The Mix You Can't Actually Play

§1 — The column (QuintAce)

Every coach piece in this pipeline puts the solver in a fixed position: authority, judge, or oracle. The coach grades, gets graded, or gets their questions answered. The solver speaks; the coach reacts.

Strategy Debate reverses the structural frame. Two minds, two methods, one spot — and neither voice gets the last word by default. A player takes a position; the solver tests it; the player responds to the verdict; together we close with what a pro actually does in-game given both outputs. Debate, not grading.

For the inaugural entry we invited Xuan Liu — a Canadian pro whose analytical voice is one of the clearest in modern cash and tournament poker — to pose the debate. She brought a position she's held for years and that most pros at the highest levels would half-agree with on one day and push back on the next.

Her claim: the solver's mixed strategies are unplayable by humans, so the practical question is never "what's the mix?" — it's "what's the best pure?"

Before the claim, a primer on what a mixed strategy actually is:

We tested Xuan's claim across three spots, in the same SRP-with-button-position framework, across three different mix shapes. Xuan responded. We closed together.

⚠️ Review note — remove before publish. v1 is a ghost-drafted scaffold. Xuan's §2 (the claim) and §4 (the response) are our best-effort drafts written in her voice, grounded in the analytical register she uses publicly. She gets final cut. If she swaps sections, reframes, or concedes differently, v2 reflects her edits. The solver data in §3 is real and pre-registered; that part is fixed.

§2 — The claim (Xuan — ghost-drafted, awaiting her edit)

The solver is a calculator. I am not.

I've spent the past fifteen years learning to think like a poker player and about four of those years trying to think like a solver. Those two projects are not the same. The solver answers the question what fraction of the time should this exact hand take each action in an equilibrium against an opponent who also plays the equilibrium? That's a real question and it has a real answer. The answer, very often, looks like this: check 42%, bet small 12%, bet medium 24%, bet large 23%.

That answer is not a move I can make at a table.

Here is what actually happens when I try. I'm deep in a session. The spot comes up. I remember — dimly, from Tuesday's study block — that the solver mixed here. I don't remember the numbers. I pick the action I think was dominant. Sometimes I'm right; sometimes I've swapped which spot mixed; sometimes I pause for two seconds too long and my whole table notices. Next hand, I'm still thinking about the last one, and I misclick a standard three-bet sizing. The aggregate cost of trying to mix at the table is larger than the cost of giving up the equilibrium's last slice of EV and playing a pure strategy cleanly.

This is my position in three parts:

1. Perfect mixing is impossible for humans, in any setting. We don't have internal randomizers. We have moods, table dynamics, fatigue, and the attentional debt of a twelve-hour session. The "mix" at the table is always conditional on things the solver doesn't know about.
2. Attempted imperfect mixing is worse than clean pure execution. Because bad mixing leaks information (timing tells, sizing inconsistencies within a session, a player's observable tendency to bet big when they think the solver says bet big). Clean pure strategies leak less because they are internally consistent.
3. Therefore the practical question, at every mixed node, is "what's the best pure?" And: when the solver's answer is already pure — no mix at all — trust it over your mixing impulse, because your mixing impulse is usually a protection instinct talking.

I've been told this is the wrong frame, often by people I respect. The standard counter is that pure strategies are exploitable — that if I always bet here and always check there, a smart opponent maps my game and punishes it. The counter is technically right and practically almost always wrong at the stakes where I play, because (a) the pool is not mapping me at this granularity in real time, and (b) the exploitability gap between a well-chosen pure strategy and the equilibrium mix is, in the nodes that matter, fractions of a big blind.

So. Three spots. The solver mixes in each of them. I want to see what the pure costs.

§3 — The solver test (QuintAce)

We took Xuan's claim and ran it through three river-and-turn decisions in a 100bb 6-max cash setup, single-raised pots, hero on the button closing action against BB. Each spot picked for a distinct mix shape. All solver data from QuintAce's GTO engine, hero-specific strategies (not range averages).

Spot 1 — KhTh on 7♥️4♣️2♠️ — J♥️ — 9♣️. The whole tree is a dice roll.

A natural triple-barrel candidate that picked up draws on the turn and bricked them on the river. The widget below shows the full tree: four actions the solver spreads across. The strategic question is whether KhTh is indifferent enough that a pure simplification works — or whether the 4-way mix is actually load-bearing.

This is the hard version of Xuan's claim. KhTh isn't mixing between two actions — it's mixing across four, including two decisions that aren't on the same axis (whether to bet, and what size if yes). A human trying to execute this answer has to randomize twice. And the two reasonable simplifications point at different pure strategies: pick by biggest single bin and you land on check; pick by aggregate category and you land on a bet-75%. The solver is not refereeing between them.

Reach check. KhTh arrives at this node with reach probability 46.1% under equilibrium play — well into reliable territory. A hand with sub-5% reach is one the solver has barely explored, so fragmented mixes there can be numerical artifacts rather than real indifference. 46.1% means the 4-way mix above is the equilibrium doing real work, not noise. (More on the reach trust gate in §6.)

Spot 2 — TsTd on A♦️7♥️3♣️ — 2♠️. Sometimes the solver isn't mixing at all.

Pocket tens on an ace-high turn, facing BB's check — one of the most-argued spots in cash-game forums. Players instinctively want to bet for protection. See if you can predict what the solver does before revealing.

This spot is Xuan's thesis from the other side. The solver says don't mix — play pure, and the human mixing impulse ("should I bet for protection?") is the thing the solver is correcting. For a pro chasing the equilibrium, the hardest discipline isn't executing a 30/70 mix. It's resisting the urge to mix when the answer is pure — especially on spots like this one, where two-thousand forum posts a year argue you should be betting TT for protection against overcards.

Spot 3 — AsJh on 8♣️6♥️3♠️ — 2♥️ — A♣️, x-x turn. Sizing-only mix.

Top pair with a medium kicker after checking turn behind. The solver wants a bet — the question is how big. Pick a size and see how fragmented the distribution actually is.

This is a different mix shape from Spot 1. No check. The solver wants a bet with 100% certainty; it just can't decide which size. The real debate about Spot 3 isn't "bet or check?" — it's "does the human simplification 'always pick the middle size' cost anything meaningful?" When 80%+ of the equilibrium lands on the middle two sizes and the extremes are sprinkled in at single digits, the answer is almost certainly no.

The three shapes together

The three spots are three different mix architectures.

• Spot 1 is a vertical mix — the solver is indifferent between checking and betting, and within betting across multiple sizes. It's the hardest to simplify because the check option and the bet option probably serve different functions in the equilibrium (give-up vs. value-extraction).
• Spot 2 is no mix at all — the solver is pure, and the human error is the instinct to mix.
• Spot 3 is a sizing-only mix — the question is settled (always bet) and the open decision is granularity, which is the easiest mix to coarsen without paying for it.

Xuan's claim treats all three as "the solver mixes, humans can't mix, pick a pure." The solver's response is: these are not the same spot.

§4 — The response (Xuan — ghost-drafted, awaiting her edit)

Fair.

Let me start with where I was wrong and work up.

Spot 2 — full concession. You've actually strengthened my position here from a direction I hadn't used. I think about mixing mostly as "the solver rolls dice, humans can't, pick one." But this spot is something else — the solver isn't mixing at all, and the human error is the impulse to bet TsTd for protection on an ace-high board because that's what 2010-era poker training told us. When I get this kind of spot at a table, I'm still catching myself reaching for the bet. So yes: the discipline is not just "pick a pure when the solver mixes" — it's also "trust the pure when the solver is pure." Both failure modes live on the same axis. I'll take that as an upgrade to the claim.

Spot 1 — partial defense. You've shown me a four-way mix and asked me to pick a pure from it. That's exactly the spot my claim was designed for, and I'll stand by the rule: pick the dominant pure and stop arguing with yourself. Here the dominant single action is check at 51.7%. If I'm forced into one move, it's check. Not because I think check is necessarily better EV than each of the bet sizes individually at the node — it might not be — but because the variance of which pure I pick is itself an EV cost, and the most committable pure is the one the solver weights highest.

I know the counter: "check is only slightly over half; you're throwing away a 48.3% bet-aggregate that may net more EV in the real pool." Fair. But the rule has to be executable at 3 a.m. against a shortened attention span, and "pick the biggest single bin" is executable; "collapse by aggregate and decide again" requires math I won't do in the moment. The cost of making a different call on different nights because my head is tired is higher than the cost of the suboptimal rule.

The reach data you included makes this easier to defend, actually. 46% reach on the hand means this spot really is one my game runs into. It's not a rare-line curiosity. If I were hitting this node at sub-1% reach, I'd almost argue the opposite — don't bother having a rule for it, it comes up once a year, the expected cost of any reasonable pure is negligible. Reach shifts the question from "do I need a rule?" to "which rule?" — and for this node my rule is check.

Spot 3 — push back. This is where I actually disagree with the framing. You've called this a "mix" and I don't think it belongs in the same category as Spot 1 at all. Spot 3 is a question with a settled answer (bet) and an open parameter (size). That's not a mix in the sense my claim addresses. A sizing-tree with four options where 81% of weight is on the middle two sizes is a continuous distribution that's been bucketed. A pro picks the weighted middle — in this case the 50% size — and commits, and the exploit cost of that commitment is almost certainly under a big blind over a career. Your own data says so: 45.3% of the tree is already on that size. I'm not giving up much by always picking it.

So: concede Spot 2 as a strengthening of the claim. Stand on Spot 1 with the dominant-pure rule. Reject Spot 3 as miscategorized — it's not a mix in any practical sense.

Where that leaves the overall claim, re-stated: the human rule is "pure when the solver is pure, pure-dominant when the solver's mix is vertical, weighted-middle when the solver's mix is sizing-only." Three different rules because the three shapes are different things. My original claim was a one-rule-fits-all. That was too coarse.

§5 — The synthesis (joint)

The piece was set up as a debate about whether to mix. What it ended up being is a debate about what "mixing" even means at the table.

The three spots exposed three architectures that a pro has been treating as one. Xuan's starting claim — that practical poker is always "pick the best pure" — holds in a nuanced form after the data lands: the rule changes with the architecture, and the architecture isn't always what the reader's intuition flagged as "a mix."

None of these rules comes from the solver alone. The solver describes the equilibrium; it doesn't care whether a human can execute the equilibrium. What the synthesis adds is the executability overlay — the recognition that every equilibrium move has an execution cost that the solver doesn't price in, and that a pro's job is to pick the cheapest-to-execute pure that survives the exploit gap.

There's a sentence we kept circling back to in the draft:

"The solver is right about the equilibrium. You still can't do what it says. So the real question is: what's the best thing you can actually do."

Strategy Debate #1 doesn't resolve that question. It sharpens it. The next entries in the column will take it into spots where the executability overlay bites harder — multi-way pots, deep-stack play, spots where the solver's mix spans five or six actions instead of three or four, and the places where the "pure pilot" discipline stops working.

Two minds, two methods, one spot. When they disagree, you learn more than when they agree.

§6 — Methodology & caveats

Solver runs. All three spots run through QuintAce's GTO engine at 100bb, 6-max, $1/$2 NLHE, no ante, no rake modeling. Hero-specific strategies (not range averages) captured from solver_tool. All three hands end with a final bet-and-fold action in the hand history so that hero resolves correctly to the button (the in-position decision-maker). Scenarios constructed 2026-04-21 against QuintAce's current solver build.

Reach as a trust gate. Every hero-specific solver output comes with a reach probability — the frequency with which this exact hand reaches this exact node under equilibrium play. Low-reach nodes (the convention is "below ~5%") are out of distribution: the solver has barely explored them, so mixed strategies returned at those nodes can be numerical artifacts rather than real indifference. The reach values for the three spots in §3:

First pass of Spot 1 used KdQd at the same node, which reaches at only 1.4% — a solver-exploration red flag. We re-ran with candidate holdings that plausibly triple-barrel the line (AhQh 23.8%, AhTh 9.5%, KhTh 46.1%) and swapped to the highest-reach hand that preserved the 4-way mix pattern. This is the methodology discipline the piece was designed around; worth surfacing here because the column's editorial premise — debating the solver on its own terms — only holds if the solver is being queried where its answers can be trusted.

Bet sizes in the solver's action set are not exactly pot-fraction labels (the solver emits discrete bb amounts). Pot-fraction conversions in §3 tables are rounded for readability.

Spot 2 is asymmetric to Spots 1 and 3 — it's a turn decision, not a river decision, and the "pure" verdict is for TsTd specifically. Other hands at that node (QsQd, AhJd) will have different strategies. The point is not that the range is pure; it's that this specific representative holding is, against the human instinct to bet.

Single-hand analysis caveat. Each spot is one hand, one holding, one equilibrium run. The claim that "the dominant-pure rule works" across the spectrum of mixed-strategy spots needs many more runs than three to be defended at full rigor. This column's function is to debate the frame; future entries will drill into wider coverage per mix type.

Xuan's §2 and §4 are agent-drafted. The positions defended in those sections are constructed to represent the analytical register Xuan uses publicly; the specific phrasings, anecdotes ("3 a.m." fatigue framing, fifteen-year timeline, training-era references), and concessions are our best-effort curation. Xuan gets final cut in v2.

What this piece deliberately did not test. Multi-way spots. Deep-stack mixes where the solver's action set expands further. Spots where the mixed range is structurally functional (bluff-catch mixes with specific blockers). ICM-pressured spots where the solver's mix carries non-chip-EV weight. Those are candidates for future entries.

Column co-authors: Xuan Liu × QuintAce. Solver: QuintAce's GTO engine. Engagement mode: live co-authorship (real-time Google Doc). Draft lineage: v1 ghost-drafted 2026-04-21; v2 pending Xuan's edits.