// Research · 2026-09-11

What lives on a launchpad.

Robinhood Chain produces a new token every five seconds. We watched them for 6.9 hours and then went back to see which ones were still breathing. 90.8% were dead within the hour. 15.6% never traded at all — a mint, and then nothing, ever.

This matters because an AI agent asked to evaluate a token is overwhelmingly being asked about a corpse. Every security tool in this space answers is this malicious? Almost none answer the question that is true far more often: is this alive?

The curve

A cohort of 250 tokens launched inside one narrow band, so every token is the same age. For each we counted ERC-20 Transfer events in three windows measured from its own launch block: the first ten minutes of life, an hour later, and the most recent thirty minutes of the chain.

The last row is the one that shows the instrument works. Tokens that survived traded a median of 216 times in their first ten minutes; tokens that died traded 13. A sixteenfold separation is what a real signal looks like. And seven tokens in the cohort were still trading in the final window, which is how we know a zero means death rather than a broken query.

Nobody knows which one you mean

Agents resolve tokens by name and symbol. On this chain those are not unique. Across 1,600 sampled tokens, 35.3% share a symbol with another token, and the collisions are not accidents spread over months — 58% of duplicate families had two members launched under ten minutes apart. One symbol, RBNHD, was claimed by six near-identical tokens inside thirty-five seconds.

The duplicates come from different deployers, which is the part that makes it a security problem rather than a spam problem. It is imitation by separate actors, not one wallet flooding the chain. When a user tells an agent to buy RBNHD, there is no correct answer.

Exact-string matching also misses an entire class. Nine tokens in the sample carry Cyrillic characters that render as Latin — one symbol displays as NYSE, built with a Cyrillic Е. Another renders as CHIP and collides with a real ASCII CHIP already on the chain. A filter looking for the string “NYSE” sees nothing at all.

The scale

For context on what any of this is filtering. All read live from the chain rather than from a dashboard.

Note the last row. Burning is widely described as a launchpad norm; on this chain 1.1% of tokens have sent any supply to a burn address at all, and none exceeded 3%. We had assumed the opposite before measuring, on the strength of a handful of tokens that burned 96.8%. Those are outliers, and we were wrong.

Method

Everything above is reproducible against the public RPC with no account and no key.

# launches: the PONS factory's launch event
topic0 = 0x8d4aad4953d0ca700d468f3753aa14432d1b35b43ec6409f051fb6aa43a89607
address = 0x7eD598BcEf8bd9Edd8C97A195C6d13f40801EC7e

# survival: Transfer events, per token, address-filtered
topic0 = 0xddf252ad1be2c89b69c2b068fc378daa952ba7f163c4a11628f55a4df523b3ef

# metadata: read straight off the token contract
name()        0x06fdde03
symbol()      0x95d89b41
description() 0x7284e416
getTokenInfo() 0xabb1dc44   # socials: website, twitter, telegram, ...

Two traps cost us real time and are worth repeating. The gateway serves a Cloudflare challenge to a default client — a browser user-agent fixes it. And eth_getLogs silently caps at 10,000 logs rather than erroring, so wide windows return truncated results that look complete. Address-filtered per-token queries avoid both.

What this instrument cannot see

The dataset spans 6.9 hours. That means 24-hour survival is not measurable here, and we are not going to estimate it. “Dead within an hour” is fully supported by the data; anything beyond six hours is not.

Duplicate-family sizes come from a 23.3% sample, so the true families are roughly four times larger than the counts above — we report what we saw, not what we extrapolate. Website liveness counted only 404s and timeouts as dead; 403s were excluded because bot-blocking is indistinguishable from absence at this remove, which makes 18.5% a floor rather than an estimate.

An earlier pass of this work reported different metadata percentages. They were computed on a subset where transient RPC failures had been silently recorded as unreadable tokens — a sampling bias, not a finding. The numbers here are re-derived at N=1,600 with zero fetch failures, and they supersede the earlier ones.

Why we measured this

We run a layer that scans token metadata before an agent reads it, and it reports a verdict of clean_by_rules or findings. This research says that vocabulary is incomplete. For the large majority of tokens an agent will ever be asked about, the honest verdict is neither — it is dead, and no scanner on the market says so.

Freshness is not a safety property in the traditional sense. Nobody is attacked by a token that never traded. But an agent that cannot tell the difference between a live market and an abandoned one will make confident decisions about nothing at all, and that failure mode is roughly nine times more common than the one everyone is selling protection against.

Chain figures read from rpc.mainnet.chain.robinhood.com (chainId 4663) on 2026-09-11. Registry figures are re-fetchable at /api/v1/stats/launches. Confidence intervals are Wilson score at 95%. Negative findings mean “not found with substantial effort,” never “proven absent.”