What Poker Solver Math Taught Us About Pricing Real Estate

In 2019, an AI named Pluribus did something that had eluded computer scientists for decades. It beat the world's best professional poker players at six-player no-limit Texas Hold'em, the most popular form of the game humans play. The breakthrough was not raw computing power. Pluribus computed its blueprint strategy in eight days using just 12,400 core hours, a tiny budget by modern AI standards. The breakthrough was a way of reasoning about decisions where the information is incomplete: weighting evidence, modeling ranges of possibilities, and pricing the unknowns directly into the strategy.

It is also exactly what real estate pricing has been missing.

How AVMs got stuck

For two decades, the most powerful idea in real estate technology has been the automated valuation model, or AVM. Zillow's Zestimate, Redfin's Estimate, and the bank-side AVMs that originate trillions of dollars in mortgages every year all share a common premise: feed in beds, baths, square footage, lot size, location, and recent comparable sales, and a model returns a number.

It works, up to a point. Zillow currently reports a median error rate of 1.74% on on-market homes and 7.20% on off-market homes across the United States. That gap is the most interesting number in residential real estate. The on-market accuracy is striking because the model gets to use the listing price as a signal. The off-market number, which is what someone considering refinancing or a portfolio acquisition is actually relying on, sits four times higher. Why? Because off-market homes are not labeled with the one dimension that most clearly differentiates one property from a comparable one: what has actually been done to it, and whether anyone can prove it.

That dimension is what poker solvers would call the hidden information of the game. And handling hidden information well is what separates a strong pricing model from a weak one.

What solver math actually does

Here is the simplified version of what Pluribus and its predecessors broke through. In poker, you cannot see your opponents' cards. So you cannot know what the right move is in any single hand. What you can do is compute, over millions of simulations, the strategy that maximizes expected value given a probability distribution across what your opponents might be holding. That distribution, in poker theory, is called a "range." Solver-grade play does not chase certainty. It weights every action by the evidence it has and the cost of being wrong.

Pricing a home involves the same structure. Two homes on the same block, both listed at the same price per square foot, are not the same asset. One has a documented HVAC replacement, a verified roof inspection, and a clean record of permitted renovations. The other has a Google Drive link the seller has been threatening to organize for three years. A solver-grade pricing model would weight the documented home meaningfully higher, not because the documents themselves are valuable, but because the buyer's downside risk is meaningfully lower.

AVMs do not currently do this. They cannot, because most of the information that would drive the differentiation has never been verified or captured in machine-readable form. The model treats the two homes as fungible inputs because, on paper, that is what they look like.

The market is already pricing what AVMs cannot see

You can watch the market quietly do this math on its own. According to Realtor.com's 2025 Consumer Attitudes and Usage Study, the single most common regret among recent homebuyers (16%) was unexpected maintenance after closing. That regret is the precise gap a verified property record closes. Buyers know they are paying a premium for the unknown, and they are starting to look for ways to take that risk off the table.

We have seen this in our own customer conversations. One recent seller listed his home with his REL Property Guidebook attached and went under contract on Day 3 with multiple competing offers and full appraisal gap coverage. The most telling line was not from him. It came from the buyer's agent who told him afterward: "If I had a dollar for every client who asked for a CARFAX-like report for a home, I'd be rich." That quote is what a market consensus sounds like before the pricing models catch up. Buyers were already paying for the verification dimension informally because the formal models could not see it.

A solver-grade AVM, the one that comes next, will weight this information natively. It will treat a property with a fingerprinted, tamper-evident maintenance record as fundamentally less risky than a property without one, in the same way Pluribus treats a known board card differently from a random one. The early movers, builders documenting at the source and homeowners maintaining a verifiable record over time, are the ones whose properties will price up first when those models arrive.

Where this goes next

The math behind Pluribus did not stay in poker. The same family of algorithms, descended from the poker AI research lineage that platforms like OpenPoker are now operationalizing, is reshaping decision-making in any domain where outcomes depend on imperfect information. AI products like TheUpsider need verifiable inputs to make those decisions trustable. The substrate that lets verification scale across consumer and institutional applications is Constellation Network, which is also what makes Real Estate Ledger possible. Pricing real estate is going to get smarter the same way poker did: by learning to weight what it knows, and being honest about what it does not.

What this means for sellers and agents

For most homeowners, this looks abstract. It is not. The shift is already happening in the parts of the market where buyers are sophisticated enough to act on it. Sellers with documented and verified property histories are getting offers faster, with fewer contingencies, at prices that reflect a measurable trust premium. The next generation of pricing models is going to formalize what the smartest buyers already know. The houses being priced inside that model in five years are the ones being documented today.