Head-to-head against settled outcomes

Each night we take the pump.fun mints our radar called the day before, draw up to 300 of them evenly over the day, read RugCheck's free public report for each, and score both against what happened to the mint: rugged within 300 seconds of launch, graduated. Window: 7 of the last 7 UTC days on record (2026-11-07 to 2026-11-13). We are one of the two sides, and the outcome labels are ours (see the caveats). Nothing here is ranked.

Read at the same moment

Method. Same mint, same second, same outcome rule: RugCheck's public report and Jupiter's token record were read a median 2.1 s after the radar's call on the mint, and all three were scored against the settled outcome at least 24 hours after launch.

170,755 mints graded, out of 175,309 read between 2026-10-01T20:54:58Z and 2026-11-14T02:19:47Z. Not graded: 3,054 read less than 24 hours ago, 1,495 whose outcome had not settled when the labels were read, 5 not found in our launch table. Our flag: rug odds of at least 50% at the radar's first call. RugCheck's flag: at least one risk at level danger, the flag used in the day-later tables. Jupiter's flags: an organicScore of 0, and a mint or freeze authority that is not disabled. Each column covers the mints that side was asked about; the counts of reads are below the tables.

All 170,755 graded mints. Rugged within 300 s: 44.6% (76089 of 170755, 95% interval 44.3 to 44.8%).

Reads made within 30 seconds of launch: 145,383 mints. Rugged within 300 s: 47.1% (68498 of 145383, 95% interval 46.9 to 47.4%).

148,135 of the 170,755 graded mints had a record at both RugCheck and Jupiter; the same four columns over those mints are in the JSON under same_moment.all.common_set.

Reads over every mint collected, graded or not: RugCheck answered 155,697, had no record for 17,562 and failed to read 2,050; Jupiter answered 171,905, had no record for 13 and failed to read 3,391. A read was made a median 2.1 s after our call (90th percentile 5.7 s), when the mint was a median 17.9 s old; a request took a median 478.0 ms at RugCheck and 83 ms at Jupiter.

Limits of this section

• The mints are a sample of the radar's first calls, not of all launches: the radar's rug table lists the mints of each minute that our odds rank highest, and at most 4 a minute were read, at even spacing through that minute's new calls. Of the 247,528 first calls the radar listed since 2026-10-01, 175,309 were read; 72,126 were over the per-minute cap, 68 were seen too late to read at the same moment and 25 fell in a back-off. A rate here does not describe launches in general.
• Both comparators were read through their free keyless public endpoints, with no account and no paid key. A paid or keyed form of either service may answer differently. A read that failed (an error, a timeout or a back-off) is counted as no reading, not as no record.
• The outcome labels are ours: our launch table, our recorder and our rule for a rug within 300 seconds. Neither side is scored by a party independent of us. About one launch in ten is quoted in an asset other than SOL, and our SOL-based labels do not measure those; such mints were not removed.
• The flag thresholds are ours to choose: odds of 50% for the radar, any danger risk for RugCheck, an organic score of 0 and a mint or freeze authority not disabled for Jupiter. Beside each precision figure the table shows the share flagged: a flag raised on most answers has a precision close to the share that rugged.
• The rows behind these figures (our odds, each comparator's answer and when it was read, one per mint) are not served on this page yet.

The tables below are the day-later reading. RugCheck was asked a median 12.9 hours after launch there, so its columns differ from the same-moment columns above: a risk listed at launch can be gone from a report read a day later.

Each side over the mints it answered for (7 of the last 7 UTC days on record (2026-11-07 to 2026-11-13))

Our flag: rug odds of at least 50% at the first call (odds of 30% and 70% are in the JSON). RugCheck flag: the report lists at least one risk at level danger (second column) or at level warn or danger (third). Intervals are Wilson 95%.

Same mints for both sides (1,823 mints both answered for)

A paid competitor on the same fresh mints

The one outside wallet we have seen paying for token-safety data in this category bought GET https://api.deepnets.ai/api/token-safety?mint= ($0.0 a call). On 2026-09-30 we bought its answer for 32 pump.fun mints our radar had just called (first call at a median age of 9.7 s, 1.2 to 28.4 s), once at that moment and once five minutes later. It had a record for 0 of 32 at our call and for 0 of 32 at five minutes; every other answer was Token safety record not found. To trigger research for this token, use POST /api/request-token-safety-analysis (requires API key and credits). (HTTP 404 after the payment handshake; nothing was charged). Of 3 older tokens probed by hand, BONK (established, 2023) had a record; pump.fun graduate, 1 day old, pump.fun graduate, 7 days old had none. On the population this page measures, that side gave no answer, so no precision can be compared; the counts and status codes are in the JSON under paid_competitor, its answers are its own and are not republished, and the read is dated: it says nothing about what that service answers later or for other tokens.

Per UTC day

Method

• Population: pump.fun mints whose first call by our radar fell on the UTC day. The radar's rug table lists covered mints under 30 seconds old and its graduation table covered mints under 15 minutes, ten rows each, refreshed every minute in the log (data/tape_api/radar_log.jsonl). Rows named in the graduation-flag repair mask are dropped from the graduation call.
• Sample: at most 300 mints per day, chosen at even spacing through the day's first calls in time order (no random draw). Population sizes and sample sizes per day are in the JSON.
• Our side: the rug odds shown at the mint's first appearance in the rug table. A mint that never appeared there has no rug number from us and counts as no answer.
• RugCheck's side: the summary report from the free public endpoint, one request per mint at no more than one request a second, with no account. A report with no risks is an answer. An error, a missing token or a timeout is no answer; reasons are counted below.
• Outcome labels: our launch table by the public rule (rugged within 300 s of launch; true graduation), settled only once the rule can settle them; unsettled mints are left out of every rate that needs the label.
• Every row is downloadable: /v1/compare/rows (one JSON object per line). Every figure above is recomputed from those rows.

RugCheck requests by result: answered: 2100.

Risk names on RugCheck reports that carried any warn or danger risk (mints per name): Creator history of rugged tokens 332, Copycat token 79, Low Liquidity 41, High holder correlation 7, Single holder ownership 2, Top 10 holders high ownership 1, High ownership 1.

Caveats

• Different moments. Our call was made when the mint was seconds old; RugCheck's report was read about a day later, after the outcome. A risk that appears because liquidity was removed or a creator sold is information from after the launch.
• The population is what our radar listed, not all pump.fun launches. Its tables rank mints by our own odds, so the sample leans toward the mints our model gave higher odds, and a rate here does not describe launches in general.
• The outcome labels are ours: our launch table, our recorder and our rule for a rug within 300 seconds. Neither side was scored by a party independent of us. About one launch in ten is quoted in an asset other than SOL (10 of 100 settled launches read on 2026-09-27), and our SOL-based labels do not measure those; such mints were not removed here.
• Small samples: at most 300 mints a day and seven days. The intervals are Wilson 95% intervals on a systematic sample, not on all launches.
• The flag thresholds are ours to choose: 50% for our odds, and 'any danger risk' or 'any warn or danger risk' for RugCheck. Other choices move precision and recall; our thresholds of 30% and 70% are in the JSON.
• RugCheck's free tier may serve a cached report or build one when asked. We did not set the cache-only option and did not use a paid key. A report with no listed risks may mean nothing was found or that little had been read.
• The job reads RugCheck's public API at no more than one request a second. If the service asks us to stop, the job stops and this page stops updating.

The same figures as JSON (free): /v1/compare. Rows: /v1/compare/rows. Our own dated scorecard: /track-record.