsheet, not of the underlying estimates. Presenting a model to investors with cents-level precision on a 24-month revenue estimate implies a level of accuracy that no early-stage model can have. Sophisticated investors read this as a calibration signal — a founder who does not understand the difference between precision and accuracy is a founder who does not understand what their model is and is not telling them. Round numbers are honest. False precision is a flag.

KEY INSIGHTHow this connects

  • 3.5 · Contract Structure — ARR, NRR, and the Terms That Define Your Business — The ARR waterfall in this lesson's model is built on contract structure. Annual contracts produce the clean churn math the model depends on; month-to-month contracts make ARR projections far noisier. The expansion rate driver (NRR > 100%) only applies in the contract structures that allow mid-term expansion — which is why contract design precedes financial modeling.
  • 3.6 · Gross Margin — The Number That Controls Your Destiny — Gross margin is one of the 8 key drivers in this model, but it is not a free variable. The gross margin driver is constrained by the cost structure, pricing, and delivery model decisions made in Lesson 3.6. The path from 61% to 75%+ gross margin described in the Forge section of this lesson is the direct consequence of the volume threshold analysis from 3.6.
  • 3.7 · Services Revenue — The Trap and the Bridge — Services revenue appears in financial models as a seductively large early-stage number that often misleads projection logic. Lesson 3.7's warning — that services revenue can mask product weakness, inflate gross revenue while depressing gross margin, and create false ARR metrics — is directly applicable to model construction. If your model uses services revenue as a driver, it must be modeled separately with its own cost structure.
  • 3.9 · Cash Flow and Runway — The Existential Numbers → — This lesson built the 24-month P&L and ARR waterfall. The next lesson focuses on the cash flow statement and runway mechanics — specifically the gap between P&L burn and cash burn, the impact of customer payment timing, and how to read the cash runway indicator as a decision instrument rather than a number to manage appearances.

Financial modeling is the synthesis of everything built in Stage 3: the pricing anchored to EVC (Lesson 3.2), the contract structure that defines ARR (Lesson 3.5), the gross margin architecture (Lesson 3.6), the services revenue decisions (Lesson 3.7), and the unit economics (CAC, LTV, payback) measured across Lessons 3.3–3.4. A driver-based model is where all of those decisions converge into a single instrument for testing the coherence of the whole design. If the model does not close — if honest inputs produce an implausible runway — that is not a modeling problem. It is a business model problem that the modeling has made visible.

GLOSSARYKey terms

Driver-based model
A financial model in which all outputs are arithmetic functions of a small set of explicit input assumptions ("drivers"). Changing a driver propagates correctly through all downstream calculations. The opposite of a model where outputs are manually entered to achieve desired results.
ARR waterfall
A period-by-period accounting of how ARR changes: starting ARR + new ARR from new customers + expansion ARR from existing customers − churned ARR = ending ARR. The waterfall makes the components of ARR change visible, which is essential for understanding whether growth is driven by new customers, retention, or expansion.
Sensitivity analysis
The practice of changing one driver at a time by a fixed percentage (e.g., ±20%) to measure its impact on the key output (typically month-24 ARR or runway to zero). Sensitivity analysis reveals which drivers dominate the model — and therefore where management attention should be concentrated.
Three-scenario structure (bear / base / bull)
A modeling convention in which each driver is assigned three values: a downside (bear), a best-honest-estimate (base), and an upside (bull). Scenarios must differ in their assumptions, not just scale the base case outputs by a constant. A real bear case models a different state of the world, not just a smaller version of the base case.
Runway to zero
The number of months until the company exhausts its available cash at the current burn rate, assuming no new revenue or fundraising. The most consequential single number in an early-stage financial model. Calculated as deployable cash divided by average monthly net burn (negative EBITDA adjusted for actual cash timing).
Fully-loaded cost per head
The total annualized cost of one employee, including base salary, payroll taxes, benefits, equipment, software seats, and any performance-based compensation. For US companies in 2024, the fully-loaded cost is typically 1.25–1.4× base salary. Using base salary in a model systematically understates the cost of headcount growth.
Reverse-engineered model
A model built by choosing a desired output first (e.g., "reach $5M ARR by month 30") and then working backward to select input assumptions that produce it. Identifiable by perfectly smooth growth curves, breakeven that coincides with the end of the raise, and a comfortable bear case. The opposite of a driver-based model built from honest inputs.

ACTIVE RECALLCheck your understanding

Answer in your own words first, then reveal.

1A founder shows you a financial model where revenue grows at exactly 15% month-over-month for 24 months. What does this tell you about how the model was built, and what would you want to see instead?›

Exactly 15% every month means the revenue line was drawn using a growth rate assumption rather than computed from underlying drivers (new customers, ACV, churn, expansion). Real B2B SaaS growth is lumpy — deals cluster around quarter-end and budget cycles, sales ramp takes time, and churn is not perfectly uniform. A smooth curve is almost certain evidence that the model was built top-down from a desired endpoint, not bottom-up from business mechanics. What you want to see instead: a driver-based model where monthly revenue is the sum of (existing customers × ACV/12 × retention rate) + (new customers this month × ACV/12) + expansion, where new customer additions are themselves a driver with variation rather than a constant. The resulting revenue curve will be lumpy and irregular — and that irregularity is a feature, not a bug. It means the model is honest about how the business actually generates revenue.

2In Forge's driver sensitivity analysis, why does churn rate become more important than new customer pace as the model moves from month 6 to month 24?›

In months 1–6, Forge has a very small customer base (starting at 2 customers). Each new customer added represents a large percentage increase in ARR — new customer pace dominates because the base is small. By month 12–18, Forge has accumulated 10–20 customers. Monthly churn of 1.5% vs. 3% now applies to a much larger ARR base. At 20 customers × $42K = $840K ARR, a 1.5% monthly churn difference is $12.6K per month, or $151K per year in compounding retained revenue difference. This is far larger than the impact of one additional new customer ($3.5K MRR). The compounding effect of churn means it becomes the dominant driver once the accumulated base is large enough. The practical implication: early stage, optimize for customer acquisition to build the base. By month 12, shift investment to customer success to hold the churn rate. If you wait until month 18 to build CS, you are fighting a compounding erosion that becomes very difficult to reverse.

3An investor looks at your model and says: "Your bear case still shows the company reaching profitability by month 22." Why is this a problem, and how should you fix it?›

A bear case that still shows a comfortable outcome means you have modeled a "slightly worse than hoped" scenario, not a genuine downside. Real bear cases involve material difficulty: the possibility of running out of runway before the next milestone, or needing a bridge round to survive. If your bear case reaches profitability by month 22, it is essentially a validation scenario, not a stress test. To fix it: revisit the bear case assumptions and ask "what would have to be true for this to be genuinely hard?" For Forge, that means modeling 0.5 new customers per month (not 0.7), 2.5% monthly churn (not 2%), and fully-loaded headcount costs that are 10–15% higher than expected. Run that combination honestly. If the resulting model shows the company running out of cash at month 18–20, that is the real bear case — and it is essential information for deciding how aggressively to deploy capital and how tightly to manage burn in the early months. A bear case should create urgency, not comfort.

4What is the difference between a model that is "wrong" in a useful way versus wrong in a damaging way? Use a specific example from the Forge model.›

A model is wrong in a useful way when its errors are visible, bounded, and informative about what to track. For example: Forge's base case assumes 1 new customer per month. This assumption will be wrong — some months will be 2, some will be 0. But the direction of error is bounded (you cannot have negative new customers) and the sensitivity analysis shows that the month-24 ARR impact of being off by ±0.5 customers/month is known. That is a useful error — it tells Forge that they need a pipeline that supports 1–1.5 new customers per month and that the model is robust to moderate variance in that assumption. A model is wrong in a damaging way when the error is hidden and systematically biased. For example: if Forge's model treats its $12K one-time hardware fee as recurring ARR (i.e., counts it toward NRR), it inflates the ARR metric in a way that corrupts every downstream calculation — runway, LTV, cohort retention. This error is not visible in the output, it is not bounded in its impact, and it leads to decisions (raising a larger round at a valuation that depends on the inflated ARR) that are structurally unsound. The difference: useful errors are explicit, bounded, and drive calibration. Damaging errors are hidden, unbounded, and drive mis-investment.

Next: cash flow and runway — the existential numbers →