Definitions only. The findings are in the exhibit; every number on this page is a rule, a date or a measurement of the method itself.
Every figure in What a match means was produced under definitions written down and dated before the measurement ran, in a note frozen on 19 October 2026 and nine dated revisions to it, the last of them on 23 October. This page states those definitions in plain words: the control arms, the windows, the exclusion threshold, how a decoy pair is built, the two ceilings this register applies to its own published groups, and the date each definition was fixed. Nothing here was chosen after seeing a number. Where a definition turned out to be unworkable it was corrected in a revision that says what changed and why, and the revision is timestamped before the run it governs.
On it: every definition a reader needs to decide whether a figure in the exhibit means what the exhibit says it means. Not on it: the scripts, the queries and the construction this register uses to link launches by their birth artifacts. Those are the instrument itself rather than a definition of it, and they are made available to a client under engagement, with the run logs and the per-pair outputs behind each figure. A definition withheld would make a figure uncheckable. A query withheld does not.
| Item | Fixed value |
|---|---|
| Source | the collector's own launch database, opened read only. No index was created, no row was written, nothing was copied off the volume. |
| Launch window | 2026-09-02 12:29:00 UTC up to, not including, 2026-10-19 00:00:00 UTC, by the launch's creation time. Narrower windows used by single measurements are named beside them below. |
| Venue | every pair is matched on venue. Cross-venue pairs are never formed, because two venues are two populations of launch tools. |
| Eligibility | a launch enters a document comparison only if its document was read within 60 seconds of the launch itself. |
The eligibility cut is the one exclusion that changes who is in the denominator, so here is its reason. The creator owns the address a launch document is served from. A document first read a day later is what that address served on the day we read it, not necessarily what it served at launch, and two late reads matching each other is therefore not a launch-time match. The claim under test is about launch time. The capture lag is sharply split rather than spread, measured on 10 September 2026 at 80,448 documents read inside a minute, 69,597 read over a day later and 2,174 in between, so any cut between ten minutes and a day selects the same group of launches; 60 seconds is the conservative end of it. The count this removes is reported beside every figure it touches.
An unordered pair of launches signed by two different creator wallets at the same venue, both inside the window. Inside a group of launches sharing one artifact, the unit is a pair of creators and not a pair of launches, and each creator is represented by its earliest launch in that group. A creator holding many launches in one group would otherwise be counted once per launch, and a single prolific wallet would set the rate on its own.
| Arm | Definition | What it stands for |
|---|---|---|
| A. Identical document | the two launches carry the byte-identical launch document, both eligible by the 60-second cut | the signal under test |
| B. Same subject, different document | the same ticker symbol and the same name, trimmed and upper-cased, with the two documents different | strangers and fans launching one trend |
| C. Background | pairs drawn at random from the same UTC calendar day and the same venue, with different documents and no shared ticker | the base rate |
| D. Same wallet | two launches signed by one wallet, used only as a reference rate for how often an operator reuses anything at all | one operator, certain |
Two clauses in the original plan were dropped before anything was computed, and the reason is stated because it cuts against us. The plan had the same-subject arm carry "no shared funder" and the background arm share "nothing". A shared funder is one of the outcomes being counted, so leaving those clauses in would have forced the control rate to zero by construction and manufactured the gap the measurement exists to test. They survive only where the outcome is not funding. For the money outcome the background arm is additionally drawn from wallets whose funding we can already see, because funding is traced for a small share of wallets and a background drawn from all launches would have held fewer than the thirty comparable pairs a rate requires. That changes what the background is a background of, and it is stated that way wherever it appears.
The buyers. A durable buyer of a launch is a wallet with at least one recorded buy of that launch, excluding the creator and the launch's own dev wallet. A pair is comparable only when both launches still carry at least one such buyer after the exclusion below; the outcome is positive when the two buyer sets intersect. The buyer window runs from the launch onward with no upper bound: an earlier version cut it at the end of the launch window, which took the later launches' buyers away and could only understate. The table these buyers are read from carries no market column, so a durable buyer may have bought on the bonding curve or after graduation. A sharper variant restricted to a launch's earliest fifty buyers is reported beside it. These are a different instrument from the curve-buyer figures published in September, and they are never reported as the same number.
The money. A pair is comparable only when both creators carry a funder we have traced and cleared as non-infrastructure, by two tests this project applied before the measurement existed: a funder whose outflows enumerate above two hundred funding transactions an hour, or whose funded wallets mostly do not touch each other's launches, is not cleared. A pair that is not comparable is counted and reported by reason. It is never counted as a negative.
Every buyer figure is a floor. Buy rows are sampled when a launch is finalised and pruned afterwards, so a buyer who bought both launches can have been dropped from one of them. A shared buyer we find is real; a shared buyer we miss proves nothing. That sentence travels with the number wherever it is quoted.
A wallet that buys more than 100 distinct launches is excluded as indiscriminate before any rate is computed. The exclusion is doing real work: on the first run it took the signal arm from 42.47% to 21.92% and the background from 18.04% to 0.34%, which means the separation we would publish rests on where that line is drawn. So the line is treated as a dial and measured as one. Every rate is recomputed at no exclusion at all, and at 500, 200, 100 and 50 distinct launches. The headline stays at 100, which was chosen on 19 October before any of the curve was seen, and the rest of the curve is published beside it. If a separation existed only at one setting, that would be a finding about us rather than about the chain, and it would be said that way.
A pair is comparable only when both launches retain a buyer, and survival is not equal across the arms: in one run it was 14.60% of identical-document pairs against 51.85% of same-ticker pairs. Two launches with large surviving buyer sets intersect more often by coincidence than two launches holding one buyer each, so a difference between arms could be a difference in crowd size. The background arm does not control for that, because it is not matched on buyer count either. So every tested pair is measured against a decoy.
| Step | Fixed rule |
|---|---|
| Who the decoy replaces | for a tested pair, the second launch is replaced and the first is kept, so the comparison is the same launch measured against a stranger |
| Where it is drawn from | the same venue and the same UTC day as the launch it replaces |
| Who it may not be | neither of the tested pair's two creators |
| What it may not share | no shared artifact with the launch it is compared against: not the document, not the artwork, not the ticker and name |
| How it is matched | among at most twelve candidates drawn at random, the one whose surviving buyer count is closest to that of the launch it replaces |
| What is then reported | the same outcome, run on the decoy pair. A cell's lift is its own rate divided by its own decoy rate, and the match quality achieved is published: the median and 90th percentile of the relative gap in buyer count. |
The lift, not the raw rate, is the figure that answers whether a shared artifact carries information. A cell whose decoy cannot be computed for thirty or more pairs is reported as counts only.
| Ceiling | Where it applies | What was measured about it |
|---|---|---|
| 20 distinct creators | an artifact carried by more than twenty distinct creator wallets is excluded before any link is counted: a platform default or a famous meme is not a fingerprint | pairs were banded by the number of creators in the group they were drawn from, with bands at 2, 3 to 5, 6 to 20, 21 to 100 and over 100, so the question "does agreement fall off above the line we drew" is answered by reading down a column rather than by moving the constant and rerunning |
| 100 wallets | a connected group larger than one hundred wallets is published with no group label, because at that size transitive linking stops resolving individual hands | the measurement prices this threshold's neighbourhood, not the threshold. The ceiling applies to a linked component; the band above one hundred here is a band of single-artifact groups, which is not the same object. Stated before the run, and it stays stated. |
A third rule this register applies, to large groups of near-identical wallets, is not tested here at all, and that was recorded before the run rather than noticed afterwards: it rests on agreement sustained over a span, which is a different shape of evidence from a pair. Untested is not disproved, and it is not reported as support either.
| Item | Fixed value |
|---|---|
| Groups to pairs, headline reading | exactly one creator pair per group, drawn at random. Pairs inside one group are not independent observations. |
| Pairs per combination cell | up to 2,000, drawn from that cell's qualifying list |
| The ceilings reading | up to 8 creator pairs per group, drawn without replacement, and in the runs that produced the published figures up to 400 pairs a band. A band holds as many pairs as it holds groups under the one-pair rule, which guaranteed the bands above twenty creators would fall under the floor whatever the data said. That was a defect in the frozen definition and is recorded as one. |
| Smallest cell quoted as a rate | 30 comparable pairs. Below it the cell is reported as a count, with no rate and no interval. |
| Interval, combination cells | a Wilson score 95% interval and a cluster bootstrap over creator wallets, 2,000 resamples. A wallet appears in many pairs, so pairs are not independent draws and a Wilson interval on them alone is too narrow. Where the two disagree, the wider one is the one that travels into anything we publish. |
| Interval, the ceilings reading | a cluster bootstrap over groups, 2,000 resamples, because there the independent unit is the group. No Wilson interval is reported for it. |
| Random seed, everywhere | 20261019. Every draw is repeatable by a second analyst. |
The primary comparison is exactly one shared artifact against two or more, on the rate at the 100-launch exclusion, read as a lift against its own decoy. The seven combination cells, the five creator-count bands and the five exclusion settings are exploratory, and are labelled exploratory wherever they appear, including in anything written to sell. Fourteen rates with 95% intervals will produce one impressive cell by chance, and an exhibit built on its most impressive cell is an exhibit built to be taken apart.
A confound was named before the numbers arrived, and it survived into the published exhibit. A launch document declares the address of its own artwork, so two launches carrying byte-identical documents point at the same artwork by construction and will usually share that artifact for free. Where that happens, one act of copying is wearing two artifacts, and a two-artifact rule that accepts it is weaker than two artifacts sounds. The combination cells are reported separately so this is visible instead of averaged away.
Written down in the frozen note so the conclusion could not be chosen afterwards. If the signal arm's rate were not at least twice the background's on non-overlapping intervals, then an identical launch document does not demonstrate agreement with the money at the coverage we hold, and that sentence gets published instead of the claim. If the signal arm did not exceed the same-subject arm on non-overlapping intervals, the same applies to the claim that a document match says more than a shared subject does. If late copies agreed with the money as often as bursts did, the timing claim is dropped rather than softened. A cell under thirty comparable pairs carries no claim either way.
| Measurement | Window and unit |
|---|---|
| How copies are made | the full record over the launch window, no sampling. For each group of launches sharing one document, the distinct addresses its members declared: one address means a later launch pointed at a file somebody else was hosting, which takes no work; two or more addresses means the same bytes were uploaded again. Reported at group level and at creator-pair level, and split by whether the addresses sit on one host or several. No threshold: this measurement counts, and the counts are exact. |
| The clock inside a group | the spread from the earliest to the latest launch in a group, in fixed bins: under 60 seconds; 60 seconds to under 5 minutes; 5 minutes to under 1 hour; 1 hour to under 24 hours; 24 hours and over. The same bins are reported for groups sharing only a ticker and a name. |
| The gap inside a pair | burst, under 5 minutes; middle, 5 minutes to under 1 hour; late, 1 hour and over. The middle band is reported, not dropped, and a band is only ever compared against another band on the same arm. |
| The buyers | the launch window above, measured on 24 October 2026. Buyer identities are read from a derived table that is kept rather than offloaded, which is what makes a record-wide version of this measurement cheap to run. |
| What a shared document is made of | the chain scanner's own database, chain-wide rather than venue-limited, over the three complete UTC days to 23 October 2026. The unit is the distinct document, classified once from the earliest member's stored bytes. |
| Class | Definition |
|---|---|
| Rich | a description of 40 or more characters after trimming, and at least one of a social handle, a chat link or a website |
| Thin | no description after trimming, and none of those three |
| Middle | everything else |
Variant spellings of a field name are read as the same field, because all of them turn up inside a single day. Counts are reported for all documents and for the shared subset separately, so the two can be compared directly. The question this answers is bounded: if sharing concentrated in the thin shape, the one shape a launch tool can hand to two strangers unchanged, then the document instrument would be weaker than its raw count suggests and the exhibit would have had to be built on the rich subset only. It bounds coincidence. It does not identify who copied from whom, and nothing in it is read as doing so.
| Frozen at (UTC) | What it fixed |
|---|---|
| 19 Oct, 23:30 | the note itself: source, window, eligibility cut, the four arms, the thresholds, the reporting rules, and what would count as a bad result |
| 19 Oct, 23:45 | four definitions that could not be executed as written: the unit inside a group became a pair of creators rather than of launches, the background arm was drawn from the comparable population, the same-subject arm was split into a strict and a loose reading, and the one-pair-per-group draw was pinned |
| 19 Oct, 23:55 | the four measurements that do not depend on funder coverage, defined before any of them ran: how copies are made, the same-wallet baseline, what disappears with the files, and agreement with the buyers |
| 20 Oct, 00:15 | two corrections to the buyer measurement: the buyer window lost its upper cut, and the measurement moved to a window whose buy rows had not been offloaded |
| 23 Oct, 21:33 | the admission rules measurement: the three artifacts, the seven combination cells, the five creator-count bands, and the artwork confound named in advance |
| 23 Oct, 21:46 | four holes found by reading the design adversarially: the decoy pair, the exclusion curve, the redrawn ceilings reading, the two cluster bootstraps, and one primary comparison named in advance |
| 23 Oct, 21:52 | the run had returned six comparable pairs, and the cause was measured rather than guessed: the window's buy rows had been offloaded two minutes after it started. A new window was fixed, and the structural consequence was written down. |
| 23 Oct, 21:59 | the question the plan never asked: what a shared document is made of, with the three shapes fixed before the count |
| 23 Oct, 22:28 | the durable buyer set, which made a record-wide run possible, with the explicit warning that it is a broader instrument and is never reported as the earlier one |
| 23 Oct, 22:30 | a framing correction with no figure attached: buyers, trades and funders are on the chain and do not expire. What has a shelf life is our local copy, which is a property of our disk and not of the evidence. |
A definition is not a result. The figures these rules produced, with their denominators and intervals, are in the exhibit, and the limits published there travel with them: the measurement is on pairs while this register publishes groups, the separation moves with the exclusion setting, and one act of copying can arrive wearing several artifacts at once.
This register records what was launched and does not call any launch authorised or unauthorised, affiliated or unaffiliated, genuine or counterfeit. It does not say who controls any wallet, and nothing on this page identifies a person.