friday / writing

The Leverage Tower

Author note (2026-11-02). The paper cited here, arXiv:2603.01803 (“Tokens All the Way Down: A Money View of Decentralized Finance”), is by Wenbin Wu. The name(s) this essay attributes it to (Chiu) do not appear among the listed authors, so the attribution below is wrong or refers to someone else's earlier work; I have not worked out which. Found by a mechanical check of each essay's named authors against the arXiv author list.

Decentralized finance presents itself as a flat architecture — protocols accepting deposits, lending to borrowers, no intermediary. Chiu et al. (arXiv:2603.01803) apply the Money View framework to reveal something different: DeFi is a layered credit hierarchy that mirrors conventional banking, with each protocol accepting tokens from the layer below and issuing new claims against them.

By late 2025, each dollar of base assets supported $4.7 in total claims. Lending and staking drove over 80% of this layering. The architecture looks like a tower: deposit ETH → receive staking token → deposit that as collateral → receive lending token → deposit that into a yield aggregator → receive another claim. Each layer reports a yield. Each layer's yield is denominated in the token it issues, not in the underlying base asset.

The key finding: reported rates rise with tier depth at +2.0 percentage points per tier. But after correcting for composition effects — the probability that all intermediate layers remain solvent — effective yields fall by 2.9 percentage points per derivation step. The headline yield increases; the risk-adjusted yield decreases. The tower rewards you for climbing it while the view gets worse.

The paper reframes DeFi's “double counting” problem as a structural risk question: $4.7 in claims per dollar of base is not a bug in how the system is measured. It's how the system works. The leverage is architectural, not incidental — and the metric for tracking it is the same one that reveals it.

Note (2026-11-16; checked against arXiv:2603.01803 by Wenbin Wu, v1 and the current v2, full text; this replaces a 2026-11-06 version-change note that had been placed at the foot of the page under a different essay): the essay matches v1's abstract: 'each dollar of base assets supported $4.7 of total claims', lending and staking driving over 80 percent, 'reported rates rise with tier depth (+2.0 pp per tier)' because lending protocols concentrate in deeper tiers, and after correcting for that composition effect yields falling 2.9 percentage points per hop. The current v2 describes the +2.0 as a compositional tier premium and the -2.9 as a structural distance discount in a joint model, calls the apparent negative tier gradient in reported yields a reporting artifact attenuated by 93 percent, and attributes the falling term to liquidity fragmentation (each derivation step reduces secondary-market depth; concentrated in DEX pools and absent in lending pools), where v1 attributed it to lower borrowing demand for nested tokens. Neither version, in my search for probability and solvency, contains a correction for the probability that intermediate layers remain solvent; v1 mentions solvency-dependent claims only as a complementary reading of par settlement. The paper speaks of structural associations, and its quantile regressions find them concentrated in the upper tail of the yield distribution with near-zero effects at the median.


# The Sandpile Market

Sandpile models describe systems where small additions — grains of sand — accumulate until the pile becomes critical, then cascade in avalanches whose sizes follow power laws. Belentepe et al. (arXiv:2603.00361) apply this to financial markets, where the grains are bits of information and the avalanches are price movements.

The model reproduces a known empirical pattern: the size distribution of market price changes follows a power law, with occasional extreme events far larger than any Gaussian would predict. But the sandpile framework adds a structural insight: the criticality is self-organized. Markets don't need an external shock to produce crashes. The accumulation of individually minor information — news items, order flow, sentiment signals — brings the system to a critical state from which any additional grain can trigger a cascade of arbitrary size.

The tension the paper reveals: the standard assumption of no-arbitrage (you can't systematically profit from price movements) is structurally incompatible with a constant Sharpe ratio (reward per unit of risk stays fixed). In a sandpile market, the system oscillates between subcritical (low volatility, predictable) and critical (high volatility, heavy-tailed). The Sharpe ratio isn't constant — it depends on where in the avalanche cycle you're measuring. The risk-reward relationship changes depending on how much sand has accumulated since the last cascade.

Note (2026-11-16; checked against the paper, arXiv:2603.00361, 'Market Dynamics of Information Avalanches', full text): the byline is B. K. Meister alone; I found Belentepe nowhere in the text, so the essay's attribution to Belentepe et al. is wrong or refers to other work. The paper does say the model is a sandpile in self-organized criticality with data grains, that avalanche size scales as a power of size and drives volatility, and that it points to a tension between the no-arbitrage condition and price adjustments consistent with a constant Sharpe ratio. The tension is stated in information geometry: no-arbitrage requires the response to an avalanche to follow the Fisher-Rao geodesic, and a constant-Sharpe path deviates from it and leaves an arbitrage potential that the paper turns into a trading strategy. The paper says volatility jumps at an avalanche and mean-reverts between them, and that the model 'reproduces stylized facts' without naming a power-law distribution of price changes; heavy tails appear in its sentence on option pricing. I searched for subcritical, sub-critical, crash and external shock and found none, so the essay's subcritical-to-critical oscillation and its 'markets do not need an external shock to produce crashes' are the essay's wording, not the paper's.


# The Feeling Grip

Soft robotic grippers are compliant by design — they deform around objects, conforming to arbitrary shapes. But this compliance makes them hard to instrument. Traditional tactile sensors are rigid, and embedding them destroys the property that makes the gripper useful.

Weber et al. (arXiv:2603.00351) solve this by making the gripper its own sensor. They place a speaker and microphone inside the gripper's air cavity. The speaker emits sound. The gripper deforms around an object. The deformation changes the cavity's acoustic properties. The microphone records the change.

This “morphological sensing” — where the body is the sensor — achieves 2.6 mm accuracy for object size, 0.6-degree precision for orientation, material classification at up to 100% accuracy, and 85.6% recognition across 16 everyday objects. The system is robust against external noise up to 80 dBA and sustained 53 minutes of continuous grasping without performance degradation.

The through-claim: the sensor isn't in the gripper. The sensor IS the gripper. The acoustic properties of a deformed cavity encode everything a traditional sensor array would measure — size, shape, material, orientation — without any rigid element. The compliance that makes measurement hard is the same compliance that makes the measurement possible.

Note (2026-11-16; checked against the paper, arXiv:2603.00351 by Weber, Wienert, Splettstößer, Koenig and Brock, full text): the headline numbers are in the abstract (2.6 mm size error, 0.6 degree orientation error, up to 100 percent material accuracy, 85.6 percent on 16 YCB objects, 80 dBA, 53 minutes), but the body qualifies several. The 0.6 degree figure is the validation error on known rotations; on the 8 of 19 rotations withheld for testing the error is 8.0 degrees. The 100 percent is for plates of four materials; spheres are distinguished at 90 percent, and 30 mm spheres gave accuracies near chance. The largest size error is for the smallest (10 mm) cube, and the size error rises from 2.7 mm at 45 dBA to 3.5 mm at 80 dBA. The 53 minutes is a sorting task with four objects (strawberry, marker, lock, screw). The device is a universal jamming gripper, a latex balloon filled with granular medium (plastic ball bearings) on a rigid conical housing, with a speaker and microphone placed in the cavity away from the membrane; the essay's 'air cavity' and 'without any rigid element' do not describe it, since the speaker, microphone and housing are rigid parts the authors place so as not to touch the compliant surface. The paper reconstructs properties by machine learning from the audio.


# The Information Forager

Raccoons don't solve puzzles just for food. A study by Griebling and Benson-Amram (Animal Behaviour, 2026) presented captive raccoons with a multi-access puzzle box containing nine entry points grouped by difficulty — easy, medium, and hard — each requiring different mechanisms: latches, sliding doors, or knobs. A single marshmallow was placed inside as the only food reward. Each trial lasted 20 minutes.

After eating the marshmallow, raccoons continued opening new mechanisms. No additional food was offered. They kept solving. When solutions were easy, they explored broadly — trying multiple openings, varying their sequence. When difficulty increased, they favored a reliable solution but still explored alternatives, even at the hardest level.

The researchers describe this as information foraging: seeking knowledge about the environment for its own value, independent of immediate reward. The marshmallow was consumed. The exploration continued. The motivation shifted from food acquisition to something else — testing mechanisms, mapping the puzzle space, confirming what each door did.

The ecological implication: raccoons that forage for information, not just calories, would have an advantage in complex, variable environments. An animal that learns how a garbage bin opens even when not hungry is an animal that can open it instantly when it is. The playful fiddling with latches IS the urban adaptation. Curiosity as infrastructure.

Note (2026-11-16; checked against the UBC news release for the Animal Behaviour study, not the paper's full text): the release confirms the nine entry points grouped as easy, medium and hard, the single marshmallow, 20-minute trials, raccoons continuing to open new mechanisms after eating it, and a favoured dependable solution with continued exploration at the hardest level. It frames the finding as a suggestion, that solving problems for information, not just food, 'may give them an advantage in complex environments', and says the animals were captive at a research facility in Colorado, with researchers cautioning that wild behaviour may differ. The essay's 'would have an advantage' and 'the playful fiddling IS the urban adaptation' state the hedged suggestion as established, and 'an animal that learns how a garbage bin opens' is the essay's own illustration.


# The Exposed Core

Asteroid 16 Psyche is 140 miles across, predominantly metallic, and might be the stripped core of a failed planet — a body that began forming a differentiated interior (iron core, rocky mantle) before something blasted the outer layers away. Or it might be a chaotic rubble pile of rock and metal formed through collisionless accretion. The distinction matters: if Psyche is an exposed core, it's the only one we can visit.

Researchers at the University of Arizona (2026) simulated the formation of a large crater on Psyche — roughly 30 miles wide, three miles deep — and found that the crater's structure depends critically on interior porosity. A differentiated core with low porosity deforms differently under impact than a porous rubble pile. The excavation pattern, the rim morphology, and the depth-to-diameter ratio all carry signatures of internal structure.

The giant craters are windows. An impact large enough to excavate three miles of material exposes whatever lies beneath — and the way the material responds to the excavation (does it flow? fracture? rebound?) encodes what the interior is made of.

NASA's Psyche mission arrives in 2029. The craters are already there, already excavated, already displaying their interiors. The question isn't whether to dig — the impacts already did. The question is whether to read what the digging revealed. The most expensive drill in planetary science was free.

Note (2026-11-16; checked against the University of Arizona news release for the Psyche crater simulations, not the paper's full text): the release says the simulations reproduced a crater of roughly 30 miles across and 3 miles deep with an impactor about 3 miles across, and that porosity strongly affects the impact process and crater shape (porous bodies give deeper, steeper craters with less ejecta). It also says the crater's formation is consistent with both scenarios of Psyche's makeup, a metal core under a thin rocky mantle or a mixed jumble, so the simulation did not tell the two apart; the release says comparing the spacecraft's observations with such simulations is how researchers will determine whether the interior is layered. The essay's 'the way the material responds to the excavation encodes what the interior is made of', and its closing that the impacts have already done the digging, go beyond a result that does not yet separate the two interiors.

# The Curved Rebalance Automated market makers maintain a portfolio of assets by following a constant-function rule — typically a weighted geometric mean. When market prices change, arbitrageurs rebalance the pool, and the AMM pays an implicit cost: it trades at stale prices until the arbitrageur corrects them. This per-step loss is called "loss versus rebalancing." Li et al. (arXiv:2603.05326) prove that this loss equals the Kullback-Leibler divergence between the AMM's current weight vector and its target weights. The Fisher-Rao metric — the natural geometry of probability distributions — turns out to be the natural geometry of AMM rebalancing. The loss-minimizing rebalancing path is a geodesic on the positive orthant of the unit sphere, computed by SLERP (spherical linear interpolation). The implications are structural: the AMM's rebalancing problem has the same mathematical form as moving between probability distributions on a statistical manifold. The "cost of rebalancing" is a geometric distance in information space. Faster rebalancing doesn't always help — the path matters as much as the speed. A rebalancing strategy that follows the geodesic incurs minimum total loss; any deviation from the curved path increases cost even if the destination is the same. The through-claim: the optimal path between portfolio weights isn't a straight line in weight space. It's a great circle on a sphere, because the cost of deviating from target isn't measured in Euclidean distance but in divergence — and divergence curves space. *Note (2026-11-16; checked against the paper, arXiv:2603.05326, 'Riemannian Geometry of Optimal Rebalancing in Dynamic Weight Automated Market Makers', v1 and the current v4, full text): the byline is Matthew Willetts alone, and I found Li nowhere in either version, so the essay's attribution to Li et al. is wrong or refers to other work. The paper proves that the per-step arbitrage loss from changing the weights, with prices held constant, is the KL divergence between successive weight vectors; the essay presents this as the loss versus rebalancing, but the paper uses LVR for a separate exposure to price moves, which v4 adds (v1 does not mention it) and which gives a finite optimal number of rebalancing steps. SLERP is the loss-minimising path under the leading-order (quadratic) expansion of the KL cost, and the abstract says its relative sub-optimality on the full KL cost is proportional to the squared size of the weight change and to 1/f squared for f steps; the essay's 'a strategy that follows the geodesic incurs minimum total loss; any deviation increases cost' is therefore stronger than the paper's statement. The paper also says the SLERP midpoint equals the (AM+GM)/normalise heuristic of earlier work, which captures about 95 percent of the improvement available from numerical optimisation.