Mission
The information was always public. Reading it in time was the problem.
Everything needed to spot most rug pulls sits on-chain, visible to anyone. What nobody has is fifteen minutes to read it while a launch is thirty seconds old. HoneyRug exists to close that gap.
The asymmetry in new-token trading is not really about information. It is about time. A developer preparing a launch has weeks. A buyer deciding whether to take part has, in practice, under a minute — and the parts of the picture that matter most, like whether the creator's wallet has done this before, are the slowest to check by hand.
We built the scoring engine to do that reading in about two seconds and to hand back not just a verdict but the reasoning, in language that means something to a person who is not a smart-contract auditor.
The goal is not to tell anyone what to buy. It is to make sure that when someone does buy, they knew what they were buying.
What we deliberately do not do
- We do not give financial advice. We are not authorised to, and a scoring engine does not know your circumstances.
- We do not predict prices. The growth model reports a probability that certain conditions are present, never a target or a forecast.
- We do not sell placement. Feeds are ranked by our own scores. No project can pay to appear, or to have a finding removed.
- We do not hide uncertainty. When data is missing, confidence drops and the report says so rather than filling the gap with an assumption.
Technology
How the analysis system is built.
Ingest
On-chain state, pool data, holder distribution, creator wallet history and trade flow are fetched in parallel through the GMGN API, then normalised into a single signal object. Absent fields become null rather than zero — a measurement and an admission are different things, and the whole confidence system depends on preserving that difference.
Score
Six risk families and seven growth factors are scored by piecewise curves rather than thresholds, so a value near a boundary does not flip the verdict. Categories are pulled toward their worst signal, and disqualifying conditions cap the total. Every run is pure: same inputs, same output, always.
Explain
Findings are generated deterministically from the same inputs that produced the score, so the reasoning can never drift from the number. The long-form written report on HoneyRug Max is a readability pass over fixed findings — the model rewrites, it never decides.
On the word "AI"
The scoring itself is a weighted statistical model, not a language model — it has to be reproducible, auditable and identical on every run, which is not what language models are good at. Where a language model is used, it is used for one job: rewriting fixed findings into readable prose on paid reports. It is given the findings as immutable input and cannot add, remove, reorder or soften any of them. We would rather say that plainly than let the label do work the system does not.