# Donk-bet frequency versus stored EV Research date: September 27, 2026. Published by the team behind GTO Gecko. ## Models and sample Eight-handed NLHE tournament chip-EV library output, without rake, ICM or bounties. The catalog configures a 1bb total-table ante. The 20bb files imply a 2bb open/call, 5.5bb flop pot, 18bb remaining and a 1.4bb lead. The 50bb files imply a 2.3bb open/call, 6.1bb pot, 47.7bb remaining and a 2bb lead. Each root offers that single lead size and check. Other lead sizes were not tested. These are native configurations with different ranges and sizes, not an isolated depth experiment. Original per-seat ante allocation was not supplied. Primary sample: 24 BTN-BB nodes. Flop ranks 654, 765, 876, 865, 754 and 543, each rainbow and two-tone at 20bb and 50bb. The 654/765 families were the eight discovery nodes; four other rank families supplied 16 held-out nodes. There are 21,036 positive-weight exact-hand rows, including 14,058 held out. They are purposefully selected files; rows from one board are not independent experiments. Secondary sample: eight boundary nodes (432 rainbow/two-tone, A72 rainbow, JT9 rainbow at both depths), and four UTG-BB comparisons (654/765 rainbow at both depths). All 36 requested files were retrieved and retained. Six files inspected during feasibility are recorded in the private plan; none were silently removed. ## Metrics and weighting For each hand: check shortfall = max(lead EV, check EV) - check EV. Lead shortfall uses the corresponding lead value. This compares one decision against stored continuations with the better stored action as benchmark; it does not compare against the exported mixed strategy. Continuation EV can exceed the current pot. EVs use the current application's conversion (raw value divided by 10). The export's 0.001bb EV resolution is not a numerical-accuracy bound. Two-decimal strategy vectors totaling 0.99 or 1.01 were normalized. All examined rows remained within that interval. No absent hands were injected or weights floored. Within a node, each legal exact combination is weighted once by its exported marginal range weight. No extra suited/pair/offsuit multiplicity is added. The 169-class grids aggregate legal suits using their weights; class entry density equals weighted mass divided by legal combinations after flop-card removal. They do not remove an opponent's exact holding. A separate compatible-opponent sensitivity, assuming factorized marginal ranges, changed primary lead shares by at most 0.358 percentage points. Unknown folded-card bunching prevents treating that as full-game occurrence weighting. Across nodes we use equal board-class means, not deal-frequency weights. Boundary and UTG cases remain separate. Quantiles use entering marginal range mass. The predeclared descriptive screen required at least 95% of each primary node's mass to have a stored check shortfall no larger than 0.1bb; sensitivities are 0.02, 0.05 and 0.2bb. The lowest observed share at 0.1bb is 99.3094%. All 24 primary nodes pass that arithmetic screen. It does not establish solver accuracy or validate a poker shortcut. ## Audit and inference limits Independent Decimal-based code read raw files and replicated all 36 root summaries and checked 108 root/first-response decisions. Executing the current app's actual decoder corroborated selected nodes, units and hand EVs. The public 338 grid cells, selected exact hands, composition counts and plotted means were checked against a separate implementation. These checks validate extraction and arithmetic, not the underlying equilibrium. The consistency screen flags an action played at least 10% while more than 0.1bb below the better stored action. Primary files contain 54 root hand/action flags across nine of 24 nodes and 344 first-response flags. Affected primary root hand mass reaches 1.25%. Both headline 654 rainbow configurations have first-response flags. The 20bb A72 rainbow boundary has 109 root flags affecting 11.32% of BB's marginal range mass. Those discrepancies are larger than display rounding alone. Their cause is undetermined: original solver version, abstraction, achieved convergence and strategy/EV export pairing are unavailable. A flag does not identify a solver bug. No flagged source rows were deleted to improve the result. The original all-check shortcut remains unsupported. The article's contribution is a study method: inspect frequency, both action EVs, exact suits and range reach separately; investigate larger gaps and inconsistent mixtures. Evaluating a coherent new strategy against an adapting opponent is a separate task. No full-tree re-solve, nodelock, river solve, human study or win-rate model was performed. ## Public artifacts - [Node summaries](node-summary.json): all 36 nodes, p50/p95/max shortfalls, threshold shares and consistency counts. - [Selected hands](selected-hands.json): 13 selected exact-hand examples supporting the article, including the diagnostic Qs2s row. Frequencies are fractions; EVs are bb. - [Range views](range-views.json): two 169-class views from the main 20bb 654 rainbow node. - [Range composition](range-composition.json): weighted made-hand categories and action shares. - [Chart](frequency-and-cost.svg): equal-node descriptive means on separately labeled scales. Full proprietary trees, private ledgers and access endpoints remain private. Header imagery is AI-generated decoration; numerical figures are deterministic. ## Primary context - [GTO Wizard: indifference](https://blog.gtowizard.com/the-three-laws-of-indifference/) - [GTO Wizard: donk bets](https://blog.gtowizard.com/is-donk-betting-for-donkeys/) - [GTO Wizard: nodelock example](https://blog.gtowizard.com/top_gto_wizard_features_the_pros_prioritize/) - [PioSolver: numbers and weighting](https://piosolver.com/docs/viewer/numbers_in_piosolver/) Accessed September 27, 2026. Their configurations differ from this study.