Filters
The bot's default is a firehose. Filtering is how it becomes yours.
Open the filter menu with /filter. Everything below is reachable from there — presets first, then dimension by dimension if you want to build your own.
Start with a preset
Most people never need more than this.
⭐ Quality
Early, sellable, liquid enough to exit.
- Safety verdict: Clean or Undetermined
- Buys 24h ≤ 146
- Market cap ≤ $85k
- Liquidity ≥ $10k
⭐ Quality · no memes
The same, with the meme category excluded.
- Everything in Quality
- Category: everything except Meme
🚀 Early
Earliness only — a wider net and more noise.
- Safety verdict: Clean or Undetermined
- Buys 24h ≤ 146
- Market cap ≤ $85k
🛡 Sellable only
Everything except confirmed-unsellable tokens.
- Safety verdict: Clean or Undetermined
📡 Everything
The raw firehose. No filtering at all.
- No constraints
⭐ Quality is the recommended starting point. It combines the only two dimensions we have actually measured an edge on with a floor on liquidity, so what arrives is early, sellable, and deep enough to exit.
⭐ Quality is the best preset we have measured, and it fails the fat-tail test. Three tokens carried 61% of its total upside in the measurement window. Strip them out and the edge largely goes with them. That does not make it useless — it makes it a preset whose returns arrive in a few large lumps rather than steadily, and you should size for that.
What the evidence labels mean
Every dimension carries one of three labels. They exist so you can tell an edge we measured from a preference we are simply letting you express.
This dimension separates outcomes on our own data, and the separation survived a fat-tail test that removes the three biggest winners.
This protects you from tokens you cannot exit. It is not a claim about upside.
A real, reliably populated property with no measured relationship to outcome. Legitimate to filter on. Not an edge.
The distinction is not decoration. Exactly two dimensions in the whole list below carry measured. Two more carry safety. Everything else is a preference — an honest, well-populated property with no demonstrated relationship to what a token does next. We would rather put that in a label than let a well-designed settings menu imply an edge that is not there.
Every dimension
Market
How much trading had already happened when we fired. Lower is better.
Measured on Robinhood, and it does not carry over. On Robinhood, at or under 146 buys the chance of reaching 5x its alert price was 23.3% against an 11.8% baseline at t+1h — a lift of 1.98 (n=450 of 1,281). On Base the same rule is flat: lift 1.21 at t+24h and 0.95 at t+6h (n=96 of 359). Use it on Robinhood; on Base it is a preference, not an edge.
Very early (≤75)Early (≤146)Any
Market cap at alert time.
Holds on both chains. Zero of 152 Base launches and zero of 116 Robinhood launches above roughly $85K market cap ever reached 10x. This is the cleanest negative filter we have — it works by ruling things out, not by picking winners.
Under $30kUnder $85k$10k – $250kAny
Pool depth at alert. Decides whether you can get in and out at size.
Robinhood pools run about three times thinner than Base — median $14.8K against $44.5K. A floor that feels reasonable on Base will silence most of Robinhood.
Min $5kMin $15kMin $50kAny
Traded volume at alert time.
Min $10kMin $50kAny
Distinct holders at alert.
Correlates about 0.80 with buys24h on Robinhood, so this is largely the earliness lever seen from another angle — and earliness itself is a Robinhood effect. Setting both does less than it appears to.
Min 25Min 100Any
Share of supply held by the ten largest non-contract wallets. Lower is more distributed.
Under 25%Under 40%Any
Safety
The collapsed verdict from the nine-layer safety stack.
RUG_RISK means unsellable — sell probes reverted at every size, or a round trip costs 50% or more. Leaving RUG_RISK switched on is how you receive tokens you cannot exit.
CleanUndeterminedRug risk (unsellable)
Share of the insider allocation already sold at alert time.
Under 25%Under 50%Any
Share of supply handed to insider wallets rather than bought on the open market.
Under 50%Under 70%Any
Deployer
How many previous tokens by this deployer we recorded as rugs.
NoneAt most 1Any
Where the deployer wallet got its money.
Deployer age is deliberately absent from this list. At launch every deployer is under a day old, so the field is a constant and carries no signal at all.
Fresh walletEstablished walletUntraced
Product evidence
A reachable repository.
Present on 24 of 913 Robinhood launches, and zero of those reached 10×. Switching this on cuts the feed to about 4% and buys no measured upside.
RequiredAny
Present on about 8% of launches. No measured relationship to outcome.
RequiredAny
Extremely rare — about 1%. Switching this on will nearly silence the feed.
RequiredAny
Chain & category
Which chains you want launches from.
BaseRobinhood
The narrative bucket, classified by a language model.
Chain-dependent, and mostly an input problem. About 34% of Robinhood and 60% of Base launches land in Unclassified — but where the token carries a description the classifier lands it 87.8% of the time on Base and 73.3% on Robinhood. Base simply publishes a description on only 31% of launches against 86.5% on Robinhood. Category shows no measured link to outcome on either chain: use it to shape your feed, not to find an edge.
AIAgentsInfrastructureDeFiRWAGamingSocialUtilityMemeUnclassified
The ones we deliberately do not offer
A filter that is always null, or always the same value, is worse than a missing one — you believe you filtered, and you did not. These were measured across the full corpus and cut:
Null on every one of 913 Robinhood launches. A filter that is always null is worse than a missing one, because you believe you filtered and you did not.
The literal string "unknown" on every row.
Zero on every non-null row.
Identically zero across all 803 Robinhood tokens measured. A dead channel.
Always true — it is an upstream hard gate. Zero variance means zero filtering power.
Set only on confirmed hard fails. Null does not mean safe, so offering it as a three-way switch would actively mislead. Use the safety verdict instead.
"Weak" on 88% of rows. Near-constant, and filtering on it implies a precision we do not have.
Excluded on purpose. Their apparent signal is an enrichment-failure artifact: "social fetch failed" tracked better outcomes than success, and inside the succeeded subset the correlation with peak was +0.017 (p=0.73) — exactly zero. Shipping these would sell a pipeline state as a token property.
The social exclusion is the one worth reading twice. On its face, "only alert me when real KOLs are talking" is the most attractive filter we could build. In our data the signal ran the wrong way: launches where the social fetch failed had better outcomes than launches where it succeeded, and inside the succeeded subset the correlation with peak price was +0.017 with a p-value of 0.73 — statistically indistinguishable from nothing. The apparent edge was an artifact of our own pipeline's failure modes. Shipping it would have sold you a fact about our infrastructure dressed as a fact about a token.
A note on stacking
Dimensions combine with AND. Every constraint you add narrows the feed, and some narrow it far more than they appear to:
- Holders and buys-24h correlate at 0.80. Setting both is close to setting one twice.
- Has GitHub cuts the feed to roughly 4% and, in our data, removed every 10× along with the noise.
- Liquidity floors are chain-specific. A $50K floor is reasonable on Base and silences most of Robinhood, where the median pool is $14.8K.
Use /filters to see what you currently have set, and the preview count before you save.
➡️ See also: Conviction score · The safety stack
Anything on this page unclear, or wrong?
Tell us →