Over the past fourteen days, derivative contracts have been trading in the United States whose settlement source is a human mouth.
The contract pays if a named person says a named word inside a defined window. It pays zero if the word is not said. No index grades the outcome. No third-party data feed publishes it. No exchange official sits in the room. The oracle is a speaker, and the speaker is simultaneously the party with the strongest economic interest in knowing whether anyone is positioned against them.
That is the instrument. Now the rule.
The Commodity Futures Trading Commission's Division of Market Oversight has issued a staff advisory on contracts of this type. The industry shorthand is mention markets. The operative language is not a prohibition. It is a burden shift. Staff presume these contracts are easy to manipulate, and an exchange that wants to list one must rebut that presumption with a heightened showing. Three predicates sit inside that showing: whether legal, professional, or fiduciary duties suppress the manipulation, whether external pressure on the subject is feasible, and whether the outcome is independently verifiable and of significant public interest. The letter carries the signature of Duncan Hennes, Acting Director of the Division of Market Oversight. It states plainly that it creates no new obligations and represents the views of staff rather than the Commission itself.
Two enforcement settlements precede it. In the first, a respondent booked roughly $107,500 in unlawful gains and paid $172,000 in total, of which $65,000 was a discounted fine, plus a three-year industry ban. In the second, a respondent paid $35,000.
This is not a war on technology. It is a plumbing question, and plumbing questions are the only ones worth auditing. We mapped the water, not the wave.
The system, as it stands, has a settlement layer with no independent verifier, a manipulation surface measured in syllables, and a compliance standard written for an operator who may not exist in the venue where these contracts are actually growing. That is the subject of this report.
Context: What a Mention Market Is, and Why It Is Not a Niche Curiosity
Start with the legal container, because the container determines everything downstream.
Event contracts are derivatives whose payoff depends on whether a specified event occurs. They have existed in regulated form for decades under a narrow set of permitted categories. The modern venue for them is the Designated Contract Market, or DCM โ a CFTC-registered exchange with a rulebook, a compliance function, surveillance obligations, and the legal capacity to list contracts for trading. A DCM is not a protocol. It is a legal person that can be sued, fined, and shut down.
The conventional resolution sources inside that container are deliberately boring. An election result certified by a state authority. A consumer price index published by a statistical agency under an embargo protocol. A game outcome recorded by a league with an integrity unit. An official unemployment print. Each of these shares one property that is rarely stated explicitly and is absolutely load-bearing: no single participant can will the outcome into existence. A candidate cannot decide to win. A statistician cannot decide the print, at least not alone and not without leaving a document trail that other statisticians will read. The settlement source sits outside the reach of any one actor's hand.
Mention markets break that property.
The category is defined by a question of the form: will person X say word Y, appear at event Z, or appear in a photograph with person W, within a stated window. The resolution source is not an institution. It is an individual's behavior, and that behavior is under that individual's direct voluntary control. A podcaster can satisfy the condition in a single sentence. A guest can satisfy it by showing up. A public figure can satisfy it by shaking a hand.
The gap is not theoretical, and the enforcement record makes that concrete. In one matter tied to this space, a respondent was found to have generated roughly $107,500 in unlawful gains and ultimately paid $172,000, of which the monetary penalty component was discounted to $65,000, alongside a three-year bar from the industry. In a second matter, a respondent paid $35,000. Two settlements are not a trend. Two settlements are, however, an established cost baseline, and a cost baseline is more useful than a principle.
The advisory also lands inside a live competitive dispute. Traditional derivatives incumbents have signaled interest in the prediction market category, and the reporting around the advisory explicitly frames the question of whether Washington is quietly choosing winners between an incumbent exchange with a century of clearing relationships and a newer, litigation-tested venue that built its franchise on event contracts. That framing matters because it tells you what the letter is actually allocating. It is allocating listing rights. Listing rights are a form of capital.
And there is a third venue class the letter does not resolve, only shadows. On-chain prediction markets settle through an optimistic oracle with a dispute window and a bond, not through a compliance department. They cannot, at the protocol layer, identify an insider or cap a position in the way a DCM can. When a regulator publishes a mitigation menu built entirely from identity and position limits, the menu itself is a statement about which architectures can comply and which cannot.
One more framing note, because it will matter later. Media coverage filed this story under crypto. The instrument is not a token. It is a cash-settled derivative in a CFTC-registered venue. The classification is a function of the outlet, not the object. But the second-order effect is real: on-chain venues that list semantically identical contracts now operate in the reputational shadow of a rule they cannot satisfy structurally. That asymmetry is where the interesting risk lives.
Core: The Resolution Source as a Trust Boundary
A taxonomy of controllability
The advisory is usually read as a warning about a weird product. The more useful reading is that it formalizes something the derivatives industry has never had to write down: a scale of resolution-source controllability.
Consider the range. At one end, a national election. No person can unilaterally produce the outcome. At the next notch, an official statistic. Publication is procedurally protected: embargoes, independent review, multiple signatories, and career consequences for deviation. Next, a corporate earnings figure. The issuer controls the number, but the number passes through an auditor, a filing regime, and securities law exposure. Next, a sports result. Players can affect it, which is precisely why leagues maintain integrity units, betting monitoring, and lifetime bans โ the manipulation surface exists and is policed rather than absent.
Then the scale leaves the institutional region entirely. Whether a specific individual says a specific word is not a low point on the controllability scale. It is the terminal point. The subject needs no accomplices, no leverage, and no luck. The subject needs a sentence.
The correct abstraction is this: a resolution source is a trust boundary, and mention markets put the boundary inside a human being who has an economic interest in the outcome. That is a different class of problem from an oracle that can be bribed or a feed that can be delayed. Those are failures of infrastructure. This is a failure of incentive geometry, and it exists before any code runs.
I spent late 2017 doing static analysis across more than 150 ERC-20 contracts from the ICO cycle, and I found twelve critical flaws, most of them arithmetic โ integer overflows in transfer logic, unchecked return values, approval races. Those were failures inside the container. The container could be fixed. What the CFTC is describing is different in kind. The contract logic here can be flawless and the settlement can still be wrong, because the event the contract references is authored by a person who knows what the contract needs.
Why the standard integrity stack does not reach it
The existing surveillance architecture for event contracts rests on a small number of mechanisms that the industry treats as sufficient. Large position reporting. Block trade review. Cross-market correlation analysis. Post-settlement review of unusual activity. Restricted lists for employees of the exchange.
Each of these fails against a mention market, and the mode of failure is worth being precise about, because imprecision is how compliance theater gets funded.
Large position reporting detects concentration. It answers the question: does anyone hold a size that would make manipulation profitable? But it operates on positions held, not on the party who can trigger the event. A speaker need not hold a single contract to move the settlement. In fact the cleanest manipulation is for the speaker to hold nothing at all and simply say the word, while a related account on the other side collects. Reporting thresholds are calibrated to find the whale. Here the whale is the microphone.
Cross-market correlation analysis detects statistical footprints after the fact. It is a forensic tool. It reconstructs intent from timing. It does not stop a sentence, and it does not prevent the price from being wrong for the entire window during which the contract trades.
Restricted lists for exchange employees are the closest existing analog to what the advisory actually demands, and they are the tell. Restricted lists work because the population is enumerable: employees, directors, and their households are known to the compliance function. The mention-market problem is that the relevant population is not enumerable through employment. It is the speaker's staff, the speaker's family, the producer who writes the script, the teleprompter operator, the person who books the guests, the agent, the publicist, the sound engineer who hears the rehearsal.
A script is a document. A document is material non-public information. An unknown number of people hold it before the market moves. There is no employment relationship that makes them visible to the exchange, and there is no registration that makes them visible to the regulator.
The three predicates, decomposed
The heightened showing is not rhetorical. It decomposes into three testable predicates, and each one has a failure mode the exchange must argue against.
Predicate one: are there legal, professional, or fiduciary duties that suppress the manipulation? This is a question about the subject's external constraints. A public company officer faces securities law. A lawyer faces bar rules. A trustee faces fiduciary exposure. A licensed broadcaster may face employment terms. The predicate is answerable only when the subject sits inside a regulated role with a credible enforcement threat behind it. For a private podcaster with an audience and no professional licensure, the honest answer is no.
Predicate two: can external pressure be brought to bear on the subject? The advisory's own example is the strongest part of the letter, because it is not hypothetical. A listener can offer a donation conditional on the word. A guest can steer an interview with a question engineered to elicit it. A caller can bait a host into repeating a phrase. In each case the pressure is legal, public, and effective. The exchange must therefore show that the subject cannot be induced. That is a hard showing for any format built on audience interaction, which is to say, for most of the formats these contracts are written on.
Predicate three: is the outcome independently verifiable and of significant public interest? This predicate is the most operationally useful because it is the one an exchange can actually engineer. A nationally televised event with an archived transcript and broad coverage is verifiable. A low-traffic stream with no archive and no audience is not. Public interest is a proxy for verification density: the more people watching, the more people who will contradict a false settlement, and the more expensive a dispute becomes.
Read together, the three predicates define a compliance test whose stringency is a direct function of how controllable the resolution source is. That is a coherent framework. It is also an admission that for some contract designs, the honest output of the test is a refusal.
The mitigation math: when manipulation stops paying
The advisory names two mitigations: identify potential controllers and known insiders, and set position limits so that manipulation becomes economically irrational.
The second one is quantifiable, and the arithmetic is unforgiving. Define the manipulator's position as N contracts paying one unit each, acquired at price p. Define the personal cost of triggering the event as c, which includes effort, reputational exposure, and legal risk. Manipulation is rational when:
N(1 โ p) > c
Therefore, to make manipulation irrational, the position limit L must satisfy:
L(1 โ p) โค c
Now examine what happens as the contract design moves toward the terminal point of the controllability scale. For a speaker repeating a scripted word, c approaches the cost of speaking. For a subject who says words professionally, c is close to zero on the reputational axis, because the word is contained inside a format where saying words is the job. As c approaches zero, the permissible position limit L approaches zero.
A position limit that approaches zero is not a control. It is a delisting with extra steps.
This is the algebra underneath the advisory, and it explains the letter's structure better than any commentary. The CFTC is not saying these contracts are bad. It is saying that the mechanism it has available โ position limits โ loses its power exactly where the product becomes interesting. The contracts people want to trade are the ones with the lowest c. The contracts the CFTC can safely allow are the ones with the highest c. The two sets barely overlap.
There is a second-order consequence worth flagging. Position limits suppress liquidity. If L is tight, market makers cannot size, spreads widen, and the contract becomes a curiosity traded by hobbyists rather than a price-discovery instrument used by institutions. So the framework does not merely narrow the product shelf. It also determines which contracts can ever develop a meaningful order book. Listing rights and liquidity are being decided in the same document.
The insider registry problem
The first mitigation โ identify potential controllers and known insiders โ sounds like a procedural step. It is actually the hardest engineering problem in the entire framework, and it is the point where the advisory implicitly separates venues by identity architecture.
To build an insider registry for a mention market, an exchange needs to answer a question that has no clean answer: who can influence whether this person says this word? The obvious entrants are the subject, the subject's household, and the subject's employer. The non-obvious entrants are everyone in the information path: the writer, the editor, the fact-checker, the producer, the booker, the agent, the publicist, the co-host, the guest, the audience member who submitted the question, and the venue staff who prepared the room.
That is a registry that cannot be closed. It can only be approximated, monitored, and periodically revised. Approximation is expensive. Monitoring is a perpetual operating cost, not a one-time build.
Here is where my 2025 work becomes directly relevant. I spent part of last year working with legal teams to build a compliance framework against new Canadian digital asset standards โ forty-five operational requirements mapped to existing securities precedents, with an eighteen-month transition window documented end to end. The single most durable finding from that exercise was a cost differential: firms that already had robust internal controls absorbed roughly 40 percent less compliance cost than firms building the same controls from scratch. The differential was not caused by the size of the rulebook. It was caused by whether the control surface already existed.
The same differential will apply here, and it will be decisive. An operator that already maintains a restricted-person process, a trade surveillance system with alerting, an incident response path, and a documented settlement-source verification policy will treat the heightened showing as an additional filing. An operator that does not will treat it as a capital project. In a bear market, that difference determines which venues keep listing and which venues quietly withdraw product.
The on-chain divergence: bonds are not position limits
Now the part the advisory does not address and cannot easily resolve.
On-chain prediction markets settle through optimistic oracles. A proposed resolution is posted with a bond. If no one disputes within the challenge window, it finalizes. If someone disputes, the question escalates to a vote or an arbitration layer, and the losing side forfeits the bond.
This is a genuinely different control mechanism from a position limit, and the difference is not cosmetic. A position limit constrains the size of the economic prize before the event. A bond constrains the cost of asserting a false settlement after the event. One is ex ante and quantity-based. The other is ex post and capital-based.
The equivalence condition is straightforward, and it is worth writing down because almost nobody does. The mechanism holds only when:
Bond > P(dispute succeeds) ร Profit from false resolution
When the bonded amount is small relative to a single actor's position in the contract, the oracle is not a security mechanism. It is a toll booth. The manipulator pays the bond and keeps the difference.
There is a further structural problem. A position limit is enforced by an entity that knows who holds what. Bonded dispute systems are enforced by anonymous capital. Anonymity is the feature that makes them permissionless, and it is precisely the property that makes the CFTC's insider-identification mitigation unimplementable at the protocol layer. Identity gating happens, if it happens at all, at the front end โ geofencing, wallet screening, terms of service โ and front-end gating is a legal posture, not a protocol property. Any user willing to interact with the contract directly bypasses it.
I will state my confidence level plainly: the advisory does not name on-chain venues, and my reading here is inference from the mitigation menu rather than a quotation. But the inference is tight. A regulatory framework whose only two instruments are identity and position limits is, by construction, a framework that applies to centralized venues and only hurts decentralized ones indirectly. That indirect pressure will show up as banking access, payment rails, exchange listings, and reputation โ not as an enforcement action against a contract.
Pricing enforcement: what Perez and Santos actually cost
The two settlements are usually reported as anecdotes. They are better read as prices.
Take the first matter. Unlawful gains of approximately $107,500. Total payment of $172,000. A monetary penalty component discounted to $65,000. A three-year industry bar.
The ratios matter more than the absolute numbers. Total payment to unlawful gains is roughly 1.6 to 1. The fine component is roughly 0.6 to 1 against gains, which is to say the penalty is real but not punitive in the way a treble-damages regime would be. The discount โ a fine below the headline exposure โ is the standard signal that cooperation was priced in. Regulators routinely trade penalty reduction for investigative assistance, and the trade is visible in the arithmetic.
Now propagate that price through a listing decision. An operator weighing whether to list a mention contract is evaluating an expected cost:
Expected cost = P(detection) ร Multiplier ร Gains
With a multiplier near 1.6, a detection probability of 0.3 produces an expected cost of roughly 0.48 times gains. That is not a deterrent at the contract level. It is a cost of doing business. The deterrent has to come from somewhere else โ from the compliance program's ability to raise detection probability, from the registry's ability to deter insiders before they trade, and from the reputational cost of being named in the next settlement.
The second matter, the $35,000 payment, cuts the other way. A small number is a cheap precedent. It tells operators that a first, contained violation is survivable. That is exactly the wrong signal if the goal is deterrence, and it is exactly the right signal if the goal is to establish jurisdiction without triggering a political fight.
The three-tier contract shelf the letter implies
No part of the advisory publishes a tiering. But the logic produces one, and any operator building a product roadmap should be reading it off the predicates rather than waiting for the rule.
Tier one: resolution sources no single party can control. Elections, official statistics, league-certified results. These sit where the framework assumes integrity already exists. Listing friction is low and expected to remain low.
Tier two: resolution sources a party can control but is constrained from controlling. Issuer guidance, award decisions by an organization with a conflict policy, decisions by a regulated professional acting inside a licensure regime. These require the heightened showing, but the showing is arguable because the predicates have real answers.
Tier three: single-actor utterance, attendance, and interaction. The subject is the resolution source. The predicates fail or come close to failing. These contracts are presumptively unlistable absent a control architecture that identifies an unenumerable insider population and imposes position limits near zero.
And then there is the shelf edge, which is where the real risk sits.
The blind spot nobody wrote down: low-salience words
A high-salience word โ a phrase everyone is waiting for, in a broadcast with millions of viewers and an immediate transcript โ is the safest mention contract on the shelf, and that is counterintuitive.
High salience raises verification density. Many people hear it, many people clip it, many people contradict a false settlement. It also raises the reputational cost of delivering the word on cue, because the delivery is conspicuous. High salience makes manipulation expensive and detection cheap.
Low salience inverts every term. A word nobody is tracking, said in a stream with a small audience, by a subject whose speech is not archived, has the lowest verification density and the lowest reputational cost. The manipulator does not need to be brazen. The manipulator needs to be boring.
The lowest-salience contracts are simultaneously the cheapest to manipulate and the least likely to be surveilled, which means the enforcement record and the risk distribution are pointing in opposite directions. The cases that get brought are the conspicuous ones. The cases that exist are the quiet ones.
This is the structural blind spot inside an otherwise coherent framework, and it is the reason I would expect the next enforcement action, if there is one, to involve a contract class that most of the market has never heard of.
The 2026 threat model: agents, not orators
The advisory is written for a human manipulator. That is the correct threat model for a letter drafted against a human-shaped problem. It is not the threat model that will be live eighteen months from now.
In 2026 I evaluated three AI-agent trading protocols operating against DeFi liquidity pools. Two of them were running latency arbitrage against human participants, and the mechanism was not exotic. The agents monitored the same public information humans monitored, parsed it faster, and routed orders ahead of the human reaction. The result was price discovery that looked functional and was actually a race with a predetermined winner.
Transfer that architecture to a speech-act contract and the failure mode changes shape entirely.
An agent with a speech-to-text pipeline and a direct order path can detect a spoken word and trade inside the latency of the human decision loop. That is not manipulation of the resolution source. The speaker said the word honestly. The agent simply extracted the entire informational content before any human could act, which means the market's price discovery stopped functioning as discovery and became a speed contest.
Now push it one step further, because this is the part that should worry a compliance officer. An agent can also be an influence instrument. The advisory's second predicate asks whether external pressure can be brought to bear on the subject. The advisory's examples are a donation and a planted question. An agent that operates audience-facing channels can generate pressure at scale and near zero marginal cost โ engineered questions, coordinated comments, repeated prompts, synthetic engagement designed to elicit a phrase. The subject still holds free will. The subject is also being optimized against by a machine with no reputational exposure and no attributable identity.
I have now audited enough of this intersection to be uncomfortable with it. My 2026 report was cited at an industry panel on ethical AI in finance, and the conclusion I stood behind then is the conclusion I stand behind here: technology that accelerates price discovery without improving the integrity of the settlement layer does not make markets better. It makes them faster at being wrong.
The mention market is the first derivative in regulated markets whose underlying is a human utterance, and it will not be the last. As AI-generated content becomes an underlying for contracts, the controllability spectrum acquires an entirely new terminal point: a subject that is machine-generated, infinitely producible, and directly owned by whoever holds the model. The framework in this advisory will be inherited by that problem, and it will need to be rebuilt.
What this does to venue economics in a bear market
The macro backdrop matters here, and it is not favorable. In a market where survival dominates return, the relevant question about any venue is not its growth narrative. It is whether its revenue base is narrowing while its fixed compliance cost is rising. That is the exact shape of this advisory's first-order effect.
A narrower product shelf means fewer contract classes. Fewer contract classes means less diversification across uncorrelated event flows, which raises revenue variance. Simultaneously, the heightened showing means a compliance function that must be staffed, documented, and audited on a schedule that does not care about volumes. For venues whose listing revenue is concentrated in a small number of contract families, this reduces to a single arithmetic problem: mandatory cost up, discretionary revenue down.
The venues with the strongest survivability characteristics in this environment are the ones that already own the control surface. That is the 40 percent differential again. The moat here is not technology. It is documentation.
For anyone holding exposure to prediction market venues, the diligence list is short and unglamorous. Does the venue publish a restricted-person policy with a defined population. Does it publish position limits by contract class, and are those limits explained in terms of manipulation economics rather than risk appetite. Is there a named compliance officer with a documented escalation path. Is the settlement-source verification policy public, and does it describe what happens when a settlement is disputed. Is there a disclosure for how the venue handles contracts whose resolution source is a single individual. If the answer to the last question is silence, that is an answer.
A note on legal weight, and why it matters less than it sounds
The advisory says it creates no new obligations and represents staff views rather than the Commission's. Both statements are true and both are strategically incomplete.
The mechanism the letter actually deploys is a presumption. Presumptions do not create obligations. They allocate the cost of proof. Before the letter, an exchange listing a mention contract had no established doctrine to rebut. After the letter, the exchange must affirmatively demonstrate that the contract is not easily manipulated. That is a real increase in cost, delivered without a rulemaking, a vote, or a Federal Register notice.
Staff advisories in this pattern historically precede formal rulemaking. The sequencing is familiar: staff guidance, then an enforcement pattern, then a notice of proposed rulemaking, then a final rule that codifies what practitioners had already been doing for a year. Anyone building on this category should model the current guidance as the shape of the eventual rule, discounted for uncertainty.
A ledger is a confession written in code. A staff advisory is a confession written in the conditional tense. It tells you what the Commission would do if it wanted to, and it leaves open whether it does.
Contrarian: The Letter Is Narrowly Bearish and Broadly Bullish, and the Consensus Has It Backwards
The consensus reading of this advisory is that it suppresses a category. I think that reading is backwards, and I think the error comes from treating mention markets as the subject rather than the probe.
Consider what the letter actually produces. It defines a trust boundary, publishes the test for crossing it, and prices the failure. Ambiguity is the most expensive input in an institutional allocation decision, and ambiguity in the settlement layer is worse than ambiguity in the fee schedule. Once a regulator states that contracts on the safe side of the controllability scale are listable and explains what safe means, the category stops being a novelty and starts being an asset class with an articulated risk floor. Institutional capital does not require low risk. It requires priced risk.
The narrow slice that loses is the terminal point of the controllability scale โ single-actor utterance contracts on low-salience content. Those contracts were never going to carry institutional flow. They were a retail curiosity with a structural manipulation surface. Losing them costs the category very little and buys it a legitimacy that no amount of growth could manufacture.
The second contrarian point is less comfortable, and it is the one I would put in front of a risk committee.
The word mention is not a stable category. It is a description of how a contract is phrased, and phrasing is a choice. Whether a central bank says a specific word about the path of rates is, in substance, a mention contract on a central banker, dressed as a monetary policy instrument. Whether a chief executive says a specific phrase about a divestiture is a mention contract on a chief executive, dressed as an equity-linked event. Whether a defendant says a specific phrase under oath is a mention contract on a witness.
The advisory draws the boundary where the resolution source is directly controllable by an individual. Drawn consistently, that boundary does not enclose a niche. It cuts across a meaningful share of corporate, political, and legal event contracts. The regulator has written a rule for the edge of the map, and the edge of the map is larger than the map itself.
We mapped the water, not the wave. The water is the twelve contracts in front of us. The wave is every contract whose settlement is decided by a sentence, and that set is growing faster than the framework that governs it.
Takeaway
The signal to track is not whether mention markets survive. It is whether the controllability boundary gets codified into a rule before the category outgrows the guidance.
Watch four things. A notice of proposed rulemaking, because that converts staff views into enforceable text. Commissioner-level statements, because they reveal whether the presumption survives a vote. Incumbent product filings in the prediction category, because the first mover inside a clarified boundary captures the moat. And the next enforcement action, which will tell you whether staff meant the low-salience blind spot or only the conspicuous cases.
The plumbing question does not change. When the settlement source is a person, the market integrity mechanism is a position limit, and the position limit that works is the one that makes the contract untradeable. The question is not whether the industry can build around that. It is whether anyone has noticed that the constraint is arithmetical rather than regulatory โ and that no venue has ever repealed an equation.
