Findings

Nearly half of the graduations we watched had no outside buyer

Of 16,442 tokens that completed a bonding curve on pump.fun, which this archive watched from the creation transaction and confirmed against the curve account itself, 7,796 (47.4%) had no outside buyer at all. Not one wallet other than the creator ever bought on the curve.

The creator's share of supply is set by the creator's own first-block trades, and it is the single choice a launch cannot later disown. In 8,043 of the 16,442 watched graduations (48.9%), at least half of total supply went to the creator's wallet in the first block.

No outside buyer, and at least half the supply taken at birth: 7,646 launches (46.5%) show both, which is to say the curve was completed by its own creator from a position that already held the float.

A bonding curve absorbs roughly 85 SOL before it completes. 4,955 of the watched graduations (30.1%) absorbed it inside a minute of the creation transaction. Speed alone concludes nothing; it is a birth fact this record holds and a later check cannot recover.

These are facts about the first blocks of a token's life, and they stop being observable by inspection almost immediately. Once an operator spreads the float across wallets and funds the pool with real SOL, a launch whose creator took the supply and one with many buyers look the same under inspection: pool depth reconciles, holder concentration looks ordinary, mint and freeze authority are clean. Every present-tense check passes.

One launch in this record was created with 79.3% of supply taken by its creator and zero outside buyers, and graduated. Hours later its pool held 2,043 SOL against the 2,027 that constant product predicts for its market cap, and its largest holder was 4% of supply. Nothing measurable that afternoon showed what the first block recorded. That is not a gap in inspection. It is a property of inspection. Recovering a first block after the fact takes a full replay of the launch's history from archival infrastructure - per launch, at cost, which is the reconstruction we offer for launches we did not watch - and no replay recovers the launch's files once their host has stopped serving them, or makes a later analysis contemporaneous.

A finding is a live measurement: every figure on this page is recomputed from the record at each build and moves as the archive grows, dated where it stands. An exhibit is one finding worked into dated evidence, frozen when published, identifiers truncated. A report is a dated publication with the queries to reproduce it, frozen the same way: the published reports, and a commissioned provenance report is the same instrument run on one launch, in full. The corrections register applies to all three.

One finding at a time, as a dated, self-contained page: what was measured, the days it happened on, and the records behind every count, identifiers truncated. The figures in each exhibit are from its stated date, not from the live record, and they do not move after publication.

A graduation reaches us as an event on a feed, and an event is not a curve. Where a graduation was independently confirmed (by the pool existing, or by the curve account's own complete bit), reading the curve again disproved 97 of the 1,756 we re-read (5.5%). Those hold up.

The graduations we could not confirm are a different population. We have read the curve for 16,360 of the 18,109 of them, and 16,248 (99.3%) had not completed. A threshold crossed on a feed, and no curve behind it.

We publish both rates because a register that corrects itself and will not say how often is asking to be taken on trust. Neither is a rate over all graduations: we check the doubtful ones far more often than the settled ones, deliberately, so an unsplit figure would describe our own sampling rather than the market. A further 110 were read after the account had already gone: we looked and learned nothing, counted as neither.

These figures were computed on 2026-09-26 01:24 UTC, from the record this page was built against, and they are a static snapshot: the archive keeps collecting and the counts keep rising, so a number here is a fact about that moment and not a live total. The queries below return the current answer from the current file.

Coverage begins 2026-09-02 12:29 UTC. A token that launched before then was not watched and this archive answers UNKNOWN for it: the honest answer, and not a useful one. Launches before coverage can be reconstructed from chain history on request, and are marked as reconstruction.

The population above deliberately excludes two kinds of row, and the exclusions matter more than the headline. A token that a detector restored after its launch carries a zero buyer count because nobody was watching it, not because nobody bought. There are 5,964 such rows and counting them would inflate this finding by 76.5%. A launch rebuilt from chain history is not an observation either. Both are excluded here and both are labelled in the file.

None of this says a token was a fraud, and none of it is advice about anything. It says what the chain recorded in the first blocks, which is a narrower claim and the only one we can support.

These are the queries above, verbatim. Run them against the record database, which downloads with an API key, and disagreeing with us is the point of publishing them.

Record database: record.db, with an API key. What every column means, and how a launch is checked.

Every number above links to its records. The lists of launches and wallets behind each count open with a Pro account; the archive behind all of them is available under a data agreement.

Cite a figure from this page as: Chain of Title, findings as of 2026-09-26, chainoftitle.org/findings.html. Figures here move as the archive grows; for a citation that must hold still, use a dated exhibit or report.