# Do Two On-Chain Observation Pipelines See the Same Tokens? Cross-Pipeline Coverage on the Solana pump.fun Launchpad

Crypto.world

> Studies of memecoin launchpads typically treat the token list produced by one data-collection pipeline as if it were the market itself, without checking whether a differently configured pipeline…

**arXiv**: 2609.18975

**Published**: 2026-07-16

**Authors**: Arati Uday Kamat

## AI summary

## Overview

**Research area:** Empirical market microstructure and on-chain measurement methodology, applied to memecoin launchpads on Solana (specifically pump.fun).

**Technical level:** Intermediate. The statistical content is deliberately minimal (counts, proportions, Jaccard and Dice indices), but readers need familiarity with blockchain data collection, mint addresses and timestamp-gated joins.

**Scope:** A descriptive comparison of the mint sets produced by two separately configured observation pipelines run over two consecutive windows on the Solana pump.fun launchpad, measuring how many tokens the two pipelines actually share.

## What This Paper Is About

Studies of memecoin launchpads typically treat the token list produced by one data-collection pipeline as if it were the market itself, without checking whether a differently configured pipeline would have seen the same tokens. This paper puts two pipelines from a single research programme side by side: one that detects coordinated early-buyer wallet cohorts, and one that logs a trader-side observer's pre-trade rejection decisions. The goal is a purely descriptive answer to one question: over the same time window, do the two pipelines observe the same mints?

## Key Contributions

1. **A first measurement of cross-pipeline token overlap on pump.fun.** The paper reports raw intersection counts, observed proportions on each side, and Jaccard/Dice indices for two consecutive, non-overlapping windows, together with a sensitivity analysis over four timestamp rules for aligning the two data streams.
2. **A documented, quantified collection gap and its effect on inference.** The rejection stream contains no events between 3 June 2026 (06:35 UTC) and 24 June 2026 (16:34 UTC), a gap the public dataset's own documentation (version v2.2) does not mention; the paper shows how this single gap renders one v1 proportion uninformative.
3. **A proposed routine reporting check for on-chain studies.** Report (a) the overlap between collectors under a stated timestamp gate and (b) each collector's own coverage of the analysis window, before making whole-market statements.
4. **An explicitly non-inferential statistical framing plus an open reproduction package.** Because inclusion in each pipeline is deterministic given collector configuration, no confidence intervals or significance tests are reported; a standard-library Python script re-derives every number from four SHA-256-pinned input files.

## Main Findings

- **The two pipelines see nearly disjoint token sets in the fully covered window.** In v2 (17.534 days, 29 June to 16 July 2026), the cohort set contains 623 mints and the rejection set 1,742, with 5 overlapping mints. That is 0.803% of the cohort side and 0.287% of the rejection side, on a union of 2,360 mints, with a Jaccard index of 0.002119 and a Dice coefficient of 0.004228.
- **The earlier window points the same direction but carries little weight.** In v1 (13.376 days, 11 to 25 June 2026), 20,162 cohort mints and 53 strict-window rejection mints share 1 mint: 0.005% of the cohort side and 1.887% of the rejection side, on a union of 20,214, with Jaccard 0.000049 and Dice 0.000099. The rejection stream covers only the final 14.25 hours of the 321.03-hour cohort window, so the cohort-side figure mainly reflects that gap rather than cross-pipeline coverage.
- **A single undocumented collection gap drives the v1 result.** No rejection events exist between 3 June 2026 (06:35 UTC) and 24 June 2026 (16:34 UTC). Restricting both mint sets to the common-coverage interval (24 June 16:34 UTC to 25 June 06:49 UTC, 14.25 hours, 4.44% coverage, longest internal gap 0.58 hours) leaves 1,113 cohort mints, the same 53 rejection mints and zero overlap. The restriction drops 19,049 cohort mints detected before the interval, including the sole v1 overlap mint (`5qjZ...pump`), which the cohort pipeline detected on 23 June and the rejection stream first rejected 25.9 hours later. At the v2 rejection-side rate of 0.287%, the expected overlap among 53 rejection mints would be about 0.15 mints, so an observed zero neither supports nor contradicts v2.
- **The result is insensitive to the timestamp rule chosen.** Four rules are tested: Spec A (`rejectTs` in window), Spec B (`sampleTs` in window), Spec C (either), Spec D (both). On v1 all four give 53 rejection mints and 1 overlap. On v2 the rejection mint count ranges only from 1,742 (Specs A and D) to 1,749 (Specs B and C), with the overlap fixed at 5.
- **Widening the extraction interval adds one overlap in v1 only.** The v1 file was extracted 17h 10m 35s beyond the strict window; at file level the counts become 83 rejection mints and 2 overlaps (2.410%). The recovered mint is `AvxF...pump`, whose earliest rejection timestamp (2026-06-25T19:49:28Z) falls 13h 00m 03s past the upper edge of the strict cohort window. On v2, strict versus file-level counts differ by at most 14 mints (1,742 to 1,749 versus 1,756), with the overlap fixed at 5.
- **The two cohort windows share no mints at all.** The v1 and v2 cohort intersets have an intersection of 0, which is expected because the v2 window begins 3d 20h after v1 ends and cohort detection is a per-launch signature that resets across windows. The paper states this says nothing about whether the underlying wallet cohorts persist.
- **Combining the pipelines materially expands what is observed.** In v2 the union of 2,360 mints is roughly 35% larger than the larger rejection stream alone (2,360 / 1,742 = 1.355) and about 3.8 times the smaller cohort stream (2,360 / 623 = 3.79).
- **The rejection stream is predominantly, but not exclusively, post-graduation.** Of 238,722 v2 rejection rows, 92.8% carry `dexId=pumpswap` (221,540 rows, 1,666 unique mints) and only 2.1% carry `dexId=pumpfun` (5,040 rows, 41 mints); 4,995 rows (2.1%, 1,612 unique mints) have no `dexId`, and the remainder split between meteora (3,684 rows, 1.5%, 41 mints) and raydium (3,463 rows, 1.5%, 29 mints).
- **No confidence intervals and no significance claims are made.** The paper argues that an observed mint is not a Bernoulli draw from an identifiable population, so the percentages are observed proportions of observed sets rather than estimates of a population parameter.

## Methodology in Plain English

The author takes two existing data streams from her own research programme and asks how much they agree. The first, the cohort-detection pipeline, works in two stages: it extracts the first ten buyers of every qualifying launch, then builds a co-occurrence graph across launches, filters edges by weight and surfaces persistent wallet groups using union-find. The second, the rejection-filtering pipeline, comes from a passive trader-side observer that scores incoming candidate mints against pre-trade filters (a 24-hour price downtrend, a multi-timeframe fade, a market-capitalisation range, time-of-day rules, a fast-exit rule, and in v2 a seller-dominance filter) and records a `rejectReason` for each mint that fails.

For each of two windows, the author defines a "strict cohort window" running from the earliest to the latest cohort detection timestamp. She then restricts the rejection data to that window under four different timestamp rules, computes the intersection, union, observed proportions and Jaccard/Dice indices, and reports them raw. Because the strict window is defined by cohort detections alone, she separately measures how much of each window the rejection stream was actually recording, and repeats the overlap calculation inside the interval both streams genuinely cover. There is no statistical model, no sampling assumption and no hypothesis test; the only inferential content is a direction-of-finding statement across two windows.

## Why This Matters

For research, the paper makes an accounting point that applies to any study built from a single on-chain collector: a token set describes the collector's configuration as much as it describes the market. Because inclusion in a pipeline is deterministic given the subscription, filters and window, findings about "memecoins" or "launches" should name the collector, and multi-collector overlap plus per-collector window coverage should be reported before whole-market claims. The paper also demonstrates how a single undocumented collection gap can silently invalidate a proportion.

Real-world applications:

- **Data due diligence for quantitative funds and trading desks.** Before backtesting a launchpad strategy, check whether the vendor or in-house feed actually observes the tokens the strategy trades, and how much of the window the feed was live.
- **Vendor and feed comparison.** The overlap-and-coverage check is cheap to compute and gives a concrete, reproducible way to compare two market-data providers on the same event stream.
- **Risk and scam screening.** Analysts building rug-pull or manipulation detectors need to know whether their candidate set is a lifecycle-stage slice (pre-graduation buyer behaviour versus post-graduation tradability) rather than the whole launch universe.
- **Replication of published on-chain studies.** Reviewers and replicators can apply the same reporting standard to check whether a published result depends on one collector's configuration, echoing Menkveld et al.'s finding (across 164 research teams testing the same hypotheses on the same data) that non-standard errors are sizeable relative to standard errors.

Industry relevance centres on data infrastructure for decentralised exchanges and launchpads: the same subscription-and-filter architecture described here underlies most commercial on-chain analytics products, and the paper quantifies how much two such configurations can diverge.

## Future Directions

1. **Measure each pipeline's recall against the full pump.fun launch universe.** The paper explicitly does not measure completeness, noting this would require a whole-launchpad collection of the kind used in large-scale manipulation studies.
2. **Model the observation-selection mechanism so that formal coverage estimation becomes possible.** The author notes that capture-recapture and dual-frame reweighting require a plausible source-inclusion model, which the current deterministic framing deliberately avoids.
3. **Test which of the four proposed mechanisms actually drives the low overlap.** Lifecycle stage, filter surface, set-size asymmetry and (v1 only) coverage are offered as plausible descriptive explanations, but the paper states that apart from coverage none is measured directly.
4. **Extend beyond two pipelines from one programme to genuinely independent collectors and teams.** The current comparison shares collector infrastructure, licence and author, so the two-window agreement is a direction-level statement rather than an independent replication.

## Target Audience

Researchers and practitioners who build or consume on-chain datasets for memecoin launchpads and decentralised exchanges: empirical finance and market-microstructure academics, quant researchers and data engineers at crypto funds, on-chain analytics vendors, and reviewers or replicators of blockchain measurement studies. It is most useful to readers who already work with Solana event data and want a concrete, low-cost reporting standard rather than a new estimator.

## Authors’ abstract

On-chain studies of memecoin launchpads usually rely on one data-collection pipeline, yet whether differently configured pipelines observe the same tokens is rarely measured. This paper compares the output mint sets of two separately configured pipelines from one research programme on the Solana pump.fun launchpad: a cohort-detection pipeline that flags coordinated early-buyer wallets, and a rejection-filtering pipeline that logs a trader-side observer's pre-trade filter…

[The full abstract on arXiv](https://arxiv.org/abs/2609.18975)

[Read the original paper](https://arxiv.org/abs/2609.18975)

---

Canonical: https://crypto.world/paper/do-two-on-chain-observation-pipelines-see-the-same-tokens-cross-2609.18975
