A free trading policy that loses at entry-state prices deserves attention. Yet the execution result advertised here cannot be checked against the supplied body, which studies another question and reports neither returns nor latency. We did not test the execution claim either. Reproduction would require Solana RPC or archive-node infrastructure, along with transaction submission using priority fees or Jito bundles to model queue priority. Price bars cannot test the accessibility constraint or the execution claim.
The proposed trade is real enough. Pump.fun launches tokens on a bonding curve. A small fraction reach the threshold and move to a real pool, while the rest decay. The abstract attached to this record freezes 1,220 tokens as they approach graduation and applies a counterfactual purchase that ignores the execution race. Its median gross return ranges from +8.5% to +17.45%.
Access changes the result. At the earliest resolvable state, 69.0% of the tokens are unreachable; by eight seconds, 89.7% are gone. Even after granting zero cost and entry-state pricing, the optimistic upper bound on expected gross return per token remains negative at every observable latency. An hour-cluster bootstrap leaves the upper 95% confidence limit below zero at every step.
The abstract qualifies that optimism. The bound is favorable whenever the curve does not retrace below its entry state during the unobserved arrival interval. It also proposes a composition account for the negative sign without establishing the mechanism. In its words, the reachable subset has worse economics because of "an adverse composition we observe rather than a selection mechanism we demonstrate."
That result has teeth. Every real trading cost would make a free policy that already loses at its own entry state look worse.
A different paper arrived
Brad M Lindsey is listed as the record's author. The supplied body instead gives Arati Uday Kamat's name, ORCID and contact address under a different title concerning coordinated sniper cohorts. Kamat analyses 1,578,333 buyer events from 166,098 mints with first-ten-buyer coverage. The sample spans 13.38 days, from 11 June to 25 June 2026.
For each launch, Kamat constructs a wallet co-occurrence graph from the first ten buyer events. Edges appearing at least three times are retained, then union-find identifies 1,012 persistent cohorts containing 2,965 addresses. The median cohort has 2 wallets. The largest has nine wallets and appeared in 42 launches over 11 days, with mean first-buyer rank 2.29.
This paper estimates whether a cohort's presence increases buyer flow from everybody else. The naive comparison shows +130.9% more first-30-minute buyer events on cohort-touched launches: 20.97 versus 9.08, based on 5,419 treated launches and 160,679 controls. Removing the cohort's own purchases from the outcome for treated and control launches reduces the lift to +63.9%.
Propensity matching uses ten launch-metadata covariates. After matching, the estimate falls to +16.1%, CI [+13.0%, +19.4%], across 5,419 pairs. Treated launches average 14.71, compared with 12.67 for controls. The corresponding SOL-inflow estimate is +6.3% with CI [-0.5%, +15.1%]. The paper plainly describes that result as indistinguishable from zero. The worst post-match standardised difference is 0.077, compared with a pre-match range of 0.43 to 0.65 for market-cap and social covariates.
None of these estimates addresses execution.
The body contains no 1,220-token sample, eight-second latency grid, return series or trading-cost assumption. This mismatch reaches the research question itself. Kamat planned to estimate the effect on graduation probability through a matched-pair Cox model, but graduation labels covered only a small fraction of the 5,419 treated mints. The sample was too thin for that model. The abstract above the body concerns a graduation race.
Does eight seconds decide the trade?
If the accessibility figures hold, they dominate the result. Losing 69.0% at the earliest resolvable state means the race selects the fill set. At 89.7%, an eight-second response budget leaves roughly one token in ten available.
A trader then needs the joint distribution of fills and outcomes. The abstract concedes the relevant weakness: it observes the composition without modelling the selection mechanism. Separating slow arrival from adverse availability requires a stated policy, timestamped state availability and fill probability at each latency. We could not find any of those three in the supplied text.
What remains after adjustment
Kamat's activity-matched placebo produces a median lift of +189.6% over 100 seeds. It exceeds the real-cohort naive upper CI of +137.4% in 100 of 100 seeds. The paper treats this as a diagnosis of the placebo estimator rather than a null baseline and retains it as a bias diagnostic.
There is another awkward result. Among the 5,419 treated launches, 382 (7.0%) attract zero non-cohort buyers within 30 minutes, versus 0.48% of controls. The paper says this finding cuts against both the picking-winners interpretation and the claim that cohorts generate flow. It also discloses that buyer rows, rather than distinct wallets, define the outcome.
Rosenbaum bounds for the remaining +16.1% estimate are not computed. The 436.6 MB buyer corpus is absent from the deposit and available on request from the corresponding author, with a mirror planned for a future Zenodo release (target 2026-Q4). Consequently, none of the lift estimates can be reproduced from the released files. The naive control means reported in Table 4, buyer 9.08 and SOL 2.33, are arithmetic derivations using the pooled lift and matched-sample treated mean. They are not direct script outputs.
The cohort label also carries a structural limitation for anyone trying to trade it. Membership is inferred from co-occurrence across the complete 13.38-day corpus. A label assigned to a 12 June launch therefore incorporates 24 June behaviour. The paper's estimand does not require a live signal, and Kamat says a cohort touch is not a reliable proxy for future non-cohort interest above a comparable mint. A trader reading the catalogue as an entry signal would still inherit the full-sample construction.
The missing execution evidence
Our price-bar history does not cover the June 2026 window. We previously examined an LLM execution controller at /articles/an-llm-tilts-twap-by-a-basis-point-in-simulation, but the fill model has to come before this bound can be judged.
Four items would resolve the execution question: the written policy, the list of 1,220 tokens and their state timestamps, fill probability at each latency step, and the return reconstruction that converts +8.5% to +17.45% gross into a negative bound. The text filed beneath the abstract supplies none of them.
The defensible causal estimate in this record is the +16.1% buyer-count lift from 5,419 matched pairs. SOL inflow comes in at +6.3%, 95% CI [-0.5%, +15.1%], and is indistinguishable from zero. The buyer-flow result is real, bounded honestly and worth retaining. It cannot verify the execution-alpha claim attached to it.