You don't define "expensive", so this is something of a shot in the dark.
Start with a commercially (including eBay) 10 MHz generator. Rubidium for choice, but whatever accuracy you can get sets your performance.
Now build a progammable divider of 28 bits length. At 10 MHz you can get away with 74HC CMOS logic, but you'll need to use a fast carry configuration. The output also triggers a divide by two flip-flop which provides bit 29.
The divider can run at a ratio of 10,274,912 or 10,274,913, depending on the state of bit 29. For a perfect 10 MHz input the effective output period for bit 28 will then be 1.02749125 seconds, which is approximately accurate to 1 ppb, or about 30 msec/yr. A less-accurate input, of course, will produce a less-accurate output.
Using bog-standard 74HC161s you can do this with 8 ICs, and if you're careful you might be able to use a standard prototyping strip board, although you'd want to be very careful about beefing up the ground system. Perfboard would be cheaper, more compact and more durable, but the wiring would be less convenient, since you'd need to solder the connections. You could then pot it in something like electronics-grade RTV (NOT the RTV you get at the hardware store), for a final module size in the range of 2 x 2 x 1/2 inches, not counting the oscillator.
Note that your performance standard, being linked to "regular" RTCs, is actually in the range of 1 sec/day accuracy, which is 30 times worse than this approach. So, first off, you can do away with the bit 29 stage or, alternatively, divide your 10 MHz to 5 MHz and use a ratio of 5,137,456. This lower clock rate at the counters will allow a simpler carry structure, avoiding the fast carry which would be necessary at 10 MHz. Your accuracy is now on the order of 60 msec/yr for a perfect clock.
A quick look at eBay shows a large number of 10 MHz OCXOs for less than 20 bucks. These will typically have stabilities of 1 ppb or better, with 0.2 ppb a fairly common spec. Get one of these and you should be in good shape. You'd want to borrow a fairly high resolution frequency/period meter to determine the actual output frequency, then adjust the division ratio to match.