# Blank-turn probe study: selected evidence

Retrieved September 12, 2026. These are selected findings from the owned Gecko library. The header is AI-generated editorial artwork; the chart, grids and tables are calculated from stored numeric output.

## Model and hypothesis

Six-max NLHE cash, 100bb starting stacks. BTN opens to 2.5bb, SB folds and BB calls. Pot: 5.5bb. No ante. Rake: 5%, capped at 4bb. BB has only a check at the flop root, and BTN checks behind. Turn options: check or bet 1.8bb, 4.1bb or 8.2bb. Source amounts and EVs divide by 10 to give bb, matching the consuming app.

Before EV analysis, we specified this screen: a 1.8bb probe should lose at most 0.1bb for at least 95% of eligible BB marginal range mass on each chosen low turn. Eligible hands have no pair, made straight or flush, flush draw, or four-to-a-straight draw, including gutshots and wheels.

Discovery flops: AsKh8d, Qs7h2d. Held-out: AsQh7d, AsJh6d, KsQh8d, Ks9h4d, Js8h3d, Ts7h4d. Low turn: 2c, except Qs7h2d uses 3c. Each flop's top-rank club pairing turn was also retrieved as a separate boundary comparison. All eight flop and sixteen turn trees were available. The primary test has eight low-turn boards, 1,388 eligible exact-hand observations and 1,008 held-out observations. Paired turns have no made-high-card cohort and do not enter the primary screen.

## Arithmetic and results

Local loss = highest available action EV minus chosen action EV. Within each board, use exported marginal combo reach. Across boards, give each board equal weight. Quantiles pool eligible hands after normalizing each board's total weight to one; they are not means of per-board quantiles.

All eight boards fail the 95%-within-0.1bb screen; six also fail at 0.2bb. Equal-board mean loss: 0.079974922bb. Median: 0.046bb. Pooled weighted 95th percentile: 0.257bb. Maximum: 0.450bb. Shares within 0.05bb, 0.1bb and 0.2bb: 53.6728%, 74.5180% and 88.8232%.

Kc9c on AsJh6d2c has EVs 0.016, 0.152, 0.134 and 0.584bb for 8.2bb, 4.1bb, 1.8bb and check. Small-bet loss: 0.584 minus 0.134 = 0.450bb. On Qs7h2d3c, the small bet has the highest stored EV, 0.408bb versus checking at 0.375bb: a 0.033bb gap. Compare alternatives within each row; different boards have different ranges and contexts.

## Range semantics and rounding

BB turn keys and weights exactly match its flop range after turn-card removal. BTN turn reach is consistent with its flop range times its check probability, allowing for two-decimal source rounding. Omitted keys are audited too: some Q8 combinations on KQ8 have products of 0.0007 or 0.0014 that round to absent. No hands were inserted or floored.

Used BB and facing-bet BTN probability vectors total 0.99, 1.00 or 1.01; divide each by its original total. Some five-action BTN nodes after a turn check total 0.98 or 1.02. Their action frequencies are unused; their entering ranges match the facing-bet nodes.

Range grids aggregate legal exact combos into hand classes by marginal reach. Entry weight = class mass / legal combo count. Action mix = sum of reach times normalized strategy / class mass. An absent class differs from missing data. Exact suits may differ.

Composition percentages describe current made hands, not showdown equity or range advantage. Flop-retention categories use the flop, with the eventual turn removed from both before/after columns. Turn categories can differ because hands improve.

Hero-specific response summaries remove villain combos conflicting with Kc9c before weighting normalized strategy. Compatible BTN mass: 252.61 on AsJh6d2c; 189.71 on Qs7h2d3c. Small-probe raise: to 8.6bb. An 8.2bb probe permits raises to 24.6bb or 97.5bb all-in; the article combines these frequencies.

## Files and limits

- board-summary.json: all sixteen turn summaries; primary eligible results only for eight low turns.
- chart-data.json and blank-turn-loss.svg: the low-turn screen.
- aggregates.json: equal-board results by discovery/held-out split.
- selected-hands.json: seven examples, including separate draw and paired-turn boundaries.
- range-views.json: two selected 169-cell grids.
- range-composition.json: both current marginal ranges in each turn case.
- flop-retention.json: two before/after check-back comparisons.
- response-composition.json: four hero-compatible response summaries.

Independent numerical reproduction of classification, source consistency, aggregates, selected examples, grids, responses and article tables passed. Full proprietary trees remain private.

Solver name/version, requested and achieved convergence, stopping criteria and full abstraction provenance are unavailable. Three decimals and the 0.01bb near-tie convention are not verified accuracy bounds. These are selected local comparisons with stored continuations, not a changed whole policy, adapted opponent, exploitability, win rate, bb/100 or occurrence frequencies. The evidence does not establish a causal equity/strong-hand-tail mechanism, universal passivity, human behavior, other positions/stacks/rake, ICM or multiway strategy. No river jobs, database changes or nodelocks were used.