the research record
we sell measurements, not dreams — including the ones that said no
UTC day 2026-08-29 · data through 2026-08-29 19:20 UTC · regenerated with each publish
The same program that built this product spent a season trying to find tradable edge on pump.fun and published every kill. A signal that dies in testing dies in print, with its window and its method. This page is the reason to believe the numbers on the other pages.
entry prediction is dead — measured, not asserted
We mapped it instead of arguing about it: 542 pre-declared predictor cells; 8 survive FDR at q=0.10; not one clears round-trip friction (exploration map, run 2026-08-15). What survived is telling — the single surviving return cell's top decile loses 0.21% in five minutes — it beats a random board coin only because that coin loses 1.63% (exploration map, replicated out of sample at p=0.005). The streams know a great deal about whether a coin will still be visible and almost nothing about what its price will do.
callout-following is refuted
caller identity alone scored AUC 0.471 [0.354, 0.568] at 1 hour — a confidence interval straddling chance — while free public columns scored 0.796 (callout-edge study, temporally split, run 2026-08-15). In plain terms: knowing who called adds nothing the public tape didn't already know, and the feed itself is an anti-signal — buying the callout feed averaged −11.9% at 1 hour and −43.6% at 8; two or more callers inside 10 minutes → −64.7% median at 8 hours (callout-edge study, 314 callouts / 222 mints, run 2026-08-15).
copy-trading is refuted — at measured power
same-direction co-trading by distinct wallets at short lag ran 0.69–1.03× expected once the market's own burstiness was held fixed; no leader→follower pair survived family-wise testing (copy-trading study, 2,937 live + 8,385 BigQuery swaps, run 2026-08-14). The detector is strong enough for the null to mean something: an injected copier mirroring 20 trades is caught 100% of the time — the null rules out a dedicated follower, and says so at measured power (copy-trading study, run 2026-08-14). What does exist is attention herding through social terminals — a herd hazard, not a mirror hazard.
we publish our corrections
a +21.77% post-entry edge reversed to −12.24% once the censored 96% of coins were priced instead of dropped (board-entry study, 35,031 entries, run 2026-08-14 — the correction stayed published). The wrong number and the right one are both in the study, in that order, because a record you can't watch being corrected is a record you can't trust.
why our nulls are structure-preserving
Coordination detectors love to hallucinate on heavy-tailed markets: with zero coordination planted, naive FDR validation still blessed a mean of 99 false wallet links per world, in 30 of 30 worlds; a degree-preserving null deleted every one (SVN co-trading study, synthetic worlds with known ground truth). Every crew-reuse and co-trading number we publish is tested against nulls that preserve the market's own structure — that is why our Jaccard 0.26 means something and a raw correlation would not.
safety and longevity order in opposite directions
The screen's verdicts rank rug risk. Lifetime, measured on the same launches, runs the other way: the modal CLEAN launch fades quietly inside six minutes — median last trade 5.7 min after birth, 16.4% still trading at 6h, collapse by 24h 0.03%, graduation 0.19% (verdict-survival study, CLEAN cohort n=8,773, fresh corpus 2026-08-26..28). At the far end, BUNDLED coins are the longest-lived, most-graduating cohort AND the most collapse-prone: median last trade 10.3 min, collapse by 24h 3.89% (130× CLEAN), graduation 13.49% (71× CLEAN) (verdict-survival study, BUNDLED cohort n=965, 2026-08-26..28) — the market's honest bargain, both tails fat by construction of who bundles. And KNOWN-CREW is the modal verdict — 85.7% of fresh launches — and those coins mostly just die fast, median last trade 184 seconds (verdict-survival study, 91,505 deployer-identified standard launches, 2026-08-26..28) — so a surface that plays KNOWN-CREW as a rare alarm is miscalibrated; it is the ordinary case, and the verdict names the actor rather than predicting the coin. One comparison is deliberately held: the CLEAN-vs-KNOWN-CREW lifetime ordering missed one registered per-day stability check, so that sentence does not ship until a longer window settles it. The rule is the rule.
mayhem mode: the counterparty is the protocol
pump.fun's opt-in "mayhem mode" is the first launch stratum we refuse to score, and the refusal is now a measurement rather than a shrug — nobody else has published this mechanism. Measured on chain: every mayhem launch mints 2× supply and parks half in one global pump-operated vault — the same address on all 30,831 coins — whose fee-exempt agent starts selling into the curve a median 2 seconds after birth, re-marking virtual reserves at roughly 500× each trade's real SOL (mayhem-arm study + decoded live tapes, 2026-08-26..29); at the 24-hour mark the vault's unsold tokens burn. On the tape it looks like this: one mayhem coin printed a 4,756 SOL market cap three seconds after birth on under 0.1 SOL of real money, and was dead 36 seconds in (full 31-transaction case tape decoded from chain, 2026-08-29). Set the administered prices aside and what remains is thin: the typical mayhem coin's entire human audience is four wallets, and only 7.39% of coins see any human trade after the 24-hour burn (mayhem real-flows amendment, n=30,831 / 18,472 exposure-complete, 2026-08-26..28). And why no re-tuned screen ships: the screen's gates, transplanted into mayhem, admit a set with MORE real rips than the stratum base (88.15% rip-free vs 93.88%), crew fingerprints are simply absent (Jaccard 0.0011 vs a 0.0010 null), and a dirty deployer record runs protective (risk ratio 0.49×) (mayhem-arm study, registered ship rules failed — recalibration 3 of 5 conditions, real-flows amendment concurring, 2026-08-26..28). The crew-and-deployer economy our gates measure does not operate where the protocol itself is the whale, and a re-thresholded arm would be a new unvalidated screen wearing this one's name — so mayhem launches carry the UNSCORED label with these stratum facts attached: facts about the group, never a score for one coin.
the crew ledger's memory holds
The KNOWN-CREW arm rests on wallet fingerprints staying stable across weeks, so we measured that too: crew fingerprints are durable: the 400 busiest returning deployers' new launches matched their own recorded crew 48.5% of the time (strangers: 0.59%), and fingerprint overlap retains 93% of its same-day strength after two weeks (crew-persistence study, window A 2026-08-05..14 to window B 2026-08-26..28, 11 unobserved days between). And recidivism cuts against folklore — 40.4% of fresh births came from deployers already on record two weeks earlier — and the danger is the unseen: no-record deployers' coins collapsed 1.03% vs 0.57% for known-dirty ones (crew-persistence study, 91,505 births 2026-08-26..28 against the 2026-08-05..14 record). The ledger names actors; absence from the ledger is the actual risk marker.
what survived the season
Four findings cleared the same bar the refutations died on, and they are the product: launch-screen validation (the screen), bundled-at-birth risk and crew fingerprints (ibid.), the callout anti-signal (the record), and the crowd's ledger — the crowd netted −738,301 SOL in 10 days; only 32.5% of wallets finished net-positive (wallet-estimator layer over 58,718,411 priced legs, 2026-08-05..14).
Data service only — nothing here is financial advice, and no number is a promise about the future. Every measurement carries its window and method; scores rank risk, they do not establish intent. Built by the DREGG desk on its own research stack.