MONARKFleetNarabiUkemiBellit abstains so DeFi can act.
Documentation menu
docs · the pieces · Ukemibuilt

Eligible is not liquidated.

Ukemi measures liquidation exposure on one lending venue: the lending book read at one declared block, the oracle price path the protocol actually consulted, and the difference between what a stress test flags and what gets liquidated. Once a stratum is committed, the gate hands back an upper bound on the amount liquidated, per stratum. Until then it abstains, and says so on the wire.

The question

A stress test on a lending book usually reports the debt that becomes eligible for liquidation when a price falls. That is not the amount that gets liquidated, and neither is a loss. Positions are liquidated in parts and re-buffer after each call; a deficit is left only when something breaks the price path.

Eligible is not liquidated: what a stress test flags, what gets liquidated and what is left as a deficit are three quantities.Three quantities a stress test tends to mergeEligible for liquidationevery position a price path pushes past its thresholdLiquidatedrepaid in parts: a position re-buffers after each callDeficitoften noneleft only when something breaks the price path: a jump no liquidator can step through,a lagging oracle, or a pool too thin to absorb the salethe gap Ukemi measures, per accountA single figure for the three would be a guess. Ukemi measures the first two on a recorded episode, account by account,and states the deficit apart.An illustration: the widths carry no value.
Three quantities a stress test tends to merge. An illustration: the widths carry no value.

What Ukemi reads

  1. Read the book at the reference block. Holders of the collateral with a debt, the health factor as the protocol computes it. Two distinct operators must agree, read by read.
  2. Record the oracle path. The feed updates over the window, from chain events. The scenario is declared on this path, not on an outside market price.
  3. Join the realized labels. Per account, what was actually repaid and what deficit was left. Eligible and not liquidated counts as nothing, and is kept.
  4. Calibrate, then emit. A split-conformal upper bound per stratum at the pre-registered level. Too few points for a population: an abstention marked under_calib; while the stratum is not committed the served answer counts no calibration point, and its measured count is on the course page. Missing input: an abstention, with the reason.

The prediction, the score, the bound

For each account, the prediction y-hat is the most the protocol’s own rule lets be liquidated in one call, taken at the first point where the recorded oracle path pushes the account past its threshold. The score is one-sided: how far the realized amount went above the prediction, and zero when it did not. So the margin measures the misses that cost, not the size of the prediction.

The prediction, the score and the upper bound: the region runs from an open floor to y-hat plus q-hat.One account in one stratumopen floory-hatupper boundq-hat, the margin of the stratuma realized amount under the boundone above it: a missy-hat: the most the protocol’s rule lets be liquidated in one call, at the first crossing on the recorded path.The score is how far the realized amount went above y-hat, zero when it did not; q-hat is a quantile of those scores.
The prediction, the score and the upper bound. The labels are those of the Ukemi page.

When a class holds enough calibration, the served region is a conformal upper bound on the amount liquidated: an open floor by construction, not a calibrated bound. How far the bound reaches says nothing about being right; a loose bound is a legitimate, informative answer.

Calibration, stratum by stratum

The accounts are cut into strata by the size of their prediction, with cuts fixed before the data. Each stratum gets its own split-conformal margin, and a stratum with too few calibration points abstains. One region per category of a taxonomy fixed in advance is a known construction:

The floor under which a population abstains follows the analysis of prediction sets for grouped data (Robin Dunn, Larry Wasserman and Aaditya Ramdas (2022)). The served gate states the terms of this class itself: a conformal upper bound on the liquidable amount for the calibrated class; the lower edge is 0 by construction, not a calibrated bound; abstains (under_calib) outside it; this class requires alpha = 0.01, nMin = 100; calibrated on one recorded episode; no coverage is claimed on any other event; the H-3 exchangeability check is a report, a YES licenses nothing more; the bound holds only if yhat was produced by the frozen close-factor rule on a mono-collateral WETH account at the first crossing, which the gate does not check. The score code was frozen, and the plan of the run written down, before any fresh data was read; what the run found is what the course page prints.

The course

The course ran the frozen method on one recorded episode, weth-2025-09-22, replayed offline from recorded reads. Each bar below is a stratum’s count of fresh calibration points; the tick is its floor. The counts and the outcomes are read from the committed course report.

The calibration course, stratum by stratum: fresh calibration points against the floor of each stratum.Fresh calibration points per stratum, against the floorAmounts are in the lending venue's oracle base currency (8 decimals); a stratum is a range of the predicted liquidable amount.stratum 0below 2000.00000000170exchangeability test with the design episode: yes (level 5/100)stratum 1at least 2000.00000000, below 100000.0000000021under_calib (n below the floor of 100)stratum 2at least 100000.00000000, below 1000000.0000000010under_calib (n below the floor of 100)stratum 3at least 1000000.000000004under_calib (n below the floor of 100)the floor: fewer points than this and the stratum abstains, under_calib
The calibration course, stratum by stratum, from the committed, hashed course report.

What a yes means: the pre-registered exchangeability test with the design episode (the recorded episode the score was designed on, never served) did not reject at its stated level. It is not a coverage claim, and the report estimates no coverage bound across episodes.

A comparison trial scored exactly zero (the realized amount did not exceed the prediction) counts as covered whatever the bound; the number of such trials is shown next to each ratio.

Stratum 0 is committed and served: the served verdict read at 2026-09-25T03:26:01.410Z carries 170 calibration points, fewer than the 199 an interior quantile rank needs at the served level, so the quantile rank equals the number of calibration points (170 of 170) and the served bound margin, 1261.84298996, is the largest calibration score observed, reported as is. The served upper bound for a prediction in this stratum is the prediction plus this margin; served calibration digest e7e673664c03e3c5d15956d864f8379b6fe4660ed689be38a85add95d4eff334.

Pooled class A: bound margin 1261.84298996 (quantile rank 204 of 205).

Every hypothesis of the course, the oracle path, the reconciliation of the positions and the digests are on the course page.

What is served today

Served state of the class liquidation-eligible-coverage, as copied from the served gate description (https://api.monarkgate.tech/openapi.json) at 2026-09-25T03:26:01.410Z and hashed in the site manifest; the live description may have changed since:

the served region is a conformal upper bound on the liquidable amount for the calibrated class; the lower edge is 0 by construction, not a calibrated bound; abstains (under_calib) outside it; this class requires alpha = 0.01, nMin = 100; calibrated on one recorded episode; no coverage is claimed on any other event; the H-3 exchangeability check is a report, a YES licenses nothing more; the bound holds only if yhat was produced by the frozen close-factor rule on a mono-collateral WETH account at the first crossing, which the gate does not check

A calibration of this class is committed; the served clause above states the bound it carries. The report below is the offline course, not the served registry.

The register’s served note for Ukemi: feeds the served gate through the cascade tool, a transitional tool, to be replaced; it abstains by construction, and the gate serves an upper bound on the liquidable amount for the committed stratum of its liquidation class, calibrated on one recorded episode, while the other strata abstain; both legs replayed by integration tests.

The engine underneath

Under the calibration, Ukemi’s engine computes the clearing fixed point of a network of obligations: each node pays what it can, in rounds, until the payments settle (Larry Eisenberg and Thomas H. Noe (2001)). With default costs the fixed point need not be unique (L. C. G. Rogers and L. A. M. Veraart (2013)), and conditions under which liquidation costs still leave a single equilibrium are known (Hamed Amini, Damir Filipovic and Andreea Minca (2016)). When forced sales do not feed back into the price the book is marked at, the fixed point is simply the amount liquidable at the block; whether such a feedback exists on a given market is an empirical question (Rodrigo Cifuentes, Gianluigi Ferrucci and Hyun Song Shin (2005) show how it arises when a market absorbs sales inelastically). Ukemi’s calibration measures the realized side of it: eligible against liquidated, account by account.

Limits, and what Ukemi is not

One venue, one collateral classOne lending venue, core market, single-collateral WETH accounts. An account holding any other collateral is excluded, marked non_evaluable. In the threshold recompute the other legs are held at their book-block price: a declared limitation, not a repricing.
Few episodesOnce a stratum is committed, its coverage holds only under exchangeability with the calibration episode. The distance to a new event is named, never estimated away.
Open questions stay openThe served-price delay seen in the design episode is unexplained; on the course episode the lag is measured against a pre-registered bound, not explained. An account in another efficiency-mode category is excluded, marked non_evaluable: whether that category covers the collateral cannot be checked off-line.
  • Not a forecast of the next price, and not a forecast that more liquidations will follow.
  • Not a probability of liquidation for an account, and not a probability that a bound is right.
  • Not a rating, a gauge or a ranking of a protocol, a market or an account.
  • Not a risk parameter: it sets no threshold and no cap. A curator stays the curator.
  • Not a claim about a new event: once a stratum is committed, its bound is calibrated on one episode; exchangeability across events is named, not assumed.

The Ukemi page states the method and the served state in full.

Sources

  • Daniel Gatto (2026) the title names the lending protocol studied, which the site describes generically. SSRN working paper 7157638. read in full Used for eligible debt is not executed volume, and executed volume is not protocol bad debt.
  • Larry Eisenberg and Thomas H. Noe (2001) Systemic Risk in Financial Systems. Management Science, vol. 47, no. 2, pp. 236-249. read in part Used for the clearing fixed point of a network of obligations.
  • L. C. G. Rogers and L. A. M. Veraart (2013) Failure and Rescue in an Interbank Network. Management Science, vol. 59, no. 4, pp. 882-898. read in part Used for default costs, under which the fixed point need not be unique.
  • Hamed Amini, Damir Filipovic and Andreea Minca (2016) Uniqueness of equilibrium in a payment system with liquidation costs. Operations Research Letters, vol. 44, no. 1, pp. 1-5. read in part Used for when liquidation costs still leave a single equilibrium.
  • Vladimir Vovk, David Lindsay, Ilia Nouretdinov and Alex Gammerman (2003) the title names the method with a word the site does not print; the paper is the working paper of the series below. On-line Compression Modelling Project (New Series), Working Paper 4. read in full Used for one region per category of a taxonomy fixed in advance, with validity per category and never per case.
  • Robin Dunn, Larry Wasserman and Aaditya Ramdas (2022) Distribution-Free Prediction Sets for Two-Layer Hierarchical Models. Journal of the American Statistical Association. read in part Used for the minimum number of calibration points below which a population abstains.
  • Anastasios N. Angelopoulos and Stephen Bates (2023) Conformal Prediction: A Gentle Introduction. Foundations and Trends in Machine Learning, vol. 16, no. 4, pp. 494-591. read in part Used for the split-conformal quantile and the meaning of marginal coverage.
  • Kaihua Qin, Liyi Zhou, Pablo Gamito, Philipp Jovanovic and Arthur Gervais (2021) An Empirical Study of DeFi Liquidations: Incentives, Risks, and Instabilities. ACM Internet Measurement Conference (IMC). read in part Used for the scale of liquidations and why an attested price matters.