YES(?, a + 3)

Initial complexity problem:
1:	T:
		(?, 1)    eval1(a, b) -> eval2(a, b) [ a >= 1 /\ b = a ]
		(?, 1)    eval2(a, b) -> eval2(a - 1, b - 1) [ a >= 1 ]
		(?, 1)    eval2(a, b) -> eval1(a, b) [ 0 >= a ]
		(1, 1)    start(a, b) -> eval1(a, b)
	start location:	start
	leaf cost:	0

Repeatedly removing leaves of the complexity graph in problem 1 produces the following problem:
2:	T:
		(?, 1)    eval1(a, b) -> eval2(a, b) [ a >= 1 /\ b = a ]
		(?, 1)    eval2(a, b) -> eval2(a - 1, b - 1) [ a >= 1 ]
		(1, 1)    start(a, b) -> eval1(a, b)
	start location:	start
	leaf cost:	1

Repeatedly propagating knowledge in problem 2 produces the following problem:
3:	T:
		(1, 1)    eval1(a, b) -> eval2(a, b) [ a >= 1 /\ b = a ]
		(?, 1)    eval2(a, b) -> eval2(a - 1, b - 1) [ a >= 1 ]
		(1, 1)    start(a, b) -> eval1(a, b)
	start location:	start
	leaf cost:	1

A polynomial rank function with
	Pol(eval1) = V_1
	Pol(eval2) = V_1
	Pol(start) = V_1
orients all transitions weakly and the transition
	eval2(a, b) -> eval2(a - 1, b - 1) [ a >= 1 ]
strictly and produces the following problem:
4:	T:
		(1, 1)    eval1(a, b) -> eval2(a, b) [ a >= 1 /\ b = a ]
		(a, 1)    eval2(a, b) -> eval2(a - 1, b - 1) [ a >= 1 ]
		(1, 1)    start(a, b) -> eval1(a, b)
	start location:	start
	leaf cost:	1

Complexity upper bound a + 3

Time: 0.101 sec (SMT: 0.097 sec)