Token reaping made the tournament market state endogenous

DX Terminal Pro's market periodically reaped its lowest-capitalization token until one graduated to public trading. That rule made the available market state depend on the population's own trading outcomes. We should carry the mechanic into any interpretation of behavior in the twelve-token tournament.

The controls paper companion reports a 21-day real-capital event on Base under one frozen prompt and harness with one model family. Twelve tokens launched into Uniswap V4 pools at genesis. As tokens were removed, the available choices and survival incentives changed. The same rendered instruction could therefore operate in a materially different choice set later in the event.

The denominator moves with survival

An illustrative market begins with twelve eligible tokens and later has six. An agent concentrated in one token represents one of twelve possible choices early and one of six later. A uniform-choice reference would move from about 8.3% to 16.7% per token. These are hypothetical availability calculations, not measured selection rates in the tournament.

The example shows why a single fixed token-availability baseline could misread increasing concentration as increasing model preference. The market rules themselves narrow the choice set. Population trades can also influence capitalization, which affects the next reaping outcome. Selection and subsequent opportunity are connected through that feedback.

A concrete state artifact would record the eligible token set at each decision, the next scheduled reap information actually available, capitalization timestamps and the relevant policy text. A point-in-time evaluation would reconstruct that state rather than exposing a model to the final survivor list. The proposed artifact tests information validity; it would still need a separate design to estimate how the rule affected profits.

Context affected pre-launch behavior

The paper companion reports that inserting the reap mechanic as structured context with payoff order leading moved capital deployment in an affected pre-launch test population from 42.9% to 78%. That is an increase of 35.1 percentage points, or roughly 1.8 times the earlier share. Exact per-arm counts are absent from the published summary, and the result belongs to that prompt-side intervention.

We can infer that representation of the mechanic changed behavior in those tests. We cannot convert the deployment share into an investment-return gain or isolate payoff order from every other element of the intervention. Capital deployment measures the amount put to work under the test contract, while profitable timing would require outcomes and costs.

The continuous record companion contrasts this bounded tournament with an open-ended perpetuals fleet. That distinction matters for transfer: an asset elimination tournament creates incentives unlike a normal persistent instrument universe. A behavioral rule learned inside the tournament deserves a transfer test under the new choice set.

Readers should therefore ask whether an apparent trading pattern follows the model, the mandate or the evolving tournament rules. A frozen runtime holds software constant; it does not freeze the market that software observes.