Rise, sequence, rise

Pyth, 31 30 29 bytes

e+0f.x!sgM.:iVczdT2ZosN^STlcz

1 byte thanks to @Jakube.

Demonstration. Test Harness.

Input is given on STDIN, space separated. If newline separated input is allowed, I can shorten the program by 2 bytes.

Explanation:

e+0f.x!sgM.:iVczdT2ZosN^STlcz
                                  Implicit: z = input(), T = 10, Z = 0, d = ' '
                        ST        [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
                          lcz     len(z.split())
                       ^          All combinations w/replacement of that length.
                    osN           Order by increasing sum.
   f                              Filter on
              czd                 z.split(' ')
            iV   T                Vectorize the "Convert to base" operation over 
                                  the integers as strings and the base sequence.
          .:      2               Take length 2 subsequences.
        gM                        Map the >= operation over them.
      !s                          Sum and logically negate.
    .x             Z              If that throws an error, returns 0 (e.g. reject)
 +0                               Prepend a 0, in case no sequences are found.
e                                 Take the end of the list.

Including 1 in the list of possible bases is safe because i, which uses Python's int builtin, doesn't allow 1 as a base, and therefore always throws an error, which is caught and filtered out.

CJam, 43 bytes

0B,2>ea,m*{:+~}${ea::~_2$.b__Q|$=*@.b=}=p];

Reads command-line arguments and prints an array.

Try it online in the CJam interpreter.

Examples

$ cjam rise.cjam 12 11 10
[6 8 10]
$ cjam rise.cjam 19 18 17
0

How it works

0       e# Push a 0 (default return value).
B,2>    e# Push [0 ... 10] and remove the first two elements.
ea,     e# Push the number of command-line arguments (n).
m*      e# Cartesian power. Pushes all vectors of {2 ... 10}^n.
{:+~}$  e# Sort by the negated sums.
{       e# Find; for each vector V in {2 ... 10}^n:
  ea::~ e#   Evaluate each character of each command-line argument.
  _2$   e#   Copy the results and V.
  .b    e#   Vectorized base conversion (list to integer).
  __    e#   Push two copies.
  Q|$   e#   Deduplicate and sort the last copy.
  =     e#   Compare it to the first. Pushes 1/0 if equal/unequal.
  *     e#   Repeat the original result of .b that many times.
  @.b   e#   Vectorized base conversion (integer to list).
  =     e#   Compare the result to the modified command-line arguments.
        e#   Equality makes sure that the base was greater than all digits.
}=      e# If pushed 1, push V and break.
p       e# Print. Either prints the last V or 0 if none matched.
];      e# Clear the stack to avoid implicitly printing the 0 (if still present).

Julia, 176 156 145 118 109 99 97 bytes

A->try p=NaN;flipud(map(i->(k=11;t=p;while t<=(p=parseint("$i",k-=1))end;k),flipud(A)))catch;0end

Ungolfed:

function anonfunc(i)
  # Start with k=11 so that it evaluates to 10 on first while iteration
  k=11
  # set t to the previous value of p
  # Note: p here gets held over between iterations within the map
  t=p
  # Iterate through, dropping k by 1 and evaluating the integer in
  # base k and stopping if the value drops below t
  # Note: "p=" expression inside conditional to ensure k-=1 is evaluated
  # at least once (to make NaN work as desired)
  while t<=(p=parseint("$i",k-=1))
  end
  # if it dropped below t, return the base, k to be the corresponding
  # element in the map
  return k
end

function f(A)
  # Using try/catch to return 0 if no acceptable base found
  try
    # This is a trick to make sure the comparison in the while loop
    # evaluates to false on the first use of it (last value in A)
    p=NaN
    # Apply anonfunc to each element of A, starting with the last element
    # and store the result in S
    S=map(anonfunc,flipud(A))
    # S is backwards, so flip it and return it
    return flipud(S)
  catch
    # Will throw to here if parseint fails with the base due to having
    # a digit not acceptable in the base
    return 0
  end
end

Used with a 1d array input. If the function is assigned to c, then you would call c([12,11,10]) and it would output [6,8,10].

Note: I had used dec(i) inside the parseint command, but because i is a single-character variable name, and I don't need to access a component, I used "$i" to get the same result.