This BIP describes a modification to the consensus-enforced semantics of the sequence number field to enable a signed transaction input to remain invalid for a defined period of time after confirmation of its corresponding outpoint, for the purpose of supporting consensus-enforced transaction replacement features.
Bitcoin has sequence number fields for each input of a transaction. The original idea appears to have been that the highest sequence number should dominate and miners should prefer it over lower sequence numbers. This was never really implemented, and the half-implemented code seemed to be making this assumption that miners would honestly prefer the higher sequence numbers, even if the lower ones were much much more profitable. That turns out to be a dangerous assumption, and so most technical people have assumed that kind of sequence number mediated replacement was useless because there was no way to enforce "honest" behavior, as even a few rational (profit maximizing) miners would break that completely. The change described by this BIP provides the missing piece that makes sequence numbers do something significant with respect to enforcing transaction replacement without assuming anything other than profit-maximizing behavior on the part of miners.
For transactions with an nVersion of 2 or greater, if the most significant bit (1 << 31) of a sequence number is clear, the remaining 31 bits are interpreted as an encoded relative lock-time. A sequence number with the most significant bit set is given no consensus meaning and can be included in any block, like normal, under all circumstances.
If the second most significant bit (1 << 30) is clear, the remaining bits reduced by 2^14 are interpreted as a minimum block-heigh constraint over the input's age. A sequence number of zero indicates a relative lock-time of zero blocks (bits 31 and 30 clear) and can be included in any block. A sequence number of 1 << 14 can be included in the next block after the input it is spending, or any block thereafter, rather than it being possible to be included in the same block. A sequence number of 2 << 14 can't be included until two blocks later, and so on.
Alternatively, if the second most significant bit (1 << 30) is set, the remaining bits reduced by 2^5 are interpreted as a minimum block-time constraint over the input's age. A sequence number with just that second most significant bit set (0x40000000) is interpreted as a relative lock-time of 0, measured in seconds, and can be included in the same block as the output being spent. Advancing that sequence number by 2^5 (0x40000020) constrains the transaction to be included in blocks with an nTime timestamp at least one second greater than the median time stamp of the 11 blocks prior to the block containing the coin being spent. Advancing the sequence number by an additional 2^5 (0x40000040) constrains the spend to be two seconds later, and so on.
This is proposed to be accomplished by replacing IsFinalTx() and CheckFinalTx(), existing consensus and non-consensus code functions that return true if a transaction's lock-time constraints are satisfied and false otherwise, with LockTime() and CheckLockTime(), new functions that return a non-zero value if a transaction's lock-time or sequence number constraints are not satisfied and zero otherwise:
Code conditional on the return value of IsFinalTx() / CheckLockTime() has to be updated as well, since the semantics of the return value has been inverted.
Using sequence numbers for locking inputs makes sense, since no transaction can be in a block before its parent transactions. This means that a lower sequence number can always be included earlier than a higher one (even if the time the original coins being spent was unknown when the transaction was authored). Because of this, even rational miners should go along with the scheme: Take the lowest sequence number and collect the fees, or skip over it in the hopes that no one else takes a lower number before the next available higher number becomes spendable. And, of course, users are not constrained to make their sequence numbers go up only one at a time. So it's "take the most updated version, now, or gamble that no one in the next dozen blocks takes the most updated and that you manage to include the next to most when it finally becomes mineable." This is similar to how lock-time works. In fact, this scheme is precisely a per-input relative lock-time.
A bidirectional payment channel can be established by two parties funding a single output in the following way: Alice funds a 1btc output which is the 2-of-2 multisig of Alice AND Bob, or Alice's key only after a sufficiently long timeout, e.g. 30 days or 4320 blocks. The channel-generating transaction is signed by Alice and broadcast to the network.
Alice desires to send Bob a payment of 0.1btc. She does so by constructing a transaction spending the 1btc output and sending 0.1btc to Bob and 0.9btc back to herself. She provides her signature for the 2-of-2 multisig constraint, and sets a relative lock-time using the sequence number field such that the transaction will become valid 24-hours or 144 blocks before the refund timeout. Two more times Alice sends Bob a payment of 0.1btc, each time generating and signing her half of a transaction spending the 1btc output and sending 0.2btc, then 0.3btc to Bob with a relative lock-time of 29 days from creation of the channel.
Bob now desires to send Alice a refund of 0.25btc. He does so by constructing a transaction spending the 1btc output and sending 0.95btc (= 0.7btc + 0.25btc) to Alice and 0.05btc to himself. Since Bob will still have in his logs the transaction giving him 0.7btc 29 days after the creation of the channel, Alice demands that this new transaction have a relative lock-time of 28 days so she has a full day to broadcast it before the next transaction matures.
Alice and Bob continue to make payments to each other, decrementing the relative lock-time by one day each time the channel switches direction, until the present time is reached or either party desires to close out the channel. A close-out is performed by finalizing the input (nSequence = MAX_INT) and both parties signing.
We reuse the double-threshold switch over mechanism from BIPs 34 and 66, with the same thresholds, but for nVersion = 4. The new rules are in effect for every block (at height H) with nVersion = 4 and at least 750 out of 1000 blocks preceding it (with heights H-1000..H-1) also have nVersion = 4. Furthermore, when 950 out of the 1000 blocks preceding a block do have nVersion = 4, nVersion = 3 blocks become invalid, and all further blocks enforce the new rules.
It is recommended that this soft-fork deployment trigger include other related proposals for improving Bitcoin's lock-time capabilities, including [https://github.com/bitcoin/bips/blob/master/bip-0065.mediawiki BIP 65]: OP_CHECKLOCKTIMEVERIFY, [https://github.com/bitcoin/bips/blob/master/bip-0112.mediawiki BIP 112]: CHECKSEQUENCEVERIFY, and [https://github.com/bitcoin/bips/blob/master/bip-0113.mediawiki BIP 113] Median time-past as endpoint for lock-time calculations.
The only use of sequence numbers by the Bitcoin Core reference client software is to disable checking the nLockTime constraints in a transaction. The semantics of that application are preserved by this BIP.
There may be other uses for the sequence number field that are not commonplace or yet to be discovered. To allow for other uses of the sequence number field, it is only interpreted as a relative lock-time as described in this BIP if the most significant bit is clear. This allows up to 31 bits of the sequence number field to be used for other purposes in applications which don't simultaneously need a per-input relative lock-time. In addition, the unused low-order bits of the relative lock-time encoding are available for use by future soft-fork extensions.
