determinism: 50 seeds regenerate identically (expression and list-function families)

expression family: error rate of a left-to-right solver that ignores precedence and brackets (2000 tasks per d)
  d  error rate
  1       0.000
  2       0.303
  3       0.593
  4       0.781
  6       0.922
  8       0.976

collisions at d = 1 (243 possible texts if operands and operators are uniform; the generator may repeat)
 N seeds  distinct   expected distinct
      50        44                45.3
     200       128               136.5
    1000       239               239.1
    5000       243               243.0
d = 3: distinct among 5000 seeds = 4986
d = 6: distinct among 5000 seeds = 5000

leakage of test tasks into the training set when the split is by seed range (train seeds 0-999, test 1000-1199)
d = 1: 0.970 of the test tasks also occur in the training set
d = 2: 0.085 of the test tasks also occur in the training set
d = 3: 0.000 of the test tasks also occur in the training set
d = 4: 0.000 of the test tasks also occur in the training set

ambiguity: for a depth-2 list program, the chance that another depth-<=2 program fits all n visible examples but is functionally different
  111 candidate programs, 80 functionally distinct on 300 probe inputs
 n visible  ambiguous rate
         1           0.328
         2           0.117
         3           0.068
         5           0.022
         8           0.015
