Two pools, one number
A $500M pool held together by a thirty-day campaign and a $500M pool that grew on its own read identically to every contract on chain. That was survivable while humans allocated.
Total value locked is the dominant summary statistic in decentralised finance. It is also close to uninformative about the property most allocators actually care about, which is whether the liquidity will still be there when they need to exit through it.
Consider two pools, both reporting $500M.
| Pool A | Pool B | |
|---|---|---|
| How it got there | Deposits accumulated over eighteen months, no external incentive | $17M → $500M in five weeks under a campaign paying $29,000 a day |
| Median depositor tenure | Nine months | Nineteen days |
| Largest holder | 4% of supply | Top three hold 61% |
| Campaign expiry | None | Eleven days |
| Reported TVL | $500M | $500M |
A lending market pricing collateral, a vault choosing an allocation, a router estimating slippage — each one reads $500M and stops. A human analyst can tell these apart in ten minutes with a browser. An autonomous agent cannot, because the distinguishing information is not in any state it reads.
This is not a hypothetical asymmetry
During the first six weeks of one recently launched Layer 2, nineteen concurrent incentive campaigns paid a combined $58,351 per day, of which two campaigns accounted for 92%. The largest single campaign, at $29,514 per day, drove one synthetic dollar from roughly $17M to roughly $253M of supply — approximately 43% of all stablecoins on the chain — displacing the previously dominant issuer.
92%
of daily emissions came from two of nineteen campaigns
43%
of the chain's stablecoins, from one campaign
−72.5%
DEX volume from peak over the same period
Over that same period the chain’s DEX volume fell 72.5% from peak and its share of global DEX volume fell from 16.65% to 4.32% before partially recovering. The reported TVL never showed the fragility. It showed a record.
Why it becomes urgent now
Three developments turn a reporting deficiency into an infrastructure requirement.
Allocation is being automated.Vault strategies, yield routers and agent frameworks now execute allocation decisions without a human in the loop on each trade. These systems consume structured state. Context that lives in a governance forum, a Twitter thread, or an analyst’s judgement is invisible to them.
Incentive programmes are the default growth mechanism. New chains and protocols routinely bootstrap with emissions, gas subsidies and points programmes. That is a reasonable strategy. It also means a large fraction of visible liquidity at any moment carries an undisclosed expiry date.
Composability propagates the error. When a lending market accepts an LP position as collateral without pricing its exit liquidity, the mispricing does not stay local. Every position built on top inherits it.
What a fix has to look like
The missing quantity is duration, and it has to arrive in a form a contract can branch on. Cleaton publishes it as a single call:
getAttestation(pool) → (horizon, confidence, expiry, signature)Concretely: safe for 9 days, confidence 0.7. The claim is falsifiable, time-bounded, cryptographically attributable and economically backed. It is not a rating, and it is not advice — it is a statement about how long the depth is expected to last, published in advance and checked in public afterwards.
The part that cannot be forked
The scoring model is publishable, and we publish it. Any competent team can reimplement the survival ensemble from the paper. What cannot be reimplemented is a public record of predictions checked against outcomes over time — including the misses.
That record is the asset, and it only exists if it accumulates from the first attestation. Which is why the ordering matters: attestations first, over REST, with the reliability diagram published beside them from day one. Contracts come when someone asks to consume them on chain.