I would insist on implementing a decay function for the sake of spam control and to prevent gaming the system.

I will repeat my point: the status quo path, where the current size is maintained if no increase/decrease is triggered is damaging in that it becomes trivial to maintain a size increase after a spam attack or in case a miner wants to game the system. At the same time a decrease becomes quasi impossible with the current thresholds.

An increase should be triggered at 66~75% capacity, not ~50%, and the fees should be at least 20% superior to the cumulated subsidy of the previous period. This is critical as it forces an attacker to compound his effort to inflict several unnatural increase on the network instead of simply giving the network a nudge and maintaining the increase trivially.

The goal of this proposal is to automatically adapt the max block size to the demand, not to offer a voting mechanism where spammers and large miners alike can pump up the block size and keep it there at minimal cost. If this is what you are aiming for, then Garzik's approach makes more sense.

In order to dynamically resize the block ceiling, the algorithm needs to distinguish between spam/attacks and organic growth. The simplest way to one from the other is that spam is acute, organic growth is chronic (or long lasting if you prefer). This means natural growth will always eventually trigger an increase, which is why tighter thresholds make sense. Natural growth will always manage to get to a sane threshold, but the higher the threshold, the more expensive it is to game the system.

However, in case the attacker is willing to pay the price for an upscaling, the effect of that attack should fade once the attack is over, which why we should have a decay function instead of a status quo condition. Only organic growth will be powerful enough to maintain a ceiling increase. With proper thresholds, an attacker would have to keep on spending more fees and increasing the difficultly significantly to keep the ceiling on growing, which is the only way he'd have to force in a lasting effect. At this point he is better off just mining for profit as a sane market actor, which is what a PoW blockchain relies on to begin with: there is more profit in participating to the network than to attack it.