Classifying a surge: the order the questions get asked in

The short answer

Surge classification is a sequence, not a lookup table. The structural questions run first because they are cheap and, when they return something, decisive. The behavioural questions run afterwards because they are expensive to measure and every one of them carries an ordinary explanation that can defeat it on its own.

Getting that order wrong is the commonest failure in practice, and it is not a measurement problem. It is a narrative problem: an analyst who starts with the interesting behavioural question has usually built an explanation before the boring structural fact arrives, and explanations are extremely hard to dislodge once written down.

Why order matters more than method

Consider two analysts looking at the same window. The first checks whether pooled reserves moved, finds a large liquidity addition, and classifies the window as plumbing in about forty seconds. The second starts by measuring spacing, notices a suspiciously regular cadence, spends twenty minutes building a case for produced turnover, and only later discovers the liquidity event that explains everything.

Both analysts measured correctly. The second one produced a worse reading, and did so because of ordering alone. Worse, the second analyst now has a written argument to abandon, which is psychologically much harder than never having written it. Sequence is a defence against your own commitment, not just an efficiency measure.

There is a second reason. Structural facts are recorded on chain and can reach the Firm confidence band. Behavioural distributions are inferences and never reach it, however many of them agree. Running the checks that can produce a Firm result before the checks that cannot is simply putting the strongest available evidence first.

The sequence, step by step

Seven steps, run in this order every time. The first three are cheap enough to complete inside two minutes on a live window; the rest need the window to have closed or to have run long enough that the distributions mean something.

1. Fix the window and the comparison periodWrite down a start, an end and the earlier period the window is being measured against, before looking at anything else. A window chosen after the data has been seen can support almost any conclusion without a single false sentence being written.
2. Check the reservesDid pooled liquidity change during the window, and was a pool created or drained? This is a recorded fact rather than an inference. When it comes back positive it settles the class as plumbing and the sequence can stop.
3. Check the venue spreadHow many programs carried the flow, and did one pool take nearly all of it? A surge that appears on one venue while a deeper venue stays quiet is a different event from one that appears everywhere at once.
4. Measure end-of-window inventoryDid participants finish holding something different from what they started with? Inventory that returns to flat after heavy turnover rules out distribution and points at routing or produced flow. Inventory that ends strongly directional does the reverse.
5. Measure address novelty and tail growthWhat share of trading addresses had never appeared in the pair, and was the address list still lengthening at the close? A crowd keeps producing addresses nobody has seen before. A fixed set of wallets exhausts itself.
6. Measure spacing and size dispersionHow are the gaps between trades distributed, and how widely are sizes spread? This is where a configuration and a crowd diverge most, and also where the most ordinary explanations are available, so it runs late rather than early.
7. Check lead-lag against the earlier moveDid this burst precede or follow the first one? An echo cannot precede what it echoes, so this is the only check that can eliminate the echo class outright, and it needs the comparison period fixed in step one.

The sequence is deliberately front-loaded with checks that can end it. Steps two and three between them resolve a meaningful share of windows with a recorded fact, and every window they resolve is a window nobody has to argue about later.

The two boundaries that cause trouble

Four of the six classes separate cleanly enough in practice. Two pairs do not, and knowing in advance which pairs are difficult is more useful than pretending the scheme is crisp.

The routing and produced boundary is the one worth understanding properly, because the two have almost the same footprint and completely different meanings. Intermediation exists to close a price difference and stops when the difference is gone. Produced activity has no such stopping condition; it stops when a budget does. Persistence in the absence of any gap is therefore the sharpest available separator.

The attention and echo boundary is easier to fix and easier to get wrong. It is fixed by choosing a comparison period that starts before the first move rather than at the exciting part. Analysts who begin watching at minute four routinely classify the echo as the event.

What sits behind the produced class

S5 is the only class with a purchasable supply chain, and understanding that chain sharpens the classification. Different tool classes produce visibly different footprints, and treating them as one category is why produced windows are so often misread.

The clearest distinction is between activity spread deliberately over time and activity packed into a single block. A tool that sequences swaps across many wallets on an interval leaves a spacing signature; a bundler that executes several transactions atomically leaves almost no spacing information at all, because everything lands in the same slot. A comparison of volume bot vs bundler behaviour is therefore a practical classification aid rather than a shopping exercise, since the two produce different shapes and require different checks.

The general point is that an automated Solana volume bot exposes a small number of parameters to an operator, and those parameters map directly onto the fields in step six. Wallet count constrains address novelty. Interval settings constrain spacing. A size band constrains dispersion. Knowing the parameter set is what makes a narrow band interpretable rather than merely odd.

How a classification is written down

A classification that is not written down in a fixed shape drifts. The shape below is not bureaucracy; each line exists because leaving it out has a specific failure attached.

• The window, as a start and an end, with the time source named.
• The comparison period, stated separately, because the multiple depends on it entirely.
• Venue coverage, including the venues knowingly left out.
• The deduplication rule applied to multi-hop routes.
• Which steps of the sequence were run, including the ones that returned nothing.
• The class, or the two candidate classes, with no third option smuggled in as a hedge.
• The confidence band, with the reason it is not the band above.
• The single observation that would move the class, written before the prose.
• A plain sentence saying what the reading does not claim.

The falsifier line does the most work and is the most often skipped. Writing it before the explanatory prose forces a decision about what evidence would have been needed, which is a far harder question than describing what was found, and it has to be answered while the reading is still uncomfortable.

When undecidable is the answer

Undecidable is a finished result, not a failure to finish. It is the correct output when the fields that would separate the surviving candidates were not measurable in the window available, and it is common for windows read live.

The temptation is always to round undecidable up to the nearest interesting class, because undecidable is unsatisfying to publish and impossible to build a narrative on. That temptation is exactly why the band exists. A desk that never publishes undecidable is not being decisive; it is quietly converting absent evidence into confident readings.

Two situations produce it most reliably. Windows too short for address novelty and tail growth to have said anything, and windows where venue coverage is known to be incomplete in a way that could invert the reading. Both are stated in the record rather than glossed.

There is a third, less obvious case: windows where every field agrees but a single ordinary explanation accounts for all of them at once. Agreement among fields is only informative when the explanations that would defeat them are incompatible. If one scheduled rebalancing process would produce the regular spacing, the narrow size band and the flat inventory together, then three agreeing fields are really one observation counted three times, and the correct output is undecidable rather than a confident produced reading.

Naming two classes at once

When two classes survive the sequence, both are named. This looks like indecision and is the opposite: a reading that lists two live candidates has said something specific about four classes that were eliminated, and it tells a reader exactly which further observation would resolve it.

The failure mode it prevents is worse than it looks. Forced to pick one of two surviving candidates, an analyst will pick the more quotable, and the more quotable is almost always the one implying somebody did something deliberate. Over time that bias accumulates into a record that systematically overstates deliberate activity.

What the sequence cannot fix

Ordering the questions well does not manufacture evidence. If venue coverage is incomplete, running the steps in the right order produces a well-organised reading of the wrong data. The sequence protects against narrative bias and wasted effort, not against missing rows.

Nor does it help with a window that is still open. Steps five and six both need the distribution to have stabilised, and a distribution measured on the first ninety seconds of a burst is dominated by whichever participants happened to be fastest. That is a real limitation rather than an inconvenience, and it is why the triage procedure for live windows stops at the structural checks and says so plainly instead of pretending the behavioural fields are ready.

It also cannot reach intent. Every step above measures what accounts did. None of them touches why, who instructed it, or whether anybody involved considered it legitimate. A window that lands squarely in the produced class has been described, not accused, and the distinction is maintained deliberately throughout this site.

Finally, the sequence assumes the six classes are adequate. They are a working scheme, not a discovery, and a surge shape that none of them handles is a genuinely useful finding. The scheme is meant to be revisable, and a case that breaks it is worth more than a case that confirms it.

Questions the desk gets asked