@Jeff,
I seem to remember seeing, in one of the BT papers or patents, a statement, vague and all-encompassing, but suggesting that all sorts of errors and stats, as well as just reconnects, are included in DLM's stability assessment.
I've been wondering whether maybe (there, I said it again!) DLM may also include corrected FECs in it's equations? Corrected FECs don't affect the data, don't need retransmissions, and so have no impact on user experience or perception of stability. But a high number of FECs may still indicate that corruption is occurring, regardless of whether it's being corrected by FEC. The authors of the DLM algorithm could have decided high FEC counts should not occur on a 'very stable line', and so take them into account before reducing target. I've always tended to ignore FECs in my own line stats, but I bet there's a lot fewer FECs with a massive margin than with a normal margin, and that could be another reason DLM apparently responds favourably?
Just a(nother) thought.
- 7LM