Thanks for your responses.
When set at 10Mbps I can send any packet size up to the maximum with 100% reliability. If I select 100Mbps line speed, regardless of packet size, the reliability is unpredictable and too low to make the network useable.
While running several ping tests under the same conditions with 16 byte packets, the results varied from 16% loss up to 80% loss.
The RJ-45 plugs were fitted by me (which is a worry) using the proper crimping tool etc. The actual cable runs were bought in. The writing on the side says it is Austel approved, verified Cat-5.
Is there a possibility that attaching the RJ-45 plugs isn't as easy as it appears, and that I may not be getting a good enough connection at each of the cable ends?
Also, is there a proven method for testing attenuation over a run?
I've fairly much ruled out the cards as being the problem -- for testing, I disconnect all but two of the nodes and I've tried various combinations. I figured testing it with the two Intel cards in the Thinkpads would be a better indication than the cheaper cards in the desktop PC's, but it is just as unreliable.