On October 10, at the next settlement window, Pump.fun will begin paying the people who talk differently. The public record is one announcement from the platform's co-founder, Alon. Three facts are legible inside it. The call reward is no longer indexed primarily to the trading activity a caller generates. It is indexed to the realized return of the followers who acted on the call. And recommendations of low-market-cap tokens earn less. No formula accompanies the change. No reference implementation. No test vector. No settlement example. No audit trail.

I have read mechanisms like this before they ship, and I have read them after. In late 2017, inside a line-by-line audit of the 0x Protocol v2 order-matching engine, I found an integer overflow that could have drained liquidity pools; the team delayed launch six weeks to fix it. The lesson I carried out of that room was never about the overflow. It was that a mechanism is only as honest as the specification you can verify. Pump.fun has shipped a settlement change with no verifiable specification. Silence is the only honest ledger. Right now, the ledger is silent.

Context: what the platform is, and what the change touches

Pump.fun is an application-layer platform on Solana. It lets a user mint a token in seconds, trade it along a bonding curve, and, if the token reaches a threshold market cap, "graduate" its liquidity to an external decentralized exchange. The platform is one of the highest-throughput consumer applications on the network. It is not a protocol in the cryptographic sense. It is a product with a server, a matching system, and a payout engine. Its value capture is fee capture: every trade pays a fee, and the platform routes a portion of that fee revenue toward the people who bring traders in.

The mechanism under revision is the call reward. A "call" is a public recommendation. A caller posts that a token is worth attention, followers buy, and the platform pays the caller for the traffic. Under the prior regime, the reward tracked activity — the volume a caller's referral produced, or the number of trades attributable to that caller's link. This is the classic referral model. It pays for attention, not for outcomes.

The new regime is different in kind. It proposes to pay for outcomes. The caller's reward becomes a function of the follower's realized profit or loss. That is the whole of the public disclosure. Everything else — how a follower is identified, how a return is computed, whether the return is realized or unrealized, whether slippage and fees are netted out, how wash trades are filtered, what the settlement cadence is — is undisclosed.

I am going to reconstruct the design space, because a mechanism that cannot be read must be inferred, and inference has to be done with the same rigor as an audit. I will be explicit about the boundary between what the announcement states and what the mechanism must contain if the stated goal is to be met.

The framing matters. Pump.fun sits at a chokepoint in the Solana meme economy. Above it is the L1 — execution, settlement, priority fees. Below it are Raydium, where graduated pools migrate; Jupiter, the aggregator; the DEX bots and snipers that specialize in new launches; and the Telegram trading terminals that route retail flow. The call reward is not a side feature bolted onto the product. It is the mechanism that converts a retail social network into a distribution channel. Change it, and the incentive gradient changes for every participant who touches the funnel.

It is also worth stating plainly that this is an incentive-design revision, not a technical upgrade. There is no new consensus, no new cryptography, no new contract to audit. There is a parameter change inside a payout engine. That makes it harder to analyze, not easier. Code has a hash. A parameter change has a policy, and a policy has no hash.

Core: the mechanism, dissected

What actually changed

Strip the language and the change is a parameter revision inside an incentive design. The prior mechanism was a volume-indexed rebate. The new mechanism is a performance-indexed rebate. In accounting terms, the platform is moving from paying a commission to paying something shaped like a carried interest — except the "carry" is not paid out of a fund's gains. It is paid out of the platform's fee pool and indexed to someone else's profit and loss.

This is a materially different instrument, and it introduces a materially different set of failure modes. A volume rebate is trivially computable. You sum the notional traded under a referral code and multiply by a rate. A performance rebate requires two engines that did not previously need to exist: an attribution engine and a P&L engine. Both must be correct under adversarial conditions. The first is arithmetic. The second is forensic accounting performed live, at high frequency, against counterparties who are actively trying to game it.

The platform has not published either engine. That is the finding. Everything downstream of this paragraph is an inference about what the engines must look like if the stated goal is to be met.

The prior regime and why it failed

The volume rebate had a known pathology. It paid for activity, and activity is trivially manufacturable. A caller who front-ran their own call — buying before publishing, letting followers bid the price up, selling into the bid — generated volume with both legs of the trade. The mechanism paid the caller for the pump and for the dump. It subsidized the exact behavior it should have deterred.

This is not unique to Pump.fun. It is the structural failure of every attention-indexed incentive. When you pay for clicks, you get click farms. When you pay for impressions, you get bot traffic. When you pay for volume, you get wash trading. The metric becomes the target, and the target becomes the fraud. I have watched this pattern repeat across four market cycles. The only durable defense is to pay for a quantity the participant cannot cheaply manufacture. The platform's attempt to pivot from volume to follower return is an attempt to find such a quantity. Whether it succeeded depends entirely on how hard follower return is to manufacture.

The attribution problem: identity

To pay a caller based on a follower's return, the platform must answer four questions per follower. Who is the follower? What did they buy? At what price? And when did they exit?

Question one is identity. A follower is a wallet, or a set of wallets, that acted on a specific call. Attribution means binding a trade to a recommendation. In a public mempool, that binding is not cryptographic. It is statistical. The platform can observe that a wallet bought after a caller posted, but "after" is not "because of." Attribution here is a correlation, not a proof. And correlations can be manufactured.

Consider the cleanest attack. A caller operates two wallets: the public calling wallet and a quiet follower wallet. The caller posts a call, the follower wallet buys, the price rises on genuine follower demand, and the follower wallet sells at a profit into that demand. The follower's realized return is positive by construction. The caller collects a reward on a return the caller manufactured. The attribution engine sees a textbook sequence: call, buy, profit, exit. Nothing in the sequence is anomalous unless the engine specifically looks for counterparty clustering between the caller's wallet and the follower's wallet. No published defense suggests it does.

Now scale it. A caller with a cluster of ten wallets can generate ten attributed followers per call, each with a manufactured positive return, and harvest a reward proportional to the aggregate. The manufacturing cost is the fee on each round trip. If the reward per follower exceeds the round-trip fee, the attack is profitable, and it is profitable by design, because the reward was sized to compensate a genuine caller's effort, and the attacker's effort is lower.

The attribution problem: pricing and cost basis

Questions two and three are pricing. The follower's entry price is observable on-chain. The exit price is observable on-chain. But the follower's return is not simply exit minus entry. It is exit minus entry, minus the platform fee, minus the curve spread, minus the priority fee, minus the slippage, minus the MEV tax extracted by the snipers who front-run the bonding curve. A meme trade on Solana is a stack of costs, and the honest return is the residual after every cost is subtracted.

This is where the cost-basis decision becomes a policy question with a payout attached. If the platform computes return gross of costs, the caller is paid on a number the follower never realized. The follower's actual outcome is lower, sometimes much lower, and the reward overstates the benefit the caller delivered. If the platform computes return net of costs, the caller is paid on a number that depends on the follower's execution quality — which the caller does not control and cannot predict. A follower who uses a slow terminal and pays heavy slippage produces a worse net return than a follower who routes through a fast bot, for the same token and the same call. The caller is penalized for the follower's infrastructure.

There is no neutral choice here. Gross return overpays the caller and misrepresents the follower's experience. Net return underweights the caller's contribution and imports the follower's execution noise into the payout. The platform must pick one, and the choice changes who gets paid. The announcement does not say which.

The attribution problem: time and the snapshot

Question four is time. A return is realized only when a position is closed. If the platform measures realized returns, it must wait for exits, which introduces settlement lag and the risk that followers hold indefinitely, leaving the reward perpetually unvested. If it measures unrealized returns, it must mark positions to a price, which means the reward becomes a function of a snapshot — and snapshots are manipulable.

The snapshot attack is the oldest trick in performance-fee fraud. A caller holds a follower's mark-to-market above water at the measurement instant, harvests the reward, and lets the price collapse afterward. The follower's final return is negative; the caller's reward was paid on the transient mark. If the settlement cadence is short and the measurement price is easy to nudge — and on a thin bonding curve, the price is easy to nudge — the attack is not exotic. It is arithmetic.

The platform could defend against this by vesting rewards over a holding period, or by measuring only closed positions, or by requiring the follower to remain flat for a window before the reward settles. Each defense has a cost. Vesting delays caller income and may drive callers away. Measuring only closed positions delays settlement indefinitely. Requiring a flat window penalizes legitimate followers who hold through volatility. The platform has disclosed none of these choices, which means the attack surface is whatever the undisclosed implementation happens to leave open.

Here is a signature I want to place precisely. Code does not lie; intent does. The code will compute whatever the platform programs it to compute. The intent — whether the number shown to the caller is the number the follower actually earned — is not in the code. It is in the specification. And the specification is absent.

The low-market-cap de-weighting as a diagnostic

The second leg of the change is more legible and, to my eye, more diagnostic. Rewards for recommending low-market-cap tokens are reduced. Why would a platform specifically de-weight low-cap recommendations?

Because low-cap tokens are where the pump-and-dump lives. A low-cap token has a thin order book. A small amount of capital moves the price a large amount. That is precisely the condition under which a caller can buy ahead of their own call, let followers bid the price up, and sell into the bid. The reward for that behavior under a volume-indexed regime was maximal, because the caller was paid for the volume their call generated, and their own front-run generated volume. The mechanism subsidized the very behavior it should have deterred.

De-weighting low-cap recommendations raises the opportunity cost of the pump-and-dump. That is the charitable reading, and it is probably correct at the level of intent. But note what the change concedes. To de-weight low-cap calls specifically, the platform must already know that low-cap calls are the problem. Which means the platform has internal data showing that low-cap calls are the problem. The announcement is a confession written as a policy.

There is a second concession buried here. If low-cap calls were the dominant source of caller revenue, then de-weighting them removes the dominant source of caller revenue. The callers whose economics depended on low-cap pumping will see their payouts fall. Some will adapt to higher-cap tokens. Some will migrate to a competitor. The platform has decided that the reputational cost of hosting the pumpers exceeds the fee revenue they generate. That is a real trade-off, and it is the kind of trade-off a platform makes only when the reputational cost has become visible in the retention data. Platforms rarely reform while the revenue is still clean. They reform when the revenue is contaminated and the contamination is showing up in churn.

The adversarial surface, enumerated

Any system that pays for outcomes invites the manufacture of outcomes. Let me enumerate the attack surface as I would in an audit, because the platform has disclosed no defenses.

First, wash trading between colluding wallets. Covered above. The engine sees a clean sequence and pays a reward on a manufactured return.

Second, selective disclosure. A caller can call many tokens and publicize only the winners. If the reward is indexed to follower returns, and followers are only counted when they follow a publicized call, the caller can curate the sample. The losers are never called, so the followers who would have lost are never attributed, so the reward reflects only the survivors. This is survivorship bias as an income strategy, and it is invisible to any engine that measures per-call rather than per-caller. The defense would be to require a caller to publish every call, or to measure the caller's reward across their full call history including the losers. The announcement does not say the platform does either.

Third, timing arbitrage. If the reward is computed on a snapshot, the caller can influence the snapshot. Buy the token into the snapshot, let the mark rise, harvest, exit after. If the reward is computed on realized exits, the caller can influence the exits by coordinating a bid at the moment the follower sells. Either way, the caller controls the timing of the measurement, and the measurement is the payout.

Fourth, the oracle problem. The return computation depends on prices. If those prices come from an off-chain feed, the feed is a dependency. In early 2024, I audited a DeFi protocol that let autonomous agents drive yield farming from off-chain data feeds. The oracle had no cryptographic verification of its inputs. The yield calculation was manipulable at the source. I flagged it publicly, and the project pivoted to a hybrid model with zero-knowledge proofs for data integrity. Pump.fun's return computation has the same shape: it ingests prices and produces a payout. If the price ingestion is unverified, the payout is unverified. Verify the hash, trust no one. There is no hash here to verify.

Fifth, the meta-attack on the metric itself. If the reward is indexed to follower returns, the caller's optimal strategy is to select followers who are already likely to profit — followers who buy early, in size, and exit well — and to avoid followers who are likely to lose. But the caller cannot choose their followers; followers self-select. The only lever the caller has is to make the call look attractive, which is to say to attract the followers most likely to buy into strength and least likely to be the last buyer. That lever is indistinguishable from the pump-and-dump lever. The metric that was supposed to punish manipulation can, under the wrong implementation, reward the same skill set.

The negative-sum ledger

Step back from the mechanics and look at the arithmetic of the venue itself. A meme trade on a bonding curve is not a zero-sum game between the caller and the follower. It is a negative-sum game among all participants, because every trade pays a fee, every trade pays slippage, and the priority-fee and MEV layer extracts a further tax. The house takes a cut on every rotation of the same capital.

Against that arithmetic, the promise "we will pay callers based on your returns" cannot make the venue positive-sum for followers. It can only redistribute who loses. If the mechanism works exactly as described — callers paid only when followers profit — then callers are incentivized to call tokens that go up. But a caller cannot make a token go up except by attracting more buyers, and attracting more buyers is the mechanism by which the last buyers lose. The structure is self-referential. The reward is funded by the churn that produces the losses.

I want to be careful with the word "Ponzi," because it is thrown around loosely and it deserves precision. The call reward itself is not a Ponzi. It is funded by real fee revenue, not by new deposits. But the venue it sits on has a Ponzi-adjacent property: the aggregate of followers' returns is negative after costs, so the average follower's outcome is a loss. A mechanism that pays callers for follower returns is therefore paying for a statistically rare event and marketing it as the norm. Ponzi schemes leave trails in the data. The trail here is not in the reward. It is in the aggregate P&L of the follower cohort. If that cohort is net negative — and the fee structure guarantees it is — then the mechanism's success metric and the venue's base rate are in direct conflict.

Let me put a number on the intuition, using only public structure and not fabricated data. Every round trip on a bonding curve pays the platform fee, the curve spread, the priority fee, and any MEV. Call that the friction constant. For the median follower to break even, the median token must appreciate by more than the friction constant. Meme tokens have no cash flows. Their price is entirely reflexive: it moves because people buy, and people buy because it moved. There is no anchor. So the median token's expected appreciation is zero minus friction. The median follower's expected return is negative. A caller paid on follower returns is paid on a distribution whose mean is below zero. The mechanism can only be profitable for the caller if the caller's selection skill beats the friction. For a few callers, it will. For the cohort, it cannot. That is the statistical ceiling on the whole design.

Downstream propagation

Pump.fun's position at the chokepoint means its parameter change propagates. The propagation path is legible: platform incentive change, then caller behavior change, then volume and graduation-rate change, then downstream DEX and infrastructure load change.

Start with the callers. The callers whose revenue depended on low-cap volume now face a lower payout. Their rational responses are three. Migrate to a competitor with a more generous caller policy. Shift their calling toward higher-cap tokens where the reward is not de-weighted. Or exit. The first response moves volume off the platform. The second moves volume up the market-cap curve, which reduces the supply of the thin, easily-pumped tokens that made the platform's content cycle so fast. The third reduces the number of callers, which reduces the top of the funnel.

Now follow the volume. If caller-driven volume falls, the volume that would have graduated into Raydium falls. Raydium's graduated-pool inflow is a function of Pump.fun's graduation rate. Jupiter's routing volume includes the flow that exits graduated pools. The trading bots and terminals that specialize in new launches see fewer launches. The impact on any single downstream is small. The platform is one application, not the L1. But the direction is coherent, and the magnitude is measurable within a quarter.

There is a second-order effect on Solana's fee market. Pump.fun is a heavy consumer of priority fees. If its activity falls, the priority-fee demand attributable to it falls, which marginally relieves congestion and marginally reduces the fee burn attributable to that activity. The magnitude is small relative to the L1's total throughput, and I would not overstate it. But a platform that is a top consumer of block space is a platform whose behavior change is visible in block-space economics.

Audit the edges, not just the center. The center of this story is the call reward. The edges are where the change shows up as volume, as graduation rate, as routing flow, as block-space demand. The center is undisclosed. The edges are observable. That is the arbitrage available to anyone who wants to verify the mechanism's effect without the mechanism's specification: measure the edges and infer the center.

Governance and rule durability

The change was announced by a single co-founder. There was no vote, no forum discussion, no staged rollout, no comment period. This is the governance model: centralized. The platform sets the parameter; the users and callers live with it.

I do not consider centralization a moral failing. A product with a server and a fee engine is centralized by construction, and centralization buys decision speed. The platform can correct a misaligned incentive in a single post, which a decentralized autonomous organization could not. But the same speed that lets it correct the rule lets it change the rule again. The relevant risk for a caller is not that the current rule is unfair. It is that the rule is not durable. A caller who invests in building a following under the new regime has no guarantee the regime persists past the next announcement. Rule uncertainty is itself a cost, and it is priced into how much a rational caller is willing to invest in the platform.

There is a deeper point about accountability. The platform now pays callers based on follower outcomes. That makes the platform the arbiter of what a "follower outcome" is. The platform defines the identity, the prices, the time window, and the cost basis. Every degree of freedom in that definition is a degree of freedom in the payout. When a single entity both computes the metric and pays against it, the metric is only as trustworthy as the entity. And the entity has not published the metric. Truth is found in the source code. Until the source is published, the truth is found in the announcement, which is to say it is found in the marketing.

I have seen this exact structure before, at larger scale. In November 2022, during the FTX bankruptcy proceedings, I was contracted to review the exchange's internal ledger discrepancies. I traced roughly eight billion dollars in missing funds through unrelated wallet addresses and linked them to a proprietary trading desk. Customer assets were not merely mixed. They were commingled and risked on speculative trades without collateral. The forensic report ran two hundred pages. The recurring finding across every chapter was the same: an entity that both held the assets and computed the accounting had no internal control that could catch its own errors, because the control and the entity were the same. Pump.fun's payout engine is smaller in scale and different in kind, but the structural lesson holds. When the party that computes the number is the party that pays the number, the number is a claim, not a fact, until an independent party can reproduce it.

Regulatory surface

I will keep this short and clinical, because the regulatory analysis is speculative and the platform has no token, which removes the most obvious securities question.

The relevant exposure is not "is the platform's token a security," because there is no token. The exposure is the marketing-and-manipulation frame. A mechanism that pays promoters based on the returns of the people they promote is, in substance, a paid promotion scheme. Paid promotion of speculative assets is regulated in most major jurisdictions, and the disclosure requirements are strict. The platform's mechanism does not appear to impose disclosure on callers. It pays them. If a regulator reads the mechanism as an unregistered promotion scheme operating at scale, the platform's centralization — the fact that it computes and pays — makes it a legible target in a way a decentralized protocol would not be.

There is an irony worth stating. The adjustment may itself be a defensive move. By tying caller pay to follower outcomes, the platform can argue it is discouraging the exploitation of retail. But the same adjustment documents, in the platform's own words, that the platform ran a paid-promotion system whose participants were exploiting retail. The defense and the evidence are the same document. Whether a regulator reads it as remediation or as admission depends on facts the platform has not disclosed, which is to say it depends on the same missing specification.

The specification gap, consolidated

I want to consolidate the core finding, because it is easy to lose in the detail. The finding is this. The platform has changed the objective function of a large distribution network from "generate activity" to "generate follower profit," and it has not published the function that computes "follower profit." Every critique I have offered — attribution, wash trading, selective disclosure, snapshot timing, oracle integrity, negative-sum base rates, rule durability — reduces to that single gap.

The mechanism may be well-designed. It may be a genuine attempt to align callers with followers. The de-weighting of low-cap calls is evidence of good intent, because it is costly and precise. But a mechanism that cannot be read cannot be audited, and an unaudited mechanism that pays money is a mechanism that will be gamed. The gap is not a footnote. It is the whole of the risk.

I spent three months in 2022 cross-referencing on-chain data against the Anchor Protocol's tokenomic whitepaper, and I found a mathematical impossibility in the reward distribution. The nineteen percent yield was not yield from trading fees. It was a distribution of newly minted tokens to early depositors, funded by dilution of late ones. I published a fifty-page breakdown, and regulators cited it in the investigations that followed. The lesson was not that the yield was high. The lesson was that the yield's funding source was undisclosed and therefore unverified, and that the disclosure, once forced, told the whole story. Pump.fun's call reward has the same shape. The reward's funding source is disclosed — fees — but the reward's computation is not. The half that is disclosed is the reassuring half. The half that is not disclosed is the half that determines who gets paid.

Contrarian: what the bulls get right

Now the part the bulls get right, and they get more right than the bearish read admits.

First, the direction of the change is correct, and it is rare. Most platforms facing the "our promoters are fleecing our users" problem do nothing, because the promoters generate the volume and the volume generates the fees. Pump.fun looked at that trade-off and chose to de-weight the behavior that was cheapest to monetize and most expensive to tolerate. That is a real cost, accepted deliberately. A platform that sacrifices near-term volume for long-term retention is a platform optimizing for survival, not for the next quarter. I did not expect it, and I should say so.

Second, the low-cap de-weighting is a precise intervention. It targets the exact condition — thin liquidity — under which manipulation is cheapest. A blunt fix would have cut all caller rewards. The platform cut the subset of rewards that subsidized the highest-harm behavior. That is surgical, and surgery implies a model of the harm. The platform knows more about its own abuse patterns than any outside analyst, and the shape of the fix reveals the shape of the problem. If you want to know what a platform's worst behavior is, read what it de-weights.

Third, the mechanism, if implemented with a credible attribution engine, would create the first real skin-in-the-game for callers in the meme vertical. The prior regime paid for attention. The new regime, done properly, pays for outcomes. A caller with genuine selection skill is rewarded. A caller who is pure noise is not. That is a selection pressure the venue did not previously have. It is the difference between a marketing channel and a talent filter, and the difference compounds over time. A venue that filters for skill attracts a different class of participant than a venue that pays for reach.

Fourth, and this is the point the bears miss entirely: the platform is doing this while it is winning. Platforms usually reform after they lose. Pump.fun is reforming from a position of strength, which means the reform is not desperation. It is a bet that the meme vertical's long-term value accrues to the venue that can be trusted, not the venue that is most permissive. That bet may be wrong. But it is a sophisticated bet, and it is the opposite of the short-termism the bears attribute to it. The bears are reading a defensive move as an admission. The bulls are reading it as a positioning move. Both readings are available from the same three facts, which is itself a signal that the disclosure is too thin to settle the question.

Fifth, and this is the strongest bull argument: a platform that can change its incentive parameters quickly can also correct them quickly. The same centralization that creates rule-durability risk creates rule-adaptability benefit. If the October 10 settlement reveals that the mechanism suppresses caller activity without improving follower outcomes, the platform can revert or adjust within days. A decentralized venue would be stuck with its mistake. The centralization that I flagged as a risk is also the platform's fastest error-correction tool. Whether that net-positive or net-negative depends on whether the platform uses it, which is a question about the team, not the mechanism. I have no data on the team's follow-through, and I will not pretend to.

Takeaway

The question is not whether the mechanism is well-intentioned. The question is whether the platform will publish the function that computes follower returns — the attribution rule, the cost basis, the time window, the anti-wash defenses — before the first settlement on October 10. If it does, the mechanism can be audited and trusted to the extent the audit holds. If it does not, the platform is asking callers and followers to trust a payout they cannot verify, and the ledger that would prove the claim is the one ledger no one can read. The block chain remembers what humans forget. The trades are on-chain. The returns are on-chain. The only thing missing is the formula that turns one into the other. Publish it, or the silence becomes the finding.