# Cold-call stack-depth study: public evidence bundle

This bundle contains the sanitized, aggregate-only results behind GTO Gecko's September 4, 2026 article comparing cold-call composition at 50bb and 100bb.

## Files

- `summary.json`: assumptions, all four aggregate action distributions, comparison metrics, verification status, and limitations.
- `four-node-summary.csv`: the four matched spot comparisons in a compact table.
- `preregistration.md`: the sanitized reporting plan and its disclosure about the earlier feasibility check.

## Fixed setup

The snapshot uses selected precomputed GTO Gecko 6-max regular no-limit Hold’em cash spots at 50bb and 100bb: no ante, no straddle, 5% rake capped at 4bb, a 2.5bb open, and first-response options of fold, call, raise to 7.5bb, or all-in. The four comparisons represent eight decision nodes. Each row in `summary.json` states the full preflop action history, including folded seats. Every aggregate is weighted by physical starting-hand combinations across the 1,326 possible combinations.

## Comparison metrics

For every starting-hand class, take the lower of its call mass at 50bb and 100bb. Sum those minima, then divide by the smaller depth's total call mass. A 75% result means three quarters of the smaller calling range survives at the other depth; it does not mean that 75% of named hand classes are identical.

Relative call-rate change is `(call rate at 100bb - call rate at 50bb) / call rate at 50bb × 100`. Non-overlap is `100% - overlap` and is likewise expressed relative to the smaller calling range.

## Verification and disclosure

An HJ-versus-UTG feasibility result was known before the reporting plan was timestamped. The plan fixed the three then-unseen comparisons and the joint four-spot reporting rule before the new extraction. A separate Python implementation re-fetched the source and completed 1,384 structural and invariant validations plus 64 numerical comparisons. All numerical comparisons passed at a tolerance of 1e-12.

The full 169-class matrices, storage locations, internal identifiers, and EV arrays are not redistributed. They remain in a private audit record available to authorized internal reviewers. The public bundle lets readers inspect the method and reconcile the published aggregate outputs; it does not contain enough hand-level input to recompute overlap independently.

## Interpretation boundary

This is a point-in-time model comparison, not a sample of human hands, so sampling confidence intervals and claims of statistical significance do not apply. It does not determine the best action at another rake, sizing, table format, stack, seat order, or player pool, and it says nothing about future postflop EV.