The first five minutes of a launch

In the first five minutes of a launch, the record contains a few dozen events at most, and every visible aggregate is computed from that handful. That is the central fact of the window: not that the information is wrong, but that there is far too little of it to carry the conclusions people routinely hang on it. A reading made at minute one is a reading of a sample, and it should be stated as one.

What follows is the window walked forward second by second, with what exists at each checkpoint, what can be inferred from it, and what would have to be true for the inference to fail.

Why five minutes is the wrong unit and still the useful one

Five minutes is not a natural boundary in a launch. Nothing in the protocol changes at three hundred seconds, and a token that has done nothing in five minutes can do a great deal in the sixth. The unit is arbitrary in every technical sense.

It is useful anyway, for a human reason. Five minutes is roughly the period in which a watcher will decide whether a launch is interesting, and therefore the period in which almost all the reading errors are made. Studying the window is studying the conditions under which people reach conclusions on the least evidence they will ever have.

Solana produces slots at a target interval of roughly four hundred milliseconds, so five minutes is on the order of seven hundred and fifty slots. That sounds like a great deal of room. It is, for the network. For a single new token with a handful of participants, it is a very short list of events spread thinly across a lot of empty blocks.

The window, minute by minute

The timeline below describes the general shape of the window rather than any particular launch. Times are approximate and the point is the ordering, not the clock.

• 00:00Creation landsThe mint exists, metadata is attached, and the curve is initialised. At this instant there is no trading history of any kind, and any figure on the screen other than the address and the time is either a constant or an assumption.
• 00:05 - 00:20The first swaps arriveThe earliest purchases land, often in a tight cluster. A cluster is the expected shape here and carries almost no information, because everyone watching a newest-first feed sees the row at the same moment.
• 00:20 - 01:00The row becomes visible on sorted screensEnough activity exists to clear activity thresholds, so the launch appears on the screens most people actually watch. The population of observers changes here, which changes what happens next.
• 01:00 - 02:00The first sellsThe earliest participants begin to take profit or to leave. This is the first moment the record contains two-sided information, and it is the first genuinely informative minute.
• 02:00 - 05:00The pattern either continues or stopsEither new addresses keep arriving or the same addresses keep trading. Distinguishing those two is the single most valuable thing available inside the window.
• 05:00The first reading that survives an hourBy here the record is large enough that a careful statement about participation is defensible, though a statement about direction still is not.

Second thirty: almost nothing exists yet

At second thirty a launch typically has a small number of buys and no sells. Every derived field on the screen is computed from that. The market value is the curve price times supply, which is arithmetic on a formula, not a market judgement. The holder count is a count of accounts that have existed for less than a minute.

The honest description of this checkpoint is that it contains one fact of any weight: someone created a token and a few purchases followed. Everything else visible is an artefact of arithmetic performed on that fact.

Minute one: the first sample worth the name

By minute one there is usually enough to count rather than merely to look at. The useful measurements at this checkpoint are all counts and spacings, not amounts.

Count the distinct addresses. Count the transactions. Divide the two. A ratio near one means many addresses each doing one thing, which is the shape of a crowd. A ratio well above one means a small number of addresses doing many things, which is the shape of an operator. Neither is a verdict, but they are genuinely different pictures and the aggregate value column shows them identically.

Look at spacing next. Purchases arriving at irregular intervals across the minute look like independent decisions. Purchases arriving in a tight burst inside a few slots look like a single decision executed several times, because that is usually what they are. Slot-level ordering is visible on any public Solana explorer if the feed does not expose it.

Minute two: the shape of the arrival

The second minute is where the record stops being one-sided. First sells appear, and the relationship between new buyers and existing holders becomes visible for the first time.

The question that matters here is not whether there are sells. There will be. It is whether the buying that absorbs them comes from addresses that have not appeared before. New addresses buying while old addresses sell describes a launch that is still recruiting. Old addresses buying and selling among themselves describes a launch that has already stopped recruiting and is now redistributing.

The numbers above are an illustration with round values and describe no real launch. The point they carry is structural: the column most screens lead with is the one column on which the two situations agree exactly.

Minute five: the first reading that survives

By minute five the record is large enough that a careful sentence about participation will usually still be true an hour later. That is a real threshold and worth naming, because it is the first point where reading stops being pattern-matching on noise.

What survives is a statement of this kind: at minute five, this launch had N distinct addresses buying, M of them appearing for the first time in the previous ninety seconds, with sells accounting for a stated share of flow. That is checkable, it is dated, and it will read the same tomorrow.

What does not survive is any statement about what happens next. The window contains no information about the future, and the fact that it feels as though it does is a property of watching a chart form in real time rather than a property of the data.

What a sample this small can support

A worked illustration makes the size problem concrete. Suppose a launch shows forty buys in its first two minutes and you want to know whether the buying is broad. If four addresses account for thirty of those forty transactions, then seventy-five per cent of the visible activity came from ten per cent of the participants, and the remaining thirty-six addresses are contributing one transaction each.

Now change one thing: suppose two of those four heavy addresses were funded from the same source eight minutes before the launch. The count of independent participants drops again, and it drops without any number on the screen changing. This is the general condition of the window. The visible aggregates are stable while the thing they are supposed to measure moves underneath them.

The correct conclusion is not that small samples are useless. It is that they support statements about what happened and not statements about what is true in general. At forty events you can say what those forty events were. You cannot say what kind of launch this is.

Why the same row reads differently at two clocks

A launch seen at second forty and the same launch seen at minute four are not the same object, and the difference is not only that more has happened. The population watching has changed, the screens carrying the row have changed, and the reasons behind each new purchase have changed with them.

In the earliest seconds the watchers are people sorting by newest, a small and unusual group. By minute two the row is on activity-sorted screens, and the watchers are people who saw it because it had already done something. Those two groups behave differently, so the second group is partly reacting to the first group rather than to the token.

This is why a launch can look completely different to two careful people who are both right. They are describing different windows of a moving process, and neither of them has access to the other window. Timestamping your own observations is the cheapest fix available.

The pressure the window applies to you

The five-minute window is designed, intentionally or not, to compress decisions. The row is visibly ageing, the counter is moving, and the interface presents everything as though delay is a cost. Under that pressure people substitute the fastest available reading for the best one.

Three specific distortions show up reliably. Confirmation gets cheap, because the sample is small enough that some subset of it supports whatever you already think. Absence gets read as evidence, so no sells becomes strong hands rather than not yet. And precision gets mistaken for accuracy, because a value displayed to two decimal places feels measured even when it is derived from nine transactions.

The countermeasure is procedural rather than analytical: decide before you open the feed what you are looking for, how long you will look, and what would make you stop. A rule set that exists before the window is the only kind that survives inside it.

It is also worth knowing what a deliberate operator is doing during the same three hundred seconds, because that is a large part of what you are looking at. A Pump.fun volume bot tool exists to place recorded activity on a curve during exactly this window, on a schedule set in advance, which is a plainer description of some of these opening patterns than any interpretation of the shapes.

A checklist for the window

• Timestamp your own observation before you write anything else down.
• Count distinct addresses and transactions separately, and keep both.
• Record the median gap between the first twenty events, not just the total.
• Note the minute at which the first sell appeared, or that none had.
• Check whether buyers after minute two are addresses you had already seen.
• Write one sentence naming what would falsify your reading.
• Do nothing with the reading for a full further minute, deliberately.

The last item is the one that changes outcomes. A reading that still looks correct sixty seconds later has passed a test that most readings made inside this window do not pass, and the cost of the test is a minute.

What the window will never tell you

The first five minutes cannot tell you who is on the other side of any transaction, because the ledger records addresses and not people. It cannot tell you whether a project intends to continue, because intent is never written to a chain. And it cannot tell you what a token is worth, because the quoted price in a shallow book is not a price at which any meaningful quantity trades.

What the window can do, read carefully, is describe participation: how many, how spread out, how new, how two-sided. That is a genuinely useful thing to be able to state, and it is a much smaller thing than the window is usually asked to deliver. The gap between those two is where most of the money goes.

The mechanics of the token accounts underneath all of this are documented in the Solana program library documentation, and reading the account model once makes it much harder to mistake a count of accounts for a count of people.