Reading solver output responsibly

A solver report answers a question about a specified model. Keep the question beside the result: the ranges, board, pot, stacks, available actions and economic assumptions. A strategy color or a small EV difference becomes difficult to interpret when the screenshot leaves those details outside the frame.

Start by checking which hand and decision the display describes. Then separate action frequency, range weight, equity and expected value. You can read a number correctly while still applying it to a different decision from the one that produced it.

Every numerical report in this lesson is fictional and authored for off-table study. No solver produced or validated these figures. The inspector checks display rounding and comparison context; it does not solve a poker game, supply ranges or recommend an action.

Ace: “I saved the colored grid. Is that enough to review the hand?”

Aurora: “Save the model and the selected node too. The colors only make sense when we know what they describe.”

Keep a model cover sheet

Record the game variant, positions, board and action sequence. Identify the model’s starting point and the decision you are inspecting. A river response after a bet is a different node from the decision before that bet, even if both displays show the same board and hand.

Keep the pot and effective stack at the stated reference point. A number called stack may describe the chips available at the model’s root or at a later decision. Reconstruct the preceding contributions before comparing that number with the hand history you meant to study.

Save both starting ranges and their weights. A hand may enter the model at partial weight, or lose weight after an earlier action. Seeing it in a starting chart does not establish that it reaches every later node with its original mass.

State the return being modeled. Chips, big blinds and tournament prize equity have different meanings. Rake and other included costs can affect returns. A chip-EV figure cannot become a tournament payout conclusion merely by changing the label on the number.

Inspect the action tree

A model offers a specified set of actions and continuations. Check which bet sizes, raises, leads and later branches were included. If a size never appears in a strategy display, determine whether it was available with zero frequency or omitted from the tree entirely. Those observations support different explanations.

An omitted action cannot be selected within the tree, so its absence does not establish inferiority. Keep the restriction visible when explaining what that model answers.

For example, a fictional model may offer check and one bet size. Reading its bet frequency tells you how that model divides between those available choices. It does not compare the offered size with a smaller size, a larger size or a line that was never modeled.

When testing a tree change, keep the original configuration. Give the revised version its own identifier and note the added or removed branch. Otherwise two reports with the same board can appear to disagree when they actually answer different action-tree questions.

Read the metric and its denominator

Action frequency describes how often a strategy chooses an action at a decision. Equity concerns outcomes under specified hands, ranges and runouts. Action EV accounts for returns under the modeled betting and responses. Read the metric label before comparing figures.

If one exact hand bets 40% at a node, that does not mean the bet wins 40% of the time. It also does not mean 40% of the entire range bets. First establish whether the selected display describes one combination, a hand class, the whole range or a filtered subset.

A range weight describes a different quantity. Suppose a hand reaches the decision with half of an original unit of mass and then bets 40% at that node. Its betting mass relative to the original unit is 0.5 x 0.4 = 0.20. The conditional action frequency remains 40%.

Filters can change the group being summarized. A report restricted to one suit or one action may use a different denominator from the unfiltered range. Keep the filter state with the export, and check the product’s definition of any aggregate rather than assuming each visible square contributes equally.

Work a small weighted example

Consider a separate fictional two-entry report. Entry A has three units of supplied matchup mass and bets 25%. Entry B has one unit and bets 100%. The masses are already supplied for this exercise; we are not estimating them from a chart or from an opponent’s behavior.

Entry A contributes 3 x 0.25 = 0.75 units of betting mass. Entry B contributes 1 x 1 = 1 unit. Total betting mass is 1.75 out of four units, giving an aggregate betting frequency of 43.75%.

A plain average of 25% and 100% gives 62.5%. That answer gives the two entries equal importance despite the supplied three-to-one masses. The difference comes from the denominator and weighting, rather than from more precise rounding.

For a genuine report, verify how the software defines the weights used for that metric. Opposing ranges, card removal and action conditioning can matter. This worked example assumes its matchup masses are the correct ones for the requested aggregate; it does not derive a universal aggregation rule for every display.

Compare the same question in the same units

A direct action-EV comparison needs a common hand, model, decision node, unit and accounting basis. One figure may be net from the current decision while another includes an earlier investment. Matching the letters EV does not establish that the reference points match.

Big blinds and chips require a verified conversion before their numerical values can be compared. If the conversion is absent, leave the comparison unresolved. More decimal places cannot supply the missing blind size or tell you whether the amounts share the same accounting convention.

Matching units are also insufficient when the records describe different nodes. A check value before a bet and a response value after a bet are not automatically alternative actions at one decision. Retain the full action path so another reviewer can locate each number.

The inspector uses labels to make these distinctions explicit. Fictional M1 names an authored model, R1 and R2 name different decisions, and Ah Kh identifies the selected hand. These labels establish the intended comparison within the exercise; they do not provide a complete runnable poker model.

Separate rounding from accuracy

The matching-context pair stores check EV as 4.2000 bb and bet EV as 4.2040 bb. Both refer to the same fictional hand and node, net from that decision. Stored bet EV minus stored check EV is 0.0040 bb. This is authored decimal arithmetic, not a solved strategic advantage.

At two decimal places, both figures display as 4.20. At three places they display as 4.200 and 4.204. Showing four produces 4.2000 and 4.2040. The stored values and their arithmetic difference stay unchanged while the display reveals more of them.

Rounded equality does not prove equal underlying values. Conversely, a visible positive difference does not establish that the underlying calculation is accurate enough for the conclusion you want. Display precision and model accuracy answer different questions.

The inspector rounds its nonnegative values to the nearest displayed unit, with halfway cases upward. It offers zero through four places and prints a supported stored gap to four places. Extra digits supply no solve evidence.

Ace: “The two numbers both show 4.20. Have I proved a tie?”

Aurora: “You have proved that this display rounds them alike. Inspect the stored values and keep the reliability question separate.”

Inspect four fictional record pairs

Choose a record pair, a display precision and, if you want to test your reading, a comparison claim. Inspect first leaves the claim ungraded. Supported means that the supplied context permits the stored numerical difference; unresolved means a context gap blocks that comparison.

All four pairs show an authored check frequency of 60% and bet frequency of 40%. These figures belong to the selected hand at the stated node. They are included to practice reading metric labels, and they do not estimate a winning chance, equity or an opponent’s actual behavior.

The matching-context pair uses Fictional M1, node R1, Ah Kh and bb for both actions, with the same net accounting basis. The inspector can therefore subtract its stored values. It also shows an illustrative exploitability label, whose limited role is discussed below.

The mixed-unit pair labels check EV in bb and bet EV in chips, without a conversion. The different-node pair labels the records R1 and R2. In both cases the inspector prints the individual records and explains the missing context, while leaving the arithmetic gap unreported.

The no-convergence pair uses Fictional M2 with matching hand, node, unit and basis. Its stored gap can be shown, but convergence information is absent. The result keeps arithmetic comparability and unassessed numerical reliability in separate sentences.

Treat convergence information within its scope

Numerical solving can stop at a finite accuracy or time target, leaving an approximate strategy. Keep the settings and reported convergence information with the result. A screenshot that shows action EVs but removes the solve status leaves an important part of the study record unknown.

The matching fictional record includes an illustrative exploitability figure of 0.0100 bb per hand. It is merely an authored label in this exercise. The inspector does not derive it, validate it or use it as an acceptance threshold for the 0.0040 bb action gap.

A solution-level metric is not automatically an error bar for each individual hand’s EV. The fact that one displayed number exceeds another does not, by itself, prove that the action gap is meaningful or meaningless. You need the definition of the reported metric and its connection to the claim being assessed.

Record missing convergence information as missing. Decimal places do not establish how long a solve ran. Every pair here also lacks a genuine solve and full model inputs.

Read mixed actions in context

A finite numerical result may mix actions whose printed EVs differ slightly. That does not require a story about deliberately sacrificing value with the lower-looking action. The values are evaluated against the strategy in the solution being inspected, and finite precision can leave small differences.

At a reached decision in an exact equilibrium, actions genuinely mixed by the same hand have equal expected value against the equilibrium opposing strategy with the appropriate continuation. An approximate report should not be treated as an exact equilibrium certificate merely because its frequencies add to 100%.

Nor does a close local comparison authorize replacing every mixed action across a range with a favorite choice. A collection of such changes alters the overall strategy an opponent faces. Keep the local action values connected to the full strategy and model from which they came.

Ace: “This action has the higher printed EV. Should I always choose it?”

Aurora: “First check what was modeled and what the precision supports. A local reading exercise does not establish a whole strategy.”

Make a sensitivity test traceable

For an off-table study question, preserve the original model and change one relevant assumption in a separate version. You might investigate a starting range, an available size or a cost setting. Record which input changed before interpreting a shift in the output.

If several assumptions change together, the resulting difference belongs to that combined change. It cannot automatically be attributed to one selected input. A cleaner-looking report is less useful when the configuration that produced it has been lost.

If a frequency changes sharply with a plausible range weight, record that sensitivity. A single colored square cannot establish a fact about every opponent on a similar board.

Leave a reproducible study note

Save the original report, model configuration, node, hand, units, reference point, available actions, convergence record and filters. Include your study question so another reviewer can reproduce the comparison or name what is missing.

Try the matching pair at two decimals, then three. Write that the display initially hid a stored difference of 0.0040 bb. Next inspect the mixed-unit and different-node pairs and name the specific missing link instead of using more digits to make the comparison look complete.

Finish with the no-convergence pair. A useful note says that its supplied context permits subtraction while reliability remains unassessed. It preserves both observations and makes no claim that the fictional higher number guarantees profit in a real hand.

Use this lesson away from live decisions. For any actual software, check the room’s current rules for the situation in which you intend to use it. This page does not determine permission, and its authored inspector has no live-hand or solver connection.

For adults of legal playing age. Keep gambling within affordable limits and use this exercise for off-table study.

Back to top