Training EML expression trees (Odrzywolek 2026) with gradient descent and then snapping soft input selectors to discrete choices should recover exact symbolic forms for elementary functions. Prior work reported commitment-based success criteria; we add a symbolic-correctness criterion and characterize the gap systematically. Three-phase training (Adam, entropy penalty, temperature annealing) drives every selector in every tree to a simplex vertex at every depth tested, so the commitment problem is solved. But vertex commitment does not guarantee correctness. We distinguish valid snaps (correct symbolic form, post-snap MAE below 1e-2) from false snaps (vertex selection but wrong form) across 240 training runs over three target functions and four tree depths. For ln(x), which requires depth 4 in a balanced binary tree, 18 of 20 seeds valid-snap to the exact form eml(1,eml(eml(1,x),1)). For exp(x) at its minimal depth 2, only 5 of 20 seeds find eml(x,1); the other 15 all collapse into a single competing basin, eml(x,x), with post-snap MAE 0.688. Extra depth correlates with higher valid-snap rate for exp (17 of 20 at depth 4, unused subtrees collapsing to constants) but with lower valid-snap rate for ln at depth 5, which has more ways to misplace its three required gate levels; the mechanism behind the exp improvement is not established here. The valid/false distinction, which prior work did not make, reveals basin selection rather than commitment as the bottleneck for symbolic recovery in EML trees. Code and data are archived separately at 10.5281/zenodo.19736075. Version history (full details in the repository CHANGELOG) v2.1 (April 24 2026) — adds Sect. 4 connection to Odrzywolek SI warm-start evidence; extended reference list. v2.2 (April 24 2026) — incorporates external review: removes the claim that prior work conflated commitment and validity; removes the unsupported gradient-escape-routes mechanism claim (Sect. 3.4); softens "solves the commitment problem" to "across all tested conditions"; explains the 0.000 variance in exp d=2 false snaps. No data changed. v2.3 (June 12 2026) — minor revision: residual causal phrasing for the exp(x) over-depth improvement replaced with correlational language, per the paper's own Sect. 3.4 disclaimer; duplicated abstract sentence removed. No data changed.
symbolic regression · EML operator · neural symbolic · expression trees · basin selection · temperature annealing · elementary functions · functional completeness