You're trusting these numbers with your win rate. So here's exactly where they come from, how accurate they are, and how we verify it - no black box, no hand-waving.
PokerCortex postflop solutions are computed by Deepsolver, a professional neural-network solving engine that we've partnered with as our computation backbone. We build the product - the interface, the preflop library, custom configurations, and everything on our roadmap - on top of that engine.
Traditional solvers like PioSolver use Counterfactual Regret Minimization (CFR): they build an enormous game tree and iterate over it until the strategy converges toward Nash equilibrium. It's accurate, but slow - solves take minutes to hours and require expensive hardware.
Deepsolver's approach keeps CFR as the core algorithm but uses artificial neural networks to estimate expected values at the terminal nodes of the tree. That removes the most expensive part of the computation, making it 10–100x faster than traditional solving - which is how a full postflop solution reaches your screen in about 5 seconds.
Solver accuracy is measured by exploitability, also called Nash Distance: how much a perfect opponent could win against the strategy, expressed as a percentage of the pot. Lower is better - 0% would be a perfectly unexploitable Nash equilibrium.
The numbers below are from Deepsolver's published benchmark study - the most recent publicly documented test of the engine. They're historical results: the engine has been continuously improved since, particularly on speed and accuracy, so treat them as a conservative baseline rather than the current ceiling. Updated public benchmarks are planned.
Illustrative distribution based on Deepsolver's historical benchmark data (mean 0.6% across 430 benchmarks). Source: Deepsolver - Speed & Precision: benchmarks and testing
For context: a strategy with under 1% pot exploitability is considered solved for most practical purposes - common convergence targets in traditional solvers sit around 0.5% of the pot. No human can identify or exploit leaks that small - the difference is far below the noise of real gameplay.
These numbers aren't theoretical. The verification process is direct: finished strategies are hand-locked into traditional previous-generation CFR solvers, which then compute exactly how much a perfect counter-strategy could win. That measured exploitability is what the benchmarks report - the neural network's output is graded by the very solvers it's compared against.
Because of the neural network approach, we can't display an exact Nash Distance for each individual solve the way a traditional solver can. Accuracy is established statistically, through the hundreds of benchmarks above, rather than per-spot. That's the trade-off that buys the ~100x speedup - and the benchmark distribution shows it's a safe one.
Correct - and that's expected. Here's why it doesn't matter:
GTO strategies are full of mixed strategies at indifference points: spots where betting and checking (or calling and folding) have virtually identical EV. Near these points, small shifts in how a solver splits between the two actions cost almost nothing in EV or exploitability - which is why different solvers, and even two runs of the same solver with different settings, legitimately disagree on the exact frequencies.
| What differs between solvers | What it means for you |
|---|---|
| Mixing frequencies (e.g. bet 55% vs 62%) | Near-zero impact - at indifference points, moderate frequency shifts give up almost no EV |
| Which "close" combos mix | Near-zero impact - similar combos are close substitutes, differing mainly in minor blocker effects |
| How much EV the overall strategy gives up | This is what matters - and it's exactly what the exploitability benchmarks measure |
The right question isn't "do the frequencies match PioSolver?" but "how much could a perfect opponent exploit this strategy?" - and the answer, measured against those same traditional solvers, is about 0.6% of the pot on average.
The engine is one part of the story. PokerCortex is the product built around it:
A study-first interface designed for learning rather than engineering - no game-tree configuration rabbit holes unless you want them. A complete 6-max preflop library, free for every user. Custom solve configurations for your exact ranges, sizes, and stack depths. And an aggressive roadmap - Trainer, Flop Reports, and more - shipping continuously to early members at their locked-in price.
Questions about our methodology? Ask us anything - we'd rather over-explain than leave doubt. Reach us at contact@poker-cortex.com or via the Contact tab inside the app.