This page is the product. Each rule we apply came out of a dated study with a sample size, and the studies stay up whether they flatter us or not.
Where we don't have a number yet, we say so. Every term on this page is in the glossary, and the small "i" next to a column name explains that column.
6 of the 18 lines below came out of a dated study on our own outcome rows. The rest were set by hand and say so. None of them is presented as measured when it wasn't. Current as of 2026-10-04.
| What it decides | Value | Unit | Basis | Set on | n | Where it came from |
|---|---|---|---|---|---|---|
| Conviction tier: the 10 largest wallets hold this much or less | 30 | % of supply | set by hand | 2026-07-23 | — | Published research on token risk read in July 2026 (over 30% in ten wallets marked 87% of high-risk cases). Citation not on file. |
| Above this the token is watch-only and never sized | 45 | % of supply | set by hand | 2026-07-23 | — | Set by hand as a buffer above the 30% line |
| Conviction needs a pool at least this deep; thinner pools are scout size | 50,000 | USD in the pool | set by hand | 2026-08-01 | — | Set by hand, lowered from 100,000 |
| Largest size we would suggest for a conviction token | 1,000 | USD | set by hand | 2026-07-23 | — | Set by hand |
| Largest size we would suggest for a scout token | 100 | USD | set by hand | 2026-07-23 | — | Set by hand |
| Never more than this share of the pool, whatever the tier | 1 | % of the pool | set by hand | 2026-07-23 | — | Set by hand |
| A deployer still holding more than this fails the token | 5 | % of supply | set by hand | 2026-07-23 | — | Published research on token risk read in July 2026 |
| Fewer holders than this fails the entry test | 100 | wallets | set by hand | 2026-07-28 | — | The one holder figure in the July 2026 research |
| Up more than this today fails the entry test | 30 | % price change in 24h | set by hand | 2026-07-28 | — | Set by hand |
| Quieter than this fails the entry test | 10,000 | USD traded in 24h | from our data | 2026-09-22 | 105 | Board review: 105 passing tokens, 42 under this floor; 39 survived both floors |
| Less turnover than this fails the entry test | 5 | % of market cap traded in 24h | from our data | 2026-09-22 | 105 | Board review: median turnover of passing tokens was 4% of market cap |
| A pool over $100,000 doing less than this reads as fabricated or dead | 0.2 | % of the pool traded in 24h | set by hand | 2026-07-28 | — | Set by hand |
| Conviction needs at least this much decline from peak | 60 | % below all-time high | set by hand | 2026-09-09 | — | Set by hand, replacing the base-proximity block |
| A pool drained more than this fails the token | 70 | % below its recorded peak | set by hand | 2026-09-09 | — | Set by hand |
| At base band ends here | 30 | % above the 30-day low | from our data | 2026-09-04 | 1013 | First-alert study: within 30% of the low read +8% median, 61% positive |
| Just off base band ends here | 100 | % above the 30-day low | from our data | 2026-09-04 | 1013 | First-alert study: 30 to 100% read -1% median, 50% positive |
| Well off base band ends here; beyond is already ran | 500 | % above the 30-day low | from our data | 2026-09-04 | 1013 | First-alert study: over 500% read -70% median, 13% positive, 31% down 80%+ |
| Bands for the halve-in-7-days read | 0.4, 0.5, 0.6, 0.7 | probability | from our data | 2026-09-24 | 423 | Backtest of 423 alerts against realized 7-day return, rank correlation -0.47 |
Next step, already in the build queue: every line re-cut every 24 hours from our own store of outcome rows, and this table updated with the new value, date and n. That needs outcome rows with the raw facts at alert time (top-10 share, volume, market cap, distance from the 30-day low, drawdown from peak) flowing into Keelage's own store. Status: not yet running.
An alert is a token that passed our structural checks at the hour our scanner saw it. We log its price then, and read it again 24 hours, 7 days, and 30 days later.
| Dataset | Rows |
|---|---|
| Alerts from 28 Jul to 16 Sep 2026, returns read 24 Sep 2026 | 423 |
| Of which on Robinhood Chain (the rest are from the earlier 5-chain period) | 127 |
| With a 30-day return | 181 |
| Without a 30-day-low reading (it wasn't read at alert time) | 255 |
| Unique tokens | 408 |
Return is the change in price from the alert to the read. We use the median, not the mean, because one 60x hides forty zeros. "Positive" is any gain. "Halved" is -50% or worse. Every table carries an "all alerts" row, because in this market most tokens go down: a group is only telling you something when it differs from that row.
Two honest limits. Only tokens that passed the checks were tracked, so we can't yet show how the ones we rejected did. And most of these rows come from before the scanner went Robinhood-only on 9 Sep 2026, so the chain-specific tables are small.
Both are fixed by time. The Robinhood-only rows grow by the hour, and from the next study on we track the rejects too.
Study of 4 Sep 2026 on alert rows. 7-day outcomes unless marked.
| Rule tested | Median | Positive | n |
|---|---|---|---|
| Within 30% of its 30-day low at alert, 7 days later | +3.8% | 63% | 27 |
| Further from the 30-day low than that, 7 days later | -29.6% | 24% | 204 |
| Within 30% of the low, 24 hours later | +0.1% | 52% | 27 |
| Further than that, 24 hours later | -11.1% | 33% | 204 |
| Admitted by a third-party safety score, Robinhood Chain only | -25% | 24% | 49 |
| Rejected by that score, Robinhood Chain only | -2% | 30% | 108 |
A rule set built on a third-party safety score admitted the worse half of the market. Reading the verified code, the launcher, the owner's powers and the tax code straight from the explorer reversed it. That's why no third-party score decides anything here.
| First-alert study, 1,013 tokens, 7 days later | Median | Positive | Down 80%+ | n |
|---|---|---|---|---|
| Within 30% of its 30-day low | +8% | 61% | 0% | subset |
| More than 500% above its 30-day low | -70% | 13% | 31% | subset |
| Carrying a paid takeover claim | -58% | 11% | — | 155 |
| No takeover claim | -13% | 37% | — | 858 |
Same study, 4 Sep 2026. Each token counted once, at its first alert. The claim rows are why a paid takeover claim shows as a caution on every Keelage verdict and counts for nothing.
How far above its lowest price of the last 30 days a token sat when we saw it. We group that distance into 4 bands. The band is on every verdict because it ranked outcomes better than our score did.
| 423 alerts, all chains, 7 days later | Median | Positive | 2x | Halved | n |
|---|---|---|---|---|---|
| All 423 alerts, for comparison | -29.3% | 24% | 5% | 35% | 423 |
| At base (within 30% of the low) | +0.9% | 57% | 5% | 0% | 21 |
| Just off base (30 to 100% above) | -9.4% | 35% | 0% | 3% | 31 |
| Well off base (100 to 500% above) | -30.8% | 18% | 4% | 31% | 45 |
| Already ran (more than 500% above) | -50.7% | 13% | 6% | 51% | 71 |
| Not read at alert time | -34.1% | 24% | 5% | 38% | 255 |
28 Jul to 16 Sep 2026, read 24 Sep 2026. The further from the low, the worse it went, in order.
| Robinhood Chain only, alerts at least 7 days old | Median | 2x | n |
|---|---|---|---|
| Already ran (more than 500% above) | +2.1% | 37% | 10 to 30 |
| Well off base (100 to 500% above) | -51.8% | 21% | 10 to 30 |
| Just off base (30 to 100% above) | -33.4% | 6% | 10 to 30 |
| At base (within 30% of the low) | -77.4% | 10% | 10 to 30 |
Read 22 Sep 2026. On this chain alone, and at these sample sizes, the order flips. We don't know yet whether that's the chain or the sample. We'll re-cut it when every band has 50 rows.
Our sizing tier comes from how much of the supply the 10 largest wallets hold, with pools and contracts removed. The 30% line comes from published research on token risk; the 45% line is where a handful of wallets can set the price alone. The tier decides how much we'd size, never whether a token passes. The table below is why we keep it as a size cap and not a predictor.
| Sizing tier at alert, 7 days later | Median | Positive | 2x | Halved | n |
|---|---|---|---|---|---|
| All 423 alerts, for comparison | -29.3% | 24% | 5% | 35% | 423 |
| Conviction: 10 largest hold 30% or less, pool $50k or more | -31.5% | 18% | 7% | 34% | 107 |
| Scout, thin pool: 30% or less, pool under $50k | -42.2% | 20% | 5% | 43% | 157 |
| Scout, cluster check: 30 to 45% | -25.0% | 22% | 3% | 35% | 60 |
| Watch only: over 45%, never sized | -8.2% | 42% | 0% | 13% | 45 |
423 alerts, read 24 Sep 2026. 54 rows from two retired tiers are left out. The tier we'd size biggest lost more often than the one we don't size at all. It still caps size, because concentration is a rug vector even when it isn't a return predictor.
| Market cap at alert, 7 days later | Median | Positive | 2x | Halved | n |
|---|---|---|---|---|---|
| All 423 alerts, for comparison | -29.3% | 24% | 5% | 35% | 423 |
| Under $100k | -37.4% | 21% | 5% | 37% | 171 |
| $100k to $500k | -30.8% | 22% | 4% | 43% | 101 |
| $500k to $2M | -38.8% | 18% | 3% | 38% | 61 |
| Over $2M | -11.4% | 34% | 8% | 20% | 90 |
Same 423 rows. Over $2M the halving rate drops by half and the 2x rate is the best of any band.
| How long after the alert we read the price | Median | Positive | 2x | Halved | n |
|---|---|---|---|---|---|
| 24 hours | -6.7% | 36% | 4% | 13% | 423 |
| 7 days | -29.3% | 24% | 5% | 35% | 423 |
| 30 days | -47.9% | 27% | 7% | 49% | 181 |
| 7 days, the same 181 rows that have a 30-day read | -19.4% | 31% | 5% | 27% | 181 |
Read 24 Sep 2026. The 30-day set is the older half of the alerts, so compare it with its own 7-day row, not the full one.
| Where the price stood 24 hours after the alert | 7-day median | Positive | 2x | Halved | n |
|---|---|---|---|---|---|
| Up after 24 hours | -8.9% | 44% | 12% | 20% | 151 |
| Flat or down after 24 hours | -44.0% | 13% | 1% | 43% | 272 |
Same 423 rows. Of the alerts that reached 2x in a week, 18 of 21 were already up after a day. A day-one loser almost never came back.
Most tokens on this chain are stamped out by a launcher, a factory contract that deploys the same code for everyone. Census of 9 Sep 2026: 384 token contracts read from the explorer, 85% with verified code, 92% made by a launcher, 5 launchers covering 68%.
| Launcher | What its code allows | Tokens in census | Median change since first seen |
|---|---|---|---|
| Pons launcher | Clean: no owner powers, no transfer tax | 141 | -57% |
| LaunchToken | Clean | 61 | -39% |
| Robinhood tokenized stock | Excluded as a class: these are shares, not memecoins | 34 | — |
| Doppler | Clean | 32 | -29% |
| UERC20 (Pools Trade) | Clean | 16 | — |
"Clean" is about the code only. It says nothing about the people behind a token, which is the next table's point.
| Launcher at alert, 7 days later | Median | Positive | 2x | Halved | n |
|---|---|---|---|---|---|
| All 75 launcher-read alerts, for comparison | -31.5% | 16% | 5% | 28% | 75 |
| Pons launcher | -44.2% | 9% | 3% | 35% | 34 |
| No launcher: written by hand, code verified | -9.4% | 33% | 13% | 20% | 15 |
| LaunchToken | -28.1% | 8% | 0% | 15% | 13 |
| Doppler | -44.8% | 25% | 13% | 38% | 8 |
| UERC20 (Pools Trade) | -31.5% | 20% | 0% | 20% | 5 |
The 75 Robinhood Chain alerts whose contract we could read back on 4 Oct 2026, all with verified code. Small. The one pattern worth a sentence: tokens somebody wrote by hand did better than tokens a factory stamped out, which fits a chain where launching costs nothing. We'll re-cut this at 50 per launcher.
Jev, a model that answers a typed question with a probability, is asked two things about every alert: "will this halve within 7 days?" and "will it be up in 7 days?" We tested both answers against what happened. 423 alerts, 24 Sep 2026.
| Jev's chance of halving within 7 days | Median | Positive | Halved | n |
|---|---|---|---|---|
| All 423 alerts, for comparison | -29.3% | 24% | 35% | 423 |
| Under 40% | -4.0% | 41% | 0% | 29 |
| 40 to 50% | -7.3% | 42% | 1% | 67 |
| 50 to 60% | -26.4% | 21% | 32% | 171 |
| 60 to 70% | -51.5% | 18% | 52% | 120 |
| 70% and up | -83.9% | 8% | 83% | 36 |
Rank correlation with the realized return: -0.47. Each band did worse than the one above it. This one earns a sentence on the verdict, as context, never as a rule.
The "up in 7 days" answer had a rank correlation of about 0.10 with what happened. That's noise. We still return the number because we said we'd show what we measure, but we put no words on it and nothing in the verdict depends on it.
Neither probability admits, blocks, sizes, or ranks a token. They ride along for context and get re-tested every month.
By 7-day return. Every one of these passed the structural checks at the time. That's the point of showing them.
| Worst | Alert date | Tier | 24h | 7d |
|---|---|---|---|---|
| PONSCUPINE | 2026-09-13 | conviction | -93.5% | -98.7% |
| LOOT | 2026-09-15 | conviction | -83.1% | -98.7% |
| MOW | 2026-09-10 | conviction | -75.8% | -98.5% |
| Sparky | 2026-08-26 | conviction | -90.3% | -97.3% |
| Dust | 2026-09-05 | scout, cluster check | -87.1% | -96.5% |
| チョロ松 | 2026-08-28 | scout, thin pool | -87.3% | -96.3% |
| CTO | 2026-09-08 | retired tier | -53.3% | -96.3% |
| DTF | 2026-09-10 | conviction | +4.3% | -96.2% |
Five of the eight were tier conviction. Seven of the eight were already down more than half after one day. That's the day-one pattern in the table above, with names on it.
| Best | Alert date | Tier | 24h | 7d |
|---|---|---|---|---|
| BLORB | 2026-09-13 | conviction | +972% | +6,096% |
| LOOM | 2026-09-06 | scout, thin pool | +3,429% | +4,576% |
| USDC (a memecoin, not the stablecoin) | 2026-09-06 | scout, thin pool | +4,280% | +2,832% |
| Buddy | 2026-08-11 | scout, thin pool | -15.8% | +788% |
| STONK | 2026-09-05 | conviction | +285% | +675% |
| INCOME | 2026-08-19 | scout, thin pool | +19.9% | +531% |
| WADDLESLIV | 2026-08-23 | scout, thin pool | +48.9% | +443% |
| OTC | 2026-09-01 | retired tier | -10.5% | +407% |
Five of these were thin-pool scouts at alert, where the size rule would have put $100 or less into each. We show the winners so you can see they exist, not so you'll expect them. 21 of 423 alerts doubled in a week.
Until now only tokens that passed the checks were tracked. From this month the scanner logs a price for every token it judges, pass or fail, so the next study can show both sides on the same dates.
Creator fees claimed, tokens bought and burned, money moved off the chain, and the 7-day return after each pattern. The tool that reads this ships next; the table needs a few weeks of its output before there's anything honest to show.