Last Wednesday, NPR published internal Meta documents revealing the design of a standalone prediction market app codenamed Arena (also "Antwerp" and "FBForecast"), being developed under Mark Zuckerberg's direction. The story broke alongside CNBC, the New York Times, and a wave of coverage from public radio stations, tech publications, and AI policy outlets. The conversation has mostly focused on whether Meta will succeed commercially, what the competitive threat means for DraftKings and FanDuel, and whether 56% of Forrester respondents not trusting Meta with prediction markets matters.

Those are real questions, but they aren't the most important one.

The most important question is buried in the technical specifications. Meta's Arena will use Llama — the company's large language model — to do three things: generate the questions automatically from trending topics, recommend markets to users via personalization, and resolve the outcomes in near-real-time, with no human review and no challenge window.

That last design choice is not a minor technical detail. It is the largest structural innovation in prediction market design in 2026, and it forces every serious participant in the category to answer a question that has been carefully avoided so far: what does trustworthy resolution actually require, and how does institutional capital make decisions about that?

The institutional adoption arc this series has documented — Goldman's analytical engagement, the operational infrastructure buildout, the Evercore framework, the prop firm wave — has been operating on top of resolution infrastructure that everyone implicitly trusts but few have analyzed structurally. Meta's Arena, by introducing a categorically different resolution model, makes the underlying question unavoidable.

This piece walks through what Meta is actually building, why it's structurally different from every existing major platform, and what the implications are for the next phase of institutional adoption.

What Meta Is Actually Building

The technical specifications from the internal documents NPR obtained are precise enough to analyze structurally. Arena, when it launches, will operate as follows:

Llama generates the markets. The LLM scans trending topics across Meta's properties and the broader internet, then automatically generates yes-or-no questions to wager on. This is a meaningful design choice: it means the question itself isn't human-curated, isn't subject to operator judgment, isn't constrained by what a compliance team thinks is appropriate. Llama decides what gets a market.

Llama recommends markets to users. Each user gets personalized market suggestions based on what Llama infers about their interests, demographic patterns, and engagement profile. This is standard Meta personalization, applied to financial-adjacent products.

Llama resolves the outcomes. When the underlying event happens (or doesn't), Llama scans available information and renders a binary verdict on whether the resolution criterion was satisfied. The verdict happens in "near-real-time" per the internal documents.

No human review. No designated team validates Llama's resolution. No moderator examines disputed calls. No editor reviews the questions Llama generated for ambiguity or manipulation potential.

No challenge window. Once Llama resolves a market, the resolution is final. There is no period during which users can flag a wrong call. There is no economic mechanism (bonded disputes, voter consensus, designated escalation) for contesting the AI's judgment.

Play-money launch initially. The documents indicate Arena will launch as play-money, which sidesteps Commodity Futures Trading Commission jurisdiction for now. Any future transition to real-money wagering would land the app inside the regulatory battle that has generated more than 30 lawsuits in 2026 alone.

Meta declined to comment on the documents NPR obtained. The internal materials cite "the operational cost of manual question curation" as the reason Meta's earlier prediction market app (Forecast, 2020-2022) was shut down. Arena, in other words, is built specifically to solve the cost problem that killed Meta's first attempt — by replacing the human work with AI.

That cost framing is the part most coverage has missed. AI resolution isn't primarily a feature. It's a cost solution to the failure mode that previously made Meta's prediction markets unsustainable. Whether it works depends not on whether it's accurate (Llama may well be accurate most of the time) but on whether the cost it introduces in return — the replacement of human-reviewed resolution with probabilistic AI verdicts — is acceptable to users, regulators, and the participants the platform is trying to attract.

Why This Is Structurally Different

Every major prediction market platform operating today uses some form of human-overseen resolution. The implementations vary substantially, but they share a structural commitment: at some point in the resolution chain, a human is in the loop, dispute is possible, and the resolver has economic skin in the game.

The contrast becomes visible when you lay out the major platforms side by side.

The Resolution Architecture Comparison

The major prediction market platforms — and what Meta's Arena is doing differently.

Read across the rows and the pattern is clear. Every existing major platform has built some combination of human review, dispute mechanism, multi-source verification, and bonded participation. The implementations differ — Polymarket's UMA system is cryptographically novel; Kalshi's internal team approach is regulatory-driven; HIP-4's single-oracle design is engineering-elegant — but they all share the structural commitment that resolution requires at least some of those properties to be trustworthy.

Meta's Arena has none of them.

This isn't a criticism of Meta in particular. Arena is, structurally, the cleanest possible application of "use AI to replace expensive human infrastructure" applied to prediction market resolution. From a pure cost-engineering standpoint, the design is internally consistent: AI generates the markets at no marginal cost, AI personalizes the recommendations at no marginal cost, AI resolves the outcomes at no marginal cost. The entire pipeline is automated.

The question is what that automation costs in terms of resolution trust. And that's the question the rest of the ecosystem now has to answer.

The Structural Problems With AI-Only Resolution

Six structural problems are worth working through carefully, because they affect not just Meta's Arena but any platform considering AI-driven resolution as a cost-reduction strategy.

Hallucination risk. Large language models are known to confabulate facts with high confidence. A confident wrong answer from Llama is more dangerous than an uncertain right answer from a human reviewer, because there's no signal that the model is wrong. Llama doesn't tell users "I'm not sure about this resolution"; it renders a verdict. The verdict feels authoritative even when it's wrong. In a financial product, an unrecognized error is the most damaging kind.

Manipulation incentive transformation. In a human-resolved market, the manipulation target is the market price itself. In an AI-resolved market, the manipulation target shifts to the information environment Llama scans to make its resolution judgment. Bad actors don't just trade against market prices anymore; they flood news sources, social media, and Meta's own platforms with content designed to steer Llama's verdict. The attack surface multiplies. Boston University political scientist Matthew Motta, quoted in Tech Times coverage, named this directly: candidates could "use prediction markets not just to make a quick buck but to potentially manipulate the outcome of elections" — and with AI resolution, the manipulation lever is broader and cheaper.

Resolution opacity. How does Llama arrive at "yes" or "no"? The reasoning trace, if it exists, isn't visible to users. There's no audit log, no source citation, no chain-of-reasoning transparency mandated by the design. Even if Meta publishes the resolution code, the actual weights determining the verdict are inside the model. Users have to accept the resolution because Llama said so. That's a different kind of trust than what existing platforms require — and it's a kind of trust that institutional capital, in particular, is historically reluctant to extend.

Single point of failure. All resolution flows through one model, controlled by one company. If Llama is misaligned, biased, hacked, or simply wrong about a class of events, every market on the platform inherits the same failure. Compare this to Polymarket's UMA voter consensus (distributed across many voters with token stakes), or even Kalshi's internal team (multiple human reviewers, escalation paths). Distributed resolution mechanisms are more robust because failures are isolated. Centralized AI resolution concentrates failure risk.

Question generation bias. Less discussed but structurally important: Llama doesn't just resolve markets, it generates them. What gets a market? What questions does the model find "interesting" enough to generate? Whose worldview, whose news sources, whose engagement patterns shape the market-generation algorithm? Every prediction market platform has implicit editorial choices in which questions exist. With Llama, those choices are encoded in the model rather than made by editors — making them harder to audit, harder to challenge, and harder to evolve.

The dispute window absence. Even setting aside everything above, the design specifies "near-real-time" resolution with no human challenge mechanism. That removes the structural feature that has historically protected users when resolution mechanisms make mistakes. A wrong call is final. There's no analog to UMA's bonded disputes, no analog to Kalshi's escalation process, no analog to even the basic customer service review most retail platforms maintain. Once Llama decides, the decision stands.

Each of these problems is real. Together they constitute a categorically different resolution architecture from anything else in the market.

What This Means For The Institutional Adoption Arc

The institutional adoption story this series has been documenting reached its fourth chapter when proprietary trading firms started entering prediction markets at 13% active and 31% considering. The Acuiti / SGX survey we covered in Intelligence Brief #17 identified the structural barriers: regulatory uncertainty cited by 57%, CFTC clarity identified by 56% as the most important catalyst.

But the prop firm wave, and the broader institutional adoption it represents, has been operating on top of resolution infrastructure that everyone implicitly trusts. The question of resolution trustworthiness hasn't been the binding constraint — partly because the major platforms have built credible human-overseen resolution, partly because the most institutional volume has gone to markets (macro, elections, regulatory decisions) where the resolution sources are independently verifiable.

Meta's Arena makes the resolution question impossible to avoid, for three reasons.

First, it introduces a new entrant with massive distribution. More than 3 billion people use at least one Meta app daily. If Arena succeeds in attracting any meaningful share of those users to a prediction market product, it will inject substantial retail volume into the category — at a structural moment when institutional capital is just beginning to figure out what infrastructure it trusts. Retail volume on an AI-resolved platform creates pricing data that diverges from institutional volume on human-resolved platforms. The signal value of prediction market prices, which institutional readers have been increasingly trusting, becomes harder to interpret when different platforms resolve differently.

Second, it sets a precedent other platforms will be tempted to follow. AI resolution is dramatically cheaper than human-overseen resolution. The cost gap is large enough that any platform competing on margin will at least consider whether to adopt similar approaches. If Arena succeeds commercially, the pressure on Kalshi, Polymarket, and others to integrate more AI into their own resolution pipelines will intensify. The question is whether the entire category drifts toward AI resolution as a cost-reduction race, or whether the institutional-grade segment commits to maintaining the human-in-the-loop standard.

Third, it forces a public conversation about what resolution actually requires. The implicit trust the institutional adoption arc has been operating on becomes explicit when Meta enters with a categorically different design. Goldman's analysts citing prediction market probabilities will now have to account for resolution architecture differences across platforms. The Evercore framework's "clear resolution rules" criterion takes on a new dimension: clear by whose standard, verified by what mechanism, disputable through what process?

The institutional adoption arc, in other words, has been working with an implicit assumption that resolution infrastructure was a solved problem. Meta's Arena demonstrates that it isn't, and forces every serious participant to make their resolution architecture commitments explicit.

What Trustworthy Resolution Actually Requires

The framework that emerges from looking at existing platforms and the structural problems with Meta's approach has six requirements. These aren't novel — they describe what existing major platforms already do, plus what would be needed for AI-augmented resolution to maintain trust at scale.

Reproducibility. Multiple independent parties should be able to verify a resolution outcome from public information. If only the resolver can see how the resolution was determined, the system isn't reproducible. The major existing platforms achieve reproducibility through designated sources (Kalshi: BLS data, certified results) or through observable on-chain verification (Polymarket: UMA voter outcomes are public).

Auditability. The reasoning behind a resolution should be inspectable. Why did the resolver conclude yes rather than no? What sources were consulted? What edge cases were considered? Human-overseen resolution provides this implicitly — humans can be asked to explain their reasoning. AI resolution requires explicit logging, source attribution, and reasoning traces that don't currently exist in standard LLM deployment.

Manipulation resistance. The resolver should be structurally hard to game. UMA achieves this through economic stake (manipulating voter outcomes requires staking against the consensus and losing tokens). Kalshi achieves it through designated authoritative sources (you can't easily manipulate the BLS). An LLM scanning the public information environment is structurally easier to manipulate than these alternatives, because the information environment is itself manipulable.

Multi-source verification. No single point of failure. Even excellent human teams make mistakes; even excellent models hallucinate. Distributed resolution mechanisms (voter consensus, multiple designated sources, cross-platform validation) absorb individual failures. Single-resolver architectures inherit every failure.

Bonded participation. Resolvers should have economic skin in the game. UMA voters lose stake if they vote against consensus. Kalshi is regulated as a CFTC-designated exchange and bears regulatory liability for resolution errors. Meta's Arena would have no comparable mechanism — Llama doesn't bear cost for a wrong call.

Time-windowed disputes. A defined period during which resolutions can be challenged. This is the structural feature that protects users when the resolver makes a mistake. Every major existing platform has some version of this. Arena's "near-real-time" design with no dispute window removes it entirely.

These six requirements form a kind of de facto institutional standard. Not because anyone has formally specified them, but because every credible existing platform implements some combination of them — and the ones that don't (rare and historically short-lived) have struggled to attract institutional capital.

Meta's Arena, as designed, satisfies none of the six. That's a categorically different position from any other major platform.

Why The Resolution Question Is The Foundational One

Step back from the specifics of Arena's design and the deeper observation is that resolution infrastructure is the foundational layer the entire prediction market category sits on. If resolution isn't trustworthy, none of the rest matters — not the analytical engagement, not the institutional plumbing, not the prop firm wave, not the data infrastructure that's still being built.

The institutional adoption arc has been documenting the buildout on top of resolution infrastructure. Goldman pricing off prediction markets assumes the markets resolve correctly. Galaxy's OTC desk facilitating block trades assumes the contracts resolve correctly. Wintermute making two-sided markets assumes the underlying contracts have credible resolution. Kalshi's $17 billion May volume rests on the platform's resolution credibility. The entire structure is load-bearing on the resolution layer at the bottom.

Meta's Arena, by making the resolution layer the design centerpiece in a new and structurally different way, forces a conversation the rest of the ecosystem has been able to defer.

The conversation isn't whether AI plays a role in prediction market resolution. AI will play a substantial role — for cost reasons, for scale reasons, for the kinds of high-volume low-complexity markets where human review is uneconomical. The conversation is what role, and under what constraints, and with what fallback mechanisms when the AI is wrong.

Arena's design represents one position in that conversation: AI does everything, no fallback, near-real-time, no human in the loop. That's a clear position. It's also a position that places enormous trust in one company's model, and that institutional capital has no historical reason to extend.

The other positions in the conversation haven't been mapped yet. AI-assisted human resolution (Llama generates a recommended verdict; humans review). AI-flagged disputable resolution (Llama resolves quickly, but flags markets where confidence is low for human review). AI under multi-model consensus (multiple LLMs from different providers must agree). AI with bonded human oversight (Llama proposes; bonded human resolvers can override). Each of these designs preserves cost reduction while addressing specific structural problems with pure AI resolution.

What's clear is that the resolution architecture conversation is now the most important design question in the prediction market category. Not because of anything Meta has shipped — Arena hasn't launched yet — but because Arena's design choice has made it impossible to avoid.

Every serious platform now has to make an explicit commitment about resolution architecture. Every serious participant — institutional capital, regulators, researchers, journalists — now has to evaluate platforms on the basis of that commitment. The implicit trust the institutional adoption arc has been operating on becomes an explicit design decision that platforms publish, defend, and compete on.

That's a structurally healthy development for the category. The conversation needed to happen. Meta's Arena forced it.

The Bigger Picture

The most important design question in any new asset class is rarely the one that gets the most coverage. Coverage tends to focus on commercial dynamics, competitive positioning, regulatory battles, and feature comparisons. The foundational questions — what makes the asset class trustworthy enough to absorb institutional capital — get less attention until something forces them into view.

Meta's Arena has just forced the foundational question into view for prediction markets. The question is what resolution architecture the category commits to as it absorbs the volume the prop firm wave brings, as the regulatory landscape clarifies, as institutional capital scales further into the asset class.

The institutional adoption story we've been documenting — Goldman, Galaxy, Polymarket block trades, Wintermute, Kalshi at $17B, the Evercore framework, the prop firm wave, the data infrastructure being built — all rests on the assumption that resolution works. Meta's Arena tests that assumption by introducing a categorically different model. The result is going to be a conversation the category needed to have, about a question the category has been able to defer.

What gets built next on resolution infrastructure is the most consequential development to watch. The platforms that commit publicly to specific resolution architectures, that defend those architectures analytically, and that build the audit and dispute infrastructure required for institutional trust — those are the platforms that will absorb the institutional adoption arc as it continues to compound.

The platforms that don't will be left competing for retail attention on a substrate that institutional capital won't sit on.

Resolution is the foundational layer. Meta just made it the design question of the year.

This is the eighteenth installment in the Assymetrix Intelligence Brief series.

Previous: "The Prop Firm Wave: Prediction Markets Just Got Their Trading Floor."

Related: "Evercore ISI Has a Framework for When to Trust Prediction Markets. Here's What It's Missing." — the analytical framework whose "clear resolution rules" criterion takes on new significance in the context of AI resolution. "Polymarket Is Right About Footprints. The Question Is Who Else Can Read Them." — predecessor on the structural integrity question that resolution architecture directly affects.

Assymetrix is building the cross-venue, on-chain intelligence layer that turns public ledgers into readable, structured market data — independent of any single platform.

assymetrix.com/blog