On day 41 after the final reward epoch closed, the pool still held $4.1 million. Forty-one days earlier, it held $12.8 million. Every dashboard that tracks this sort of thing had already written the obituary — minus 68 percent, a corpse, another incentive program that rented a number and then lost the lease.
I pulled the depositor set instead of the dollar figure. Distinct addresses that had ever touched the pool: down 9 percent.
That gap — 68 percent of the money, 9 percent of the people — is the entire subsidy era compressed into two integers. The liquidity did not leave because the users left. The liquidity left because a small number of very large wallets, which were never users in any meaningful sense, executed a scheduled withdrawal. The community — the sticky part, the part that actually pays fees — barely moved at all.
I have spent the last four months pulling apart this exact question across 47 pools on two Layer 2s and one Layer 1. The finding that has survived every robustness check I could throw at it is uncomfortable for everyone involved: most DeFi incentive programs are not acquiring users, they are acquiring float, and float has a price the protocol is not measuring. Silence in the code speaks louder than the hype. The emission schedule ends, and the audit trail tells you what was ever really there.
Liquidity mining is a subsidy with a laundering step. A protocol wants depth in a pool; depth is a public good that every arbitrageur and every leveraged borrower free-rides on; so the protocol pays for it directly, minting a token and distributing it pro-rata to whoever parks capital in a contract. In 2020 this was a genuine discovery. Compound's launch that June turned a subsidy into a growth engine, and the industry spent the next four years industrializing the copy. The mechanism has not changed since. What changed is the marginal return on a subsidized dollar — and the bear market is the thing that made that marginal return visible to anyone who bothered to look at event logs instead of dashboards.
The accounting is where the story starts. Total value locked measures assets inside a contract at a point in time. It does not measure commitment, duration, or ownership. An address that deposits on epoch-hour and withdraws on epoch-hour plus ninety is counted identically to a treasury that deposited in 2021 and has never touched a governance vote. If your only metric is a snapshot, those two actors are indistinguishable — and the entire incentive-design industry has been optimizing against a snapshot.
My dataset is 47 pools across three chains, fourteen months of block-level data pulled from archive nodes through a Python pipeline I wrote and then rewrote twice. I index Mint, Burn, Transfer, and the pool-specific Sync and Swap events, and I reconstruct each depositor's full position history rather than sampling weekly snapshots. For entity clustering I use a fairly standard stack: common-input ownership signatures where they exist, gas-price and nonce fingerprints, funding-path tracing through two hops, and withdrawal-address re-use across pools. None of the heuristics are novel. Applied together, they are enough to break the naive assumption that one address equals one user — an assumption I stopped trusting in 2021, when I traced 100 Bored Ape wallets and found fifteen percent of the alleged community resolved into a single operator's cluster of addresses.
Caveats first, because this is where most on-chain claims die. Clustering produces false positives, especially for custodian-generated deposit addresses. I cannot see positions that never touch the pool contract — a wrapper, a vault, a centralized exchange's omnibus wallet. I cannot see intent at all; every behavioral label in this piece is inference from timing and size, not testimony. And the window is the window. Fourteen months in a market that has already changed regime twice is a sample, not a law.
Chaos is just data waiting for a lens. What the lens showed me was a two-population market hiding in plain sight.
The elasticity of a rented dollar.
I built one number per pool: emission elasticity, defined as the percentage change in TVL associated with a ten percent change in the pool's quoted annualized reward rate, measured with a two-week lag so capital has time to actually move. This is a crude regression, and I am the first to say so — it conflates reward-rate changes with the token price moves that usually cause them. But the distribution is not ambiguous. Across the 47 pools, median elasticity came in at 0.71. The interquartile range ran from 0.38 to 1.44. The top decile sat above 2.1, meaning a ten percent cut in quoted yield preceded a better-than-twenty-percent decline in deposits within a fortnight.
Elasticity is a price. If deposits respond that violently to a yield quote, then those deposits are not customers, they are suppliers of a commodity — blockspace-adjacent, perfectly substitutable, priced to the basis point. The pools that behaved differently were the interesting ones. Where the underlying pair generated real fee revenue and the reward rate was a minority of total yield, elasticity clustered near 0.3, and in three cases the sign flipped: deposits grew during a reward cut, presumably because the cut signaled a longer runway. That is the signature of capital evaluating a business rather than a spreadsheet cell.
The behavioral fingerprints matched the arithmetic with almost comical precision. Median time from the first emission of a new epoch to the first deposit from the high-elasticity cohort: 6.4 hours. Median time from a program-end announcement to withdrawal initiation: 31 hours. Median time from actual program end to final withdrawal: 4.2 days. This is not a group of people making decisions about protocols. It is a group of contracts responding to a parameter change, and the parameter is the only thing they read.
The ghost hands of the depositor set.
Here is the part that should worry every governance forum. In the highest-elasticity decile, 41 percent of apparently unique depositors collapsed into nine entities under clustering. Nine. In one pool, the reported depositor count was 2,140 addresses; the clustered count was closer to 380 entities, and the top nine controlled 71 percent of peak TVL.
This is a different ghost than the one I chased in 2021. The BAYC clusters were about hiding ownership to manufacture the appearance of distribution. These clusters are about routing — the same operator deploying capital across three chains, farming identical epoch structures, and diversifying chain risk while concentrating protocol risk. The wallets are not pretending to be independent people. They are just being counted as if they were. We trace the ghost in the machine's memory, and the memory says one desk, nine vaults, three chains.
The practical consequence is that depositor-count metrics are close to worthless as a health indicator in any subsidized pool. A program can announce 2,000 new depositors and have acquired, functionally, a handful of renters. Worse, those renters now have governance surface, because most programs weight voting by deposit size at the time of snapshot, and nine coordinated entities with 71 percent of a pool's liquidity is nine signatures away from controlling a quorum.
Cost per sticky dollar.
I computed the number nobody publishes: total emission value distributed over a program's life, divided by TVL retained sixty days after the program ended. Median across the sample was $0.47 of emissions to retain $1 of post-program TVL. The dispersion is enormous — the best quartile retained a dollar for under $0.12, the worst quartile paid more than $0.80 and still watched the majority of it drain within the following quarter.
Two things hide inside that figure. First, emissions are paid in a volatile asset, so the cost is quoted at the price on the distribution date, not the price the recipient realized. Given that the high-elasticity cohort's median holding period for a reward token in my sample was under nine days, the realized discount was brutal, and the sell pressure was concentrated in exactly the window when the protocol most needed price stability to look credible. Second, the retained dollar is not necessarily a good dollar. Retention measured sixty days out includes positions that are merely slow to unwind.
I have watched this same shape before. In 2022 I spent three weeks documenting reserve volatility in an algorithmic stablecoin while the market treated the metric as noise, and the death spiral arrived on schedule. The lesson was not that I was right. The lesson was that degradation is legible in the data long before it is legible in the price — you just have to be measuring the thing that degrades rather than the thing that gets quoted.
The negative-margin tail on Layer 2.
Incentive programs on rollups deserve their own paragraph, because the economics are worse there and almost nobody accounts for it.
A Layer 2 operator has two cost centers that scale with usage: data availability and proving. Proving is the expensive one for validity rollups, and it has an ugly property — it does not scale down gracefully with transaction value. A transaction that moves $12 costs the operator nearly as much to prove as one that moves $12 million, and the fixed cost per batch amortizes only if batches stay full. In a bear market, blockspace demand has thinned enough that batch utilization has fallen across the rollups I track, which means the effective per-transaction proving cost has risen even as fee revenue fell.
Now add incentives. Reward-claim transactions, deposit-and-withdraw loops, and the near-zero-value transfers that farming infrastructure generates are the cheapest possible transactions to produce and among the most expensive to prove relative to the value they carry. When an L2 subsidizes deposits to grow its activity metrics, it is paying emissions to generate transactions that consume its scarcest cost center at a negative margin, then reporting the resulting activity number as traction. The subsidy has a tail that lands on the operator's proving budget, and that budget does not care about your TVL chart.
Unless gas returns to bull-market levels — and I see no structural reason it should in the near term — this is a bleeding wound dressed as growth. The rollups that survive the next twelve months will be the ones that stop paying for transactions and start pricing them.
Recursive TVL.
One more measurement artifact, because it distorts everything above. LP tokens are collateral. In 11 of the 47 pools, a meaningful share of reported TVL consisted of LP positions re-deposited into lending markets and then counted again at the pool level. My conservative double-count ratio for those pools ranged from 1.3x to 1.9x. That number inflates the denominator of every cost-per-dollar calculation, flatters retention, and makes a pool look deeper exactly when leverage is doing the work. When the market moves against the collateral, the unwind is not a withdrawal — it is a liquidation, and it arrives all at once.
Here is where I have to argue against my own conclusion, because the data does not license the story I just told.
Renting liquidity is a legitimate strategy. If depth enables fee revenue that exceeds the rental cost, then paying for float is simply buying an input at a market price, and the alternative — no depth, worse execution, no routing, no fee income — is strictly worse. The comparison is not sticky capital versus mercenary capital. The comparison is rented depth versus no depth, and my dataset says nothing about the counterfactual. I can measure what the subsidy cost. I cannot measure what it bought in routing volume, integrator attention, or the optionality of being the venue people tried first.
Second, the "sticky" cohort may be sticky for reasons that are not virtue. A position can be retained because it is underwater, because exit costs exceed the remaining value, because the depositor is inattentive, or because the LP token is locked in a vesting contract. That is not loyalty. That is a hostage. If I re-ran the retention metric using only positions that were in profit at the sixty-day mark, I suspect the sticky cohort would shrink substantially — and I have not run it, which is a gap in this analysis I am flagging rather than hiding.
Third, and most importantly: correlation is not causation, and elasticity is not proof. Both the reward rate and the deposits may be responding to a third variable — the trajectory of the token's price, which drives both the quoted yield and the willingness to hold. My two-week lag is a gesture toward identification, not a solution. It is entirely possible that what I am calling emission elasticity is mostly price momentum wearing a costume.
And there is a version of this where the mercenary cohort is the sophisticated one and the sticky cohort is simply slow. The renters read the schedule, priced it, executed, and left. The loyalists stayed through a drawdown and are now down forty percent on a position they never modeled. Which population would you rather have acquired?
The honest answer is that I do not know. What I do know is that the industry has been reporting one number and making decisions about a different one.
Watch three signals this week. First, emission elasticity disclosed alongside the reward rate — if a team will not publish what happens to deposits when they cut, they already know. Second, cluster-adjusted depositor counts, because nine entities is not a community and the governance implications arrive before the market implications. Third, the cost-per-retained-dollar figure, which converts a marketing decision into a P&L line where it belongs.
The ledger remembers what the market forgets, and what it currently remembers is a very large number of wallets that arrived within six hours of a parameter change and left within five days of its reversal. The question worth sitting with is not whether your protocol's liquidity is real. It is whether, when the emissions end, anyone is still there who would have come without them.
