Tracer

Which new Solana tokens really ran, which wallets were early on them, and every signal on the record.

Get signals live

Does a flag mean anything?

A signal is raised when several tracked wallets buy the same new token within minutes. This page tests whether a token flagged that way ran more often than a token picked from the record at random, and says so only as far as the record carries it.

Replay of history

Not shown yet. Over 217 tokens read between 10 Oct 2026 and 22 Oct 2026, the rule flagged 0; of all the tokens read, 22% ran. That is not enough to say the rule picks tokens that run more often than chance.

  • A claim here needs at least 300 judged tokens across 21 days, 30 of them tokens that ran and 30 of them flagged. So far: 217 tokens across 12 days, 47 that ran, 0 flagged.
  • The random wallets raised too few flags to compare with (302 across 20 lists; 30 are needed), so the part that choosing wallets accounts for cannot be measured.

Replay of history: tokens launched from to

no share

win rate: 0 of 0 flagged tokens ran (range no range)

22%

base rate: 47 of 217 tokens read ran (range 17% to 28%)

none

win rate over base rate (range range not measurable)

0%

recall: 0 of 47 tokens that ran were flagged (range 0.0% to 7.6%)

Win rate, on this page, means one thing: the share of flagged tokens that ran. It is always given beside the base rate, the share of all the tokens read that ran, because a win rate without it says nothing. Every range is a 95% interval: the true figure plausibly lies inside it, and a wide one means there is little data.

How much, and how sure

Flagged tokens against all tokens read: not measurable on this record.

Scored wallets against random wallets: not measurable on this record.

The mark is the ratio and the bar is its range. The tick at 1 is no difference at all. A bar that crosses the tick means the record cannot tell the rule from chance, and nothing is claimed.

The second line is the placebo. The same rule was run 20 more times with wallets drawn at random in place of the scored ones, each about as busy as the wallet it stood in for. Those random lists raised 302 flags between them, and 28% of the flagged tokens ran (from 20% to 38% across the lists). A token that ran has more early buyers on record than one that went nowhere, so any list of busy wallets flags runners more often than chance. Only what the scored wallets do beyond the random ones comes from choosing wallets by their record.

Week by week

The same counts split by the week a token launched, so that one good week cannot carry the whole. No week has a flag yet.

For each week: tokens judged, the share that ran, tokens flagged, the share of those that ran, and the same for the random wallets.
Week ofTokensBase rateFlaggedWin rateRandom walletsWin rate over base rate, drawn
3816% (6)0no flags100% (5 of 5)
8822% (19)0no flags21% (18 of 87)
9124% (22)0no flags30% (62 of 210)

In the last column the upper line is that week's win rate and the lower one its base rate, on one scale from none to all. A week with a handful of flags says little by itself; its counts are there so that you can see how few they are.

How many it caught, and how early

Of the 47 tokens that ran, the rule flagged 0: 0% (range 0.0% to 7.6%). The rest ran without a flag.

From launch to the flag, among flagged tokens that ran: the median
not measured
SOL in the token's curve once the flag's last buy had landed: the median
not measured

The replay places a flag at the second its last buy landed. The live engine reads that buy a few seconds later, and the public sees a signal 30 minutes after that.

What a buyer at the flag would have seen

Not shown. The record this run read does not hold the market cap of a token at the moment of its flag, and this page does not work one out from a formula or from a price of SOL nobody recorded. Until the record holds it, the question on this page is only whether flagged tokens ran, not what they were worth.

Other rules, tried for comparison

The claim on this page is for one rule, the one the live engine runs: 3 separate voices within 20 minutes, among the best 150 wallets. It was set before this record was replayed and was not fitted to it. The table shows eight neighbours of it, as exploration only. With nine rules tried, one of them looks good by luck far more easily than a single rule does, and the ranges here are not widened for that. Do not read the best cell as a finding.

Each rule of the grid: how many tokens it flagged, the share that ran, and that share over the base rate with its range.
RuleFlaggedWin rateOver base rateRange
2 voices in 10 min1100% (1 of 1)4.6xrange not measurable
2 voices in 20 min1100% (1 of 1)4.6xrange not measurable
2 voices in 30 min1100% (1 of 1)4.6xrange not measurable
3 voices in 10 min0no flagsnonerange not measurable
3 voices in 20 min (the live rule)0no flagsnonerange not measurable
3 voices in 30 min0no flagsnonerange not measurable
4 voices in 10 min0no flagsnonerange not measurable
4 voices in 20 min0no flagsnonerange not measurable
4 voices in 30 min0no flagsnonerange not measurable

Since launch (live): signals raised from

Not shown yet. 4,540 signals have been raised live and closed (72 more are still open and not counted), and 10 of the tokens ran (0.2%); of the tokens read over the same days, 22% ran. That is not enough to say the live signals pick tokens that run more often than chance.

  • The ratio so far is 0.01 times, but with this much data the true figure could be anywhere between 0.01 and 0.02. That range includes 1, which is no difference at all.
Win rate: closed signals whose token ran
0.2% (10 of 4,540, range 0.1% to 0.4%)
Base rate: tokens read over the same days that ran
22% (45 of 205, range 17% to 28%)
Win rate over base rate
0.01 times, 0.01 to 0.02
Highest market cap seen in the day after the signal, over the market cap at it: the median of 3,027 measured
1.8x (range 1.8x to 1.9x, from 3,027); 88% were below the signal when closed (2,678 of 3,027)

These are flags the engine raised and recorded before anyone knew how they would do. They are counted apart from the replay and never added to it. A live signal can be on any token, while the base rate is of the tokens read (every runner, and a sample of the rest), so the base rate here is far above that of all launches and the ratio leans against the engine. Its range treats the two as separate sets (Katz's interval).

What this was computed on

  • Computed from the record as it stood at .
  • 282 tokens on record with 11,828 early buys. 65 launched before anyone could be scored and 0 in the last two days are not judged; 217 are.
  • The watchlist was drawn 40 times; at the median token it held 40 wallets. On average 0% of them had no activity measurement on file and 0% no funding read.
  • Sales by early buyers: not on record, so the rule that a sale cancels a flag could not be applied. Where each buy was made: on record for at least part of the record.
  • Random draws: 20 lists, started from the number 20261010. Ranges of ratios: 2,000 resamplings of whole days.
  • Fingerprint of everything read and every setting (SHA-256): a73b95c056a13ca6983e93f5ae20a5912489edb0c9c2b6f5acde572cefd65cbd

How the replay works, and why it cannot see the future

The replay goes through every token on record in the order they launched. Before each one it stops the clock at the start of the hour the token launched in, and builds the watchlist the engine would have had at that moment, with the same steps and the same code the live engine uses.

A token teaches the replay nothing until a full day after its launch. That is how long the live engine waits before it judges a token that did not run, and the replay gives a token that ran the same wait. A wallet's buy of a token launched ten hours earlier is in nobody's record yet, whether that token went on to run or not. This is the leak that makes backtests flatter themselves, and a test in the code builds a record where using later outcomes raises a flag and checks that the replay raises none.

Then it asks the live rule, at each second one of the watchlist's buys lands and with only the buys up to that second, whether a flag would have been raised. Tokens launched before anyone could be scored are not judged, and neither are tokens of the last two days, whose own outcome is not settled.

Flags are raised only from buys made in the early part of a token's curve, the part read for every token alike. A token that ran has a longer record, into its pool, after it had in effect already run. A flag raised from those later buys would be a flag on a token already known to have run, so they are left out.

What this does not show

The tokens read are not all tokens. Every token that ran is on record, and only a sample of the thousands that did not. Both shares on this page are therefore far higher than they are among all launches. The ratio between them is pulled toward 1 for the same reason: counting every token that did not run would move it further from 1, not closer.

Survivorship in the source. A token counts as having run when it passed the checks on how it works within a day of launch. One that ran and died between two readings may be on record as not having run. Tokens removed from pump.fun's listings before they were read are not on record at all.

Ran is not the same as paid. A token that ran may have peaked before the flag, or minutes after it. Fees, slippage, failed transactions and the time it takes to act are in none of these figures.

Copying wallets changes their results. The record is of wallets nobody was following. Once many people buy what a wallet buys, the price it gets and the price they get are no longer the same, and its record from before says less.

The replay is not the live engine. It sees one early buy per wallet per token, where the live engine sees every trade. Wallets measured as machines are left out using a measurement made recently, not at the time. Wallets found to buy together are counted as one voice by the live engine and are not by the replay, which only has the funding links. Each of these could move the figures either way.

The ranges assume days are alike. They come from drawing whole days again and again, which allows for good and bad days and not for a market that changes for months.

The past is not the future. Everything here describes tokens that have already launched. It says nothing certain about the next one, and the scorecard lists every signal that went to nothing.

Method, to reproduce it

  1. Take every token whose early buyers were read and which has at least one. It ran if a reading found it running; it did not if it was sampled without regard to how it would do and judged, a day after launch, not to have.
  2. Sort by launch time. For each token, set the clock to the start of its launch hour. The record at that clock is every early buy, made in the 90 days before it, of a token launched at least 24 hours before it.
  3. Base rate at the clock: tokens that ran over all tokens in that record, known only once 20 tokens that did not run are in it. Score each wallet with five or more early buys: the lower end of Wilson's 95% interval for its hit rate, less the base rate, scaled to 0 to 100.
  4. Watchlist: the highest scores, up to the size the live engine is set to, leaving out wallets measured at 1,500 transactions a day or more. Wallets one of which funded the other, or funded by the same wallet within two hours, are one voice.
  5. A flag: at the moment a watchlist buy lands, the set number of separate voices have bought the token within the set window, each at least 0.05 SOL, with less than 30 SOL in the curve, and no watchlist wallet sold it in the window. The first such moment is the flag.
  6. Judge only tokens launched once a base rate existed and at least 48 hours before the record was read.
  7. Win rate, base rate and recall carry Wilson's 95% interval. The two ratios carry the 2.5th and 97.5th percentile of 2,000 resamplings of whole UTC launch days, drawn with replacement. Medians carry the distribution-free interval from order statistics.
  8. Placebo: at every redrawing of the watchlist, for each followed wallet draw at random a wallet with five or more early buys whose count is in the same power of two, not known to be a machine. Run step 5 on that list. Repeat for each list.
  9. State a claim only with 300 judged tokens over 21 days, 30 that ran, 30 flagged, 30 flags from the random lists, and both ratios' ranges wholly above 1. These minimums were fixed before any history was replayed.

The replay is recomputed about once a day as the record grows. Each run is stored with its settings and a fingerprint of everything it read, so the figures above can be traced to exactly one record.

Tracer is a reading of public data. It is not advice and not a guarantee, and a signal is a record of what some wallets did, not a call to do the same. Most new tokens go to zero, including ones that ran; a token appearing here says what happened to it, not what will.