I froze 1,220 tokens as they approached the pump.fun graduation line and asked a narrow question: if you ignored the race to execute, was there a return? There was — a counterfactual buy that gets filled for free shows a median gross return between +8.5% and +17.45%. Visible, in prices, on the same tokens anyone could watch.
Then I put the execution back in. Access to that return disappears within seconds: 69.0% of the tokens are already unreachable at the earliest resolvable state, and 89.7% by eight seconds. I reconstructed that collapse exactly from graduation timing — it isn't a modeling artifact, it's the mechanism.
So I gave a fixed policy every advantage I could and still asked whether it could win. Zero cost. Entry-state pricing. A granted counterfactual fill. Under that deliberately optimistic bound — favorable whenever the curve doesn't retrace below its entry state inside the unobserved arrival window — the upper bound on expected gross return per token is negative at every latency I can observe, and its upper 95% confidence limit (hour-cluster bootstrap) stays below zero at every step.
The negative sign isn't latency alone. Reaching more of the population fills more of a subset whose realized economics are worse — an adverse composition I observe, not a selection mechanism I claim to have proven. That distinction is the whole discipline of the paper: observed, accessible, and realizable returns are three different numbers here, and I keep them apart.
What I am not saying: I don't identify who got the good tokens. I make no priority or novelty claim — execution-access effects and AMM adverse selection (loss-versus-rebalancing) are established prior art. I don't claim profit is impossible in general. One policy, one roughly two-day window, one venue. The bound is conditional and one-sided by construction, and a confidence interval below zero does not rescue the one assumption it can't see — the unobserved path between arrival and fill.
This is the same discipline the rest of this site runs on, pointed at a different domain: everything reproduces from one frozen pull of public on-chain data, every figure and number resolves to a hash, and the record carries the paper, the data, the code, the preregistration, and the manifests. You don't trust the analyst; you trust the trail.
→ The paper + reproduction packet (Zenodo, CC BY) · the methodology it inherits
Method note: language-model tools assisted with prose and analysis-code generation under author-defined estimands and reporting constraints; I reviewed and take responsibility for every analysis and claim.