Shadow Futures

Methodology and scope

How the argument works

The animation uses familiar self-reinforcement to show the setup. The theorem asks a different question: when that process closes off other paths, does one realized market still contain enough comparison to recover contribution? This page shows the shortest route from the familiar mechanism to the paper’s distinct result.

Read the source paperSee the central equation
  1. 01The market chooses

    Contribution-related inputs and accumulated advantage shape who wins next.

  2. 02The contest can close

    An early favorite may receive nearly all later opportunities.

  3. 03Evidence can run out

    More activity may repeat the lead without meaningfully testing anyone else.

01 / Simulation

What the app shows

The general model gives each person or firm a score. That score combines a verified input with advantage already accumulated. The market turns the scores into the probabilities of receiving the next opportunity.

Who receives the next opportunity
pit(β)=exp⁡(βxi)(a+Ni(t))ρ∑jexp⁡(βxj)(a+Nj(t))ρp_{it}(\beta)=\frac{\exp(\beta x_i)(a+N_i(t))^\rho}{\sum_j\exp(\beta x_j)(a+N_j(t))^\rho}

Contribution-related signal × accumulated advantage → chance of winning next.

The story example

Twenty-four creators with modeled audience-response multipliers from 0.84 to 1.18 compete for 1,600 recommendations. The feedback strength is fixed at 1.55, so both creator differences and accumulated exposure affect the next ranking.

From creators to firms

For a firm, the accumulated advantage might be customers, contracts, an installed base or past sales. The exact measure must match the real market.

Fixed random seeds make every replay reproducible. The numbers illustrate the mechanism; they aren’t forecasts for a real platform or industry.

02 / Evidence

What counts as a real comparison

A recommendation, contract or sale is informative only when more than one competitor has a meaningful chance. If the favorite is almost certain to win, the market reveals almost nothing about everyone else.

How open the contest remains
εt(β)=1−max⁡ipit(β),BT(β)=∑t=0T−1εt(β)\varepsilon_t(\beta)=1-\max_i p_{it}(\beta),\qquad B_T(\beta)=\sum_{t=0}^{T-1}\varepsilon_t(\beta)

εt\varepsilon_t is the chance left for a competitor other than the favorite. BTB_T adds those chances over time.

Why this limits information
tr⁡It(β)≤DX2εt(β)\operatorname{tr}I_t(\beta)\le D_X^2\varepsilon_t(\beta)

As the chance left for everyone else approaches zero, the new information about contribution must also approach zero.

03 / The theorem

What the paper proves

Under the formal conditions below, finite total comparison means that one complete history can’t support a method that consistently learns every nonconstant measure of contribution.

The result in one line
B∞(β)<∞ for every β⟹Pβ∼Pβ′⟹ no universal consistent recovery of F(β)\begin{aligned} B_\infty(\beta)<\infty\ \text{for every }\beta &\Longrightarrow \mathbb P_\beta\sim\mathbb P_{\beta'} \\[4pt] &\Longrightarrow\ \text{no universal consistent recovery of }F(\beta) \end{aligned}

The histories aren’t identical. They overlap too much for one realized history to identify every contribution measure consistently.

Formal conditions and boundaries
  • The design is common and predictable from the same observed past.
  • Nearby parameter values have locally equivalent one-step laws.
  • Hellinger separation in both directions is controlled by the remaining comparison.
  • Total comparison is finite under every parameter being compared.
  • Any additional observed process with parameter-dependent information must be included.
  • The conclusion is mutual absolute continuity of complete-history laws, not equality of distributions.
04 / Design

What can keep learning alive

The result isn’t a law of nature. Market and platform design can preserve new opportunities to compare people and firms.

Give newcomers real exposure

Random discovery prevents the current favorite from becoming nearly certain.

Create independent starts

Resets and separate channels produce evidence that one continuous ranking can’t.

Let people and firms reach buyers elsewhere

Portability, open standards and multihoming stop one platform or distribution channel from becoming the only record.

Limit control over discovery

Public options and structural separation can create genuinely different paths.

Random exposure is an intervention: it changes the platform rule rather than merely measuring the original one.

05 / Policy

What the result changes

The economic conclusion

Market rankings can’t by themselves settle moral or political questions about desert. They don’t reveal a clean earned-versus-unearned split.

The democratic conclusion

Tax rates, public ownership, UBI and social dividends remain collective choices. They should be decided openly around power, security, freedom and shared prosperity, not outsourced to a market score.