FlopForgeTexas Hold'em strategy data laboratory
EnglishEN

QUALITY METRICS

Poker exploitability explained in practical terms

Exploitability asks how much a perfect best response could gain by deviating against the current strategy inside the modeled game. It is a quality metric, not a promise of real-table profit.

No accountEditor-published evidenceInputs stay local
GuidesPublished by FlopForgeEditorial method

From best responses to NashConv

For each player, compute the value of a best response against the other player’s current strategy and compare it with the on-policy value. Adding both unilateral gains gives NashConv.

In a two-player zero-sum game, exploitability is commonly reported as NashConv ÷ 2.

Always keep the unit visible

A value in chips cannot be compared across games with different blind sizes without conversion. FlopForge reports the chip value and a big-blind-normalized value where appropriate.

A threshold such as 0.05 BB is meaningful only with the exact game identity, iteration count, and stopping rule.

A low value can coexist with model error

A strategy can be nearly unexploitable in a tree that omits an important bet size, compresses future streets, or uses representative private hands.

The correct claim is therefore “converged inside the declared bounded model,” not “solved unrestricted poker.”

Use a fail-closed display gate

If a result misses the declared threshold, the interface should not reveal the preferred action indirectly through EV ordering, frequencies, highlights, or a strategy matrix.

Non-qualified output can still show iterations, measured exploitability, the target, and why the model was withheld.

CONTINUE IN FLOPFORGE

Inspect a measured solve

Run a study and read the quality metric beside the ranges, action tree, and engine identity that produced it.
Open Study Tools

PRIMARY SOURCES

Evidence used by this explanation

FAQ

Questions this page should not hide

Is lower always better?

For the same finite game and unit, lower means closer to equilibrium. It does not compare model fidelity between different trees.

Can exploitability be negative?

A correctly defined non-negative deviation metric should not be negative beyond tiny numerical tolerance.

Why hide action details above the gate?

Because frequencies and EV ordering can reconstruct the same recommendation the quality gate is supposed to withhold.