Skip to content
finalstate.aiWhen Uncertainty Pays You
002Atlas

002TWO BETS

When Uncertainty Pays You

A studio has ten free weekends: one fixed-fee retainer, or ten small software experiments with repeat-sales rights. The terms change what surprise can cost and what success can earn.

Ten weekend tokens committed together to fixed-fee work, or separately to ten software experiments.The same development budget appears twice as a comparison, not twenty available weekends. The retainer bundles the tokens under a fixed fee while revisions can continue. The tools keep separate budgets and retained repeat-sales rights. The sale branch is a possibility, not an observed winner or a probability.Fixed feeRepeat salesRevisions
Illustrative case. Each tool has a one-weekend development budget; support, selling costs and other obligations still need to be counted. Ten attempts imply no success rate.

JENSEN

A convex payoff stays below its chords

Equal-weight states separate around an unchanged mean on the convex payoff f(x)=x².A straight chord joins the two endpoint payoffs. Its midpoint is their average payoff, E[f(X)], above the curve value f(μ) at the fixed mean μ. The accent marks this Jensen gap. Equality remains possible for convex payoffs on a straight interval; the quadratic example has a positive gap once its states differ. Payoff½ per stateState μ E[f(X)] f(μ)
Jensen, Acta Mathematica 30 (1906), pp. 176, 179–180, equations (1), (4)–(5), doi:10.1007/BF02418571. Modern weak-convexity notation; finite expectations and a stated domain.

The curvature test, developed in Jensen's paper, uses the same weights for the states and their payoffs. Here two equally likely states move apart around a fixed mean. This is a mathematical example, not a fitted studio payoff. Capped loss and open upside are neither necessary nor sufficient for convexity.

  • For 0 ≤ λ ≤ 1: f(λx + (1−λ)y) ≤ λf(x) + (1−λ)f(y)
  • The inequality permits equality; a positive gap requires a bend across the compared states

Terms, curvature and price answer different questions

A loss cap and room for repeat sales describe practical asymmetry. Curvature is a separate property of the payoff over a stated domain. Reliable probabilities help price that exposure. To compare distributions for all convex payoffs, use a mean-preserving spread in convex order, with finite expectations; greater variance alone is insufficient.

THE BARBELL

Protect the reserve; limit the experiments

A large reserve remains behind a boundary while ten weekend tokens move into a separate experiment budget.The reserve is deliberately uncounted; its drawn area is schematic, not an allocation recommendation. The ten tokens repeat the illustrative development budget. Their separation is conditional on contracts and shared liabilities actually leaving the reserve unexposed.ReserveExperiments
Taleb, Antifragile (2012), chapter 11, “Seneca’s Barbell”; author-hosted Appendix I, pp. 438–441. Allocation heuristic, not an optimal percentage or a guarantee of safety.
  • Keep the resources needed for existing obligations low-risk and available
  • Set an experiment budget whose complete loss is survivable
  • Look for shared liabilities that could reach the reserve

Taleb's barbell separates a protected bulk from speculative exposure and avoids a middle that hides large downside. Applied to the studio, the ten weekends must sit inside an affordable experiment budget. Several tools can fail together; separate project names do not separate liability.

AFFORDABLE LOSS

Count what total failure can reach

A guarantee extends exposure below a stated loss cap.Two payoff sketches remain fixed while an additional guarantee appears on the right. Once the guarantee is identified as exceeding the listed stake, the stated boundary opens and an exposure arrow crosses it. The comparison questions the loss limit; it does not infer whether either payoff is mathematically convex.Stated capCapGuaranteeExposure
Sarasvathy, Academy of Management Review 26(2), 2001, p. 252, principle 1. The exposure inventory and survivability question here are Atlas applications, not a validated diagnostic from the paper.
  • Include cash, time, attention, guarantees and material reputational exposure
  • Check what obligations and capacity remain after total failure
  • Review upside ceilings separately from the loss limit

Affordable loss starts with a predetermined commitment limit. Sarasvathy states that principle on page 252 of her 2001 paper. A guarantee beyond the studio's listed stake breaks its claimed cap. A capacity limit or partner veto restricts upside instead. Neither observation settles curvature.

RISK AND UNCERTAINTY

A convex payoff does not require unknown odds

Comparable case records form a reference class; an unmatched case has no supported group.The left comparison has records with the same relevant features, enclosed by a common bracket. The right candidate differs and stays outside that bracket. The records are schematic, with no count used to estimate a frequency. This is a distinction in evidence for odds, not a measured risk distribution.Comparable casesUnmatched case
Knight, Risk, Uncertainty and Profit (1921), III.VII, probability and estimates; III.VIII, opening paragraph. Black–Scholes (1973), pp. 640–644, assumptions and equation (13): European exercise, no dividends, constant rate and volatility, frictionless continuous hedging.
  • Knight's distinction: a reliable reference class can support probabilities
  • Without one, a precise-looking probability can conceal an unsupported estimate

Knightian uncertainty concerns the missing basis for a reliable probability. It does not create convexity. In the Black–Scholes model, greater volatility raises a European call's value with other inputs fixed. That result depends on the model's assumptions; it is not a general pricing rule for software ventures.

BRING A VERIFIER

Price the proof into the bet

A checking-cost portion separates from a possible-proceeds ticket.A detachable portion marked by the recurring weekend token moves away from a possible-proceeds ticket. Its movement makes a cost deduction visible without numerical bars or an assumed net return. The remaining proceeds are still conditional on a successful result and an adequate check.Possible proceedsChecking
Schematic deduction, not a measured margin. The weekend token marks time required for checking; the size of the removed portion specifies no cost share or success probability.
  • Include money and time spent establishing the result
  • Allow for a check that fails or arrives after the deadline

A check can reject the result as well as consume its value. Deduct the full cost of verification when judging the opportunity. If proving the result costs more than the result is worth, the curve may be convex and the trade still wrong.

Attention belongs in the carrying budget

In Bowman and Hurry's option account, recognition and securing access are distinct. A candidate opportunity may require further investment before a usable right exists. For the studio, account for that work and for monitoring an existing right before expiry. Recognition alone grants no new entitlement. Existing resources may already support an option; a new purchase is not a universal requirement.

WHEN IT PAYS

Leave the decision with the owner

A weekend token accompanies three unanswered checks: survive, upside and carry.A commitment record shows empty answer boxes for survivability, upside limits and carrying costs. The owner's decision line remains blank. The figure is a record of questions, not a scored instrument or permission to proceed; payoff curvature and price require separate judgment.SurviveUpsideCarryOwner
Atlas synthesis. Unfilled checks preserve the owner’s responsibility; the record supplies no odds, curvature proof or automatic approval.
  • Survivable: can the studio meet its obligations after every tool fails?
  • Open: what limits repeat sales after capacity constraints and vetoes?
  • Holdable: can checking and monitoring fit inside the commitment budget?

These questions are an Atlas synthesis for assessing practical asymmetry. Passing them establishes neither convexity nor a worthwhile price. The studio owner still has to test the payoff and judge the available evidence. Where a system cannot support its estimate, abstention must remain available.

Read the transcript

01 · TWO BETS

A freelance studio has ten free weekends. This is an illustrative choice. It can pour all ten into a year-long, fixed-fee retainer. The fee cannot rise, but the revisions can. Or it can spend one weekend on each of ten tiny software tools and keep the right to sell every copy. Each tool has its own development budget. Nine could fail and one could keep selling without rebuilding the other nine. That is a possible outcome, not a forecast or a one-in-ten hit rate. Support, selling costs and other obligations still have to fit. The starting development budget is the same; the terms give surprise a different effect.

02 · JENSEN

Across the bottom runs the state of the world; up the side, the payoff. This mathematical example uses the square of the state, not an estimate of studio returns. Take two equally likely states and pull them apart while keeping their average fixed. The average endpoint payoff sits above the payoff at that average state. The distance between them is the Jensen gap. On a straight interval the gap would be zero. Jensen published the inequality in 1906. More generally, choose any two states and weights between zero and one that sum to one. The payoff at their weighted-average state cannot exceed the same weighted average of their payoffs. That chord inequality must hold across the relevant domain. Finite expectations are assumed when applying it to distributions.

03 · SHAPE, NOT ODDS

The studio still needs a judgment about demand. When a reference class is reliable, use it to price the bet. The terms do a different job. Ask what this can cost, including the price and the waiting, and what ceiling could stop the win. A capped loss and open upside describe a useful asymmetric profile. They are neither sufficient nor necessary for mathematical convexity. The chord tests the shape. A larger variance by itself is insufficient to rank distributions for all convex payoffs. The distribution comparison requires a mean-preserving spread in convex order: all convex payoffs have at least as high an expectation, where those expectations exist. A single volatility number does not establish that ordering.

04 · THE BARBELL

Nassim Taleb calls the separation of protected resources and speculative exposure a barbell. Keep the resources needed for existing obligations low-risk and available. Set aside only an experiment budget you can lose completely. The studio can then divide that budget among its ten tools, keeping the reserve outside their exposure. Taleb warns against the middle that looks steady while concealing a ruinous break. This is an allocation heuristic, not a prescribed percentage. Separate projects can still fail together. A shared guarantee can also carry their losses into the reserve. The owner has to establish the separation; drawing a boundary does not create one.

05 · AFFORDABLE LOSS

Saras Sarasvathy describes an affordable-loss principle. On page two hundred fifty-two of her 2001 article, she describes predetermining how much loss is affordable. For the studio, the relevant question is what remains after complete failure. Count cash, time, attention and material reputational exposure, as well as explicit obligations. A guarantee can add an obligation that was absent from the listed stake. If that obligation extends beyond the stake, the stated cap no longer holds. A capacity limit or a partner's veto does different damage: it cuts off the upside. The owner must establish both boundaries. Affordable-loss sizing does not establish convexity.

06 · RISK AND UNCERTAINTY

A convex payoff does not require unknown odds. In the Black–Scholes model, a European call gains value when volatility rises and the other inputs stay fixed. That result assumes no dividends, constant volatility and interest rates, and frictionless continuous hedging. Software ventures do not inherit those conditions. Frank Knight distinguished probabilities supported by calculation or a suitable class of cases from estimates about situations that lack such a basis. A reference class earns its place through relevant similarity, not by containing a large number of records. Where that basis is missing, convexity cannot manufacture odds. The studio can still limit its commitment, examine the payoff and seek better evidence.

07 · Advertisement · Bubble AI App Builder

Some ideas do not need another document before they become testable. With Bubble AI, you describe the app you want, and Bubble creates a working starting point: interface, data, and logic you can inspect. From there, you refine visually, connect AI models and services, and turn the first version into something real enough to use.

08 · BRING A VERIFIER

Count the cost of collecting the win. A possible upside is not the same as a collectible one. A proof may check a construction. A lab may measure a binding. A registry may confirm a filing. Each check costs money and time, and the expiry clock keeps moving. Charge that checking work against the possible proceeds before valuing the opportunity. The check might also fail, or finish too late. If proving the result costs more than the result is worth, the curve may be convex and the trade still wrong. For the studio, a weekend spent verifying a tool's result is a weekend unavailable for another attempt. That time belongs on the bill.

09 · SEEN FIRST

Attention belongs on that bill too. Bowman and Hurry's option account distinguishes recognising a candidate from securing access to it. Depending on what the studio already holds, further investment may be needed before it has a usable right. Recognition alone grants no new entitlement. Once a right exists, it can be cheap to own and still expire before anyone acts. Monitoring and exercise have costs. If the position needs a weekly watch, a filing, or a decision before a deadline, fund that work when you size it. Otherwise the nominally affordable development budget understates the commitment.

10 · WHEN IT PAYS

Go back to the studio's ten weekends. The owner has three practical questions to answer. Survivable: if every tool fails, can the studio still meet its obligations? Open: what limits repeat sales after capacity constraints and partner vetoes? Holdable: can checking, monitoring and any deadlines fit inside the commitment budget? These questions are the Atlas's synthesis for assessing practical asymmetry. They leave the answers open. Even favourable answers prove neither convexity nor a good price. The owner still needs a payoff over a stated domain and a judgment about the evidence. A system that cannot support its estimate must be able to abstain.

01 / 10 · TWO BETS0:00 / 7:25