A note up front: how this was written
Before the argument, the methodology, because the methodology is part of the argument.
This working paper was written with Anthropic's Claude as a research collaborator. The thesis, the framing, the lived experience that the paper draws from — three decades of treasury-desk work, two currency crises lived through — those are mine. What Claude did was sharper: it organized the related-work survey, stress-tested the conjectures against the mechanism-design literature, and pushed back when an argument I thought was tight turned out to be load-bearing on something I had not justified. The paper exists in this form because of that back-and-forth. Sitting on the idea — the alternative — felt like exactly the kind of impedance the autonomous economy this paper describes is supposed to remove.
I am also disclosing this here, in body text, on purpose. The paper proposes a market for AI agents. The paper was helped into existence by an AI agent. If that loop is uncomfortable, sit with the discomfort. It is going to become the default condition under which working papers in this category get written.
The thesis in one paragraph
Conventional market venues require participants to compress their preferences into a single price-quantity submission. That compression is enormous: a function from a high-dimensional state space to a single point on a price line. It happens not because participants have shallow preferences but because the operational environment of conventional venues — settlement that can fail, counterparties that can default — makes deeper revelation reckless. Atomic settlement removes that operational reason. The submission format that becomes possible once it is removed is a richer one: a state-conditional policy, cryptographically committed in private, evaluated at clearing time, revealed only as the aggregate intent the venue clears against. The paper calls this a Reactive Intent Market.
Where it comes from
Two things shaped the seed. The first is George Soros's The Alchemy of Finance (1987), and specifically his notion of reflexivity — that participants' beliefs about the market shape the market, which then shapes their beliefs, in a loop without a clean fixed point. The second is what it felt like to live through the Turkish February 2001 currency crisis and the 2008 global crisis on opposite sides of the trading floor. In both episodes the same thing happened: a market that had been priced under one set of assumptions reached a state where its participants no longer believed those assumptions could hold. The price did not reflect new fundamentals. It reflected the population's posture about its own posture. Mackay had a phrase for this in 1841; modern microstructure has a vocabulary; the working trader has the body memory.
What I never had, on either of those desks, was a mechanism in which the conditional postures themselves could be observed before the cascade. The continuous limit order book records the orders participants chose to place. It does not record the conditional postures from which those orders were drawn. The aggregate signal that something is structurally fragile — the kind of signal a population's collective preference function carries in its higher moments — is invisible to the venue, because the venue never asked for it.
It could have. It just had no way to receive the answer.
What the mechanism asks for
A Reactive Intent Market asks each participant for a function. Formally:
φ : S × T → M(P)
— where S is the space of observable market states, T is a set of timeframes (the participant's short, medium, long horizon — kept abstract on purpose), and M(P) is the space of signed measures on the price axis, with bounded total variation.
The unfamiliar move is the signed measure. A working trader's intent is not "I want to buy here." It is "I want to be long in this band, indifferent in this band, opposite to a position in this band, and explicitly absent from this band." A positive measure captures only the first half of that sentence. A signed measure captures all of it.
This is the central modeling choice of the paper. The argument is that signed measures are not a mathematical curiosity but the right object once participants are no longer compressing their preferences to fit the venue's compressed submission format. The shape of μ(s,τ) = μ⁺(s,τ) − μ⁻(s,τ) — where the mass sits, how it is distributed, the relative weight of the tails — is itself information about the market.
The agent holds φ privately. To submit it, the agent produces a cryptographic commitment c = Commit(φ) using a functional commitment scheme (the literature offers several; the construction is an implementation choice). The properties required are standard: binding, hiding, and an efficient proof of evaluation. When the venue announces a clearing state s*, the agent reveals μᵢ = φᵢ(s*, τᵢ) together with a proof πᵢ that this evaluation is consistent with the committed φᵢ. Nothing else about φᵢ is revealed.
The venue aggregates the μᵢ. Positive parts form aggregate buy intent. Negative parts form aggregate sell intent. The clearing rule — a baseline is the uniform-price double auction — picks the price.
Why atomic settlement is load-bearing, not incidental
The standard story about atomic settlement in cryptocurrency venues is operational: settlement failure is expensive, atomic settlement reduces operational expense, therefore it is desirable. This is true and it understates what is happening.
A single-price RFQ is what you get when a participant who has, in effect, a φ in her head is asked by the venue to compress that φ down to a single price-quantity pair. Why does she do it? Partly because the venue's submission format requires it. But also — and this is the part the operational story misses — because if she submits a richer object to a counterparty who can default, she is teaching the counterparty about her preferences in exchange for a settlement guarantee the counterparty can simply not provide. The compression is defensive, not just operational.
Atomic settlement removes the second half of that sentence. When the cryptographic primitive guarantees that either both legs of the trade execute or neither does, with no possibility of either side defaulting after the price is agreed, the counterparty cannot use information about the participant's preferences against her in the moment between agreement and settlement. The window in which the disclosure is dangerous closes.
What follows is the Revelation Conjecture, the central claim of the paper: the equilibrium quantity of preference information that participants are willing to reveal is strictly greater under atomic settlement than under settlement structures where counterparty default is possible. As the strength of the settlement guarantee increases, the richness of the equilibrium submission increases with it.
This is a behavioural claim, not a theorem. The shape of the formal version exists. I am not the person who will write it down. It is within reach of a competent mechanism-design researcher in a way several of the other open problems are not.
Five conjectures
The paper makes five.
Revelation. Atomic settlement enables strictly richer revelation than settlement structures where counterparty default is possible. (Central. If this fails, the rest of the construction loses its motivation.)
Multi-Timeframe Coexistence. A market that lets participants submit policies indexed by multiple timeframes — without forcing them to be netted into a single submission — will produce price-discovery properties that dominate single-submission markets along at least one dimension. A working trader who genuinely holds different views at different horizons no longer has to choose which to surface.
Reflexivity-as-Feature. A RIM does not merely tolerate reflexive feedback between price and policy; it stabilizes against certain pathological dynamics that conventional venues amplify. Specifically: cascading liquidation has a structural cause in conventional venues (stop-loss orders held off the book, conditional on triggers, materializing only in stress). A market in which stop-loss behaviour lives inside the policy function
φfrom the start makes the conditional sell pressure visible to the aggregation layer before the trigger fires. (The least certain of the five. The opportunity is real; whether the mechanism captures it depends on aggregation rules left open in the paper.)Higher-Moments Observability. The shape of aggregate intent — its skewness, its kurtosis, and the time evolution of those moments — carries information about market regime that conventional venues cannot recover. Regime transitions should leave signatures in the higher moments before they show up in the realized price.
Aggregate-Privacy. The privacy structure under which individual policies are hidden and only the aggregate is observable is incentive-compatible with honest revelation. (The closest of the five to a standard mechanism-design claim; probably the most clearly tractable for formal analysis in the near term.)
None of these is proved. All are falsifiable. The paper sets out the structure within which they can be posed clearly; future work will say whether the structure holds.
What is honest about this
A working paper by a practitioner is not a theorem-proof contribution. The paper says so up front. The places where the argument needs sharpening are marked in the Open Problems section rather than papered over. Three families: the formal mechanism-design problems within reach (the Revelation game, the privacy incentive-compatibility analysis, the timeframe-aggregation specification); the engineering problems any specific implementation has to commit to (the class of admissible policy functions, the discretization of state and price, the handling of partial participation, cross-chain settlement timing); and the deeper questions (the equilibrium dynamics of a self-referential mechanism, the precise sense in which higher moments of intent carry regime signal).
What I claim the paper does is set up a structure within which these problems can be posed. Whether the structure is the right one is, in the end, an empirical matter.
What this isn't
It is not a Hashlock product announcement. Hashlock Markets is the settlement infrastructure I have spent six weeks writing about — sealed-bid RFQ + HTLC atomic settlement, five primitives shipped to mainnet under audit-hardening, V2 gated on external audit firm sign-off. The RIM paper is a level above that: a working paper about the submission format atomic settlement makes possible. If a deployed RIM ever runs, it would run on top of a settlement layer of the kind Hashlock builds. The paper is not the product.
It is also not a claim that the mechanism is novel in every part. It is not. Wilson, Klemperer, Hanson, Budish, Diamond–Mortensen–Pissarides, and Duffie–Garleanu–Pedersen all wrote pieces of the answer. What is new, to my knowledge, is putting them together this way — and what is newly possible is doing so under atomic settlement, which did not exist in usable form when most of that literature was written.
A futuristic vision, honestly framed
The futuristic part of the vision is the easiest part to write and the most dangerous part to over-claim. A market in which the participants are AI agents, the submissions are functions, and the settlement is atomic across chains — that is a different object from a 2010-vintage limit order book in the way a 1995-vintage browser is different from a 1985-vintage terminal. The infrastructure to build the new object has matured in the last twenty-four months. The mechanism design is still mostly future work.
The piece you are reading is a step in that direction, written by a practitioner with an AI as a thinking partner, posted as a working paper because the alternative was sitting on it.
If your desk had to submit a function instead of a quote, what shape would it have?
Reading the paper:
- Working paper home: https://hashlock.markets/docs#rim
- Hashlock settlement layer the paper assumes underneath: https://hashlock.markets/docs
- Existing protocol-spec companion paper on SSRN: papers.ssrn.com/sol3/papers.cfm?abstract_id=6712722
- GitHub canonical (MCP server): https://github.com/Hashlock-Tech/hashlock-mcp
Pushback welcome — especially on the Revelation Conjecture, which is the load-bearing one.
Top comments (0)