Cite the
method.
The scoring method as a paper, the launch-outcome dataset behind it, and the live results that test it. Written so that anyone can check, reuse and cite it.
The paper
in brief.
Reading launch risk from public chain data: the Cyclops scoring method for Torch launches on Robinhood Chain. Cyclops Labs, 2026. Version 1.0.
We describe a rule-based method that reads a newly launched token on Robinhood Chain (chain id 4663) from public chain data only, and summarises what it reads as a score from 0 to 100 with every contributing observation listed. It combines holder concentration, launch-window buying, wallet funding links, the creator’s record of earlier launches, liquidity, contract permissions and trading patterns into 24 published checks with fixed thresholds and weights, plus depth checks added in later versions. Separately, every Torch launch is followed after it opens and assigned one of four outcomes by fixed rules. Results are served live and change as launches resolve; this page reports none as a fixed figure.
The paper and dataset are prepared for deposit on Zenodo; the DOI appears here once it is assigned. Until then, cite this page and the date you read it.
How it
reads.
- Checks and weights. Each check that fires adds its weight; the score is the sum, capped at 100. Bands: Low 0–19, Moderate 20–39, Risky 40–64, High 65–100. The full table is in the methodology; the code is the separate scoring package.
- Wallet links. Funded-by, shared non-hub funder, direct transfers, same-transaction buys and deployer links. Funders labelled as exchanges, bridges, routers or launchpads, and any funder with 300 or more transactions, are hubs and never link wallets.
- Evidence strength. Strong, moderate or weak on every flag, with the transactions it rests on; checks that could not run are unavailable, never passed.
- Completeness band. A range around the score from how much of each input was read. It never changes the score.
- Creator sold out
- The price fell 80% or more from its peak while the creator wallet sold.
- Collapsed
- The price fell 90% or more from its peak.
- Went quiet
- No trades for three days.
- Still trading
- None of the above; trading continues.
Outcomes describe what happened on chain. They are not findings of intent. Ground truth is pre-labelled from outcomes and marked unreviewed until a person checks it.
How it has
done so far.
- Labelled launches
- 0
- Checked by a person
- 0
- Labelled bad / ok / unclear
- 0 / 0 / 0
| Flag | Fired | Followed by a collapse |
|---|---|---|
| Few holders | 104 | 8% |
| Thin liquidity | 84 | 3% |
| Creator tied to other creators | 69 | 0% |
| Less than an hour old | 54 | 3% |
| Dev sold most of their bag | 30 | 25% |
| Copies an earlier name or ticker | 19 | 8% |
| Fresh wallets among top holders | 5 | 0% |
| Snipers in the first seconds | 4 | 100% |
Base rate across judged launches: 2% of 1208. Rate is the share of judged launches with this flag that later collapsed or were dumped by the creator. Compare it with the base rate; small samples say little.
Read live: GET /v1/ground-truth (precision and recall per flag, all labels and the human-reviewed subset), GET /v1/flags/accuracy, GET /v1/score/calibration and GET /v1/score/bias (see calibration). Cite the endpoint and the date, or a versioned release.
Download
the record.
Every Torch launch Cyclops has indexed on Robinhood Chain, one row each, oldest first. Add ?since=2026-10-01 for launches on or after a date. Outcomes describe what happened on-chain, not intent, and keep updating while a launch is young. Credit “Cyclops Labs” when you publish with it.
| Column | Meaning |
|---|---|
token | Token contract address |
symbol | Ticker as written in the launch event |
name | Name as written in the launch event |
creator | Wallet that created the launch |
block | Block of the launch |
launched_at | Launch time, ISO 8601 UTC |
tax_bps | Launch tax setting in basis points |
swaps | Swaps counted so far |
volume_eth | Traded volume in ETH |
first_eth | Price of the first trade, ETH per token |
ath_eth | Highest observed price, ETH per token |
last_eth | Latest observed price, ETH per token |
last_swap_at | Time of the latest swap |
dev_sold_pct | Share of the creator’s tokens sold, percent |
outcome | alive, inactive, dev_dumped or collapsed; empty while too new to judge |
outcome_at | When the outcome was recorded |
Versioned releases bundle the dataset, the score changelog, ground-truth and accuracy snapshots, the paper, a manifest with the read time and row counts, and SHA-256 checksums.
How to
cite it.
Cyclops Labs (2026). Reading launch risk from public chain data: the Cyclops scoring method for Torch launches on Robinhood Chain (Version 1.0). Zenodo. DOI pending.
@misc{cyclops2026method,
author = {{Cyclops Labs}},
title = {Reading launch risk from public chain data: the Cyclops scoring method for Torch launches on Robinhood Chain},
year = {2026},
version = {1.0},
publisher = {Zenodo},
url = {https://cyclopslabs.dev/research},
note = {Score method v5; DOI pending}
}Machine-readable: CITATION.cff and .zenodo.json in the repository. The deposit is prepared under CC BY 4.0; the licence is final once the record is published on Zenodo. Credit “Cyclops Labs”.
What
changed.
- Score method v5 · 2026-10-29. New flags: bytecode that differs from every Torch token template (instructions compared line by line, constants ignored), contracts that can be upgraded with no timelock, names or tickers using lookalike characters that read as an earlier token, and logos or descriptions reused from other creators’ launches. Scans add locks from every locker found on Robinhood Chain, an exit curve for the creator’s and linked wallets’ bags at current depth, creator wallets that probably act together (with confidence), survival of past launches by creator class, the full swap tape, a confidence band from data completeness, and each flag’s contribution to the score.
- Score method v4 · 2026-10-29. Every flag carries an evidence strength (weak, moderate, strong) and the transactions or wallets it rests on. New flags: repeated round-trip trading and back-and-forth transfers (wash), wallets funded alike or airdropped in bulk (sybil), three or more wallets gassed by one source within two blocks, earlier tokens with the same name or ticker, and creators who share a funding source with other launch creators. Scans add snipers by behaviour across launches, supply that could leave against pool depth, observed sell routes and token-side taxes, tax changes, holder concentration over time, sandwiches in the pool, and socials taken only from the launch metadata committed on-chain.
- Score method v3 · 2026-10-28. Creator's earlier launches that collapsed or were dumped count against the score. Wallets linked to the dev by funding, transfers or same-transaction buys (insiders) and the largest linked group are scored. Early buyers who were also early in the creator's earlier launches, and the dev's off-market transfers, are scored. Contract powers are read from bytecode, and selling and transfers are simulated from a real holder.
- Score method v2 · 2026-10-28. The creator's other recent launches and how they scored feed the score.
- Score method v1 · 2026-10-28. First score: holder concentration, the largest wallet, launch bundles, fresh wallets, snipers, the dev bag and its selling, liquidity, LP lock, buy tax, age and holder count.