amplified bond penalty (complementary with liquid alpha for delay-based weight copying mitigation)

[rad2520]

context:
Yuma Consensus clips validator weights set above the consensus weight down to the consensus weight. weights that had to be clipped to consensus are penalized by slashing a bond portion of validators based on the distance from consensus.
weights set below the consensus are not clipped or penalized via bond slashing, because they lead to a lower bond (ownership) in a higher incentive miner, inherently causing a reward penalty. the impact on rewards for distance above and below consensus is 1:1.

the dynamic encryption window (weight copying solution) depends on exceeding a threshold of reward penalty for out-of-consensus weights to decrypt. hence we can reduce the required encryption duration by amplifying the penalty for out-of-consensus weights (bond slashing).

important tradeoff:
since subnet validation is rarely perfectly uniform and has e.g. probabilistic components or otherwise sources of variance between validators, bond penalty comes with the inefficiency of also penalizing honest validators to some extend, leading to differences and fluctuations in APR, which i.e. distort the decision making process behind stake delegations.

this is why amplifying the bond penalty has to be a tradeoff decision local to subnets, depending on the expected level of variance in their validation methodology.

spec:

• new subnet consensus hyperparameter "boosted_beta" expects 0-1
• 0.1 = 10% increased bond slashing etc, so max is 2x bond penalty meaning halved encryption-length

with a non-zero parameter value, bond slashing now also has to be applied to below-consensus weights in order maintain 1:1 relationship, so

• for below-consensus weights, slash the boosted_beta portion (e.g. 0.1 = 10%) of bonds that the same distance above-consensus would've slashed. or from the other perspective, the bb portion of bonds the validator is naturally missing due to lower weight on the miner

open research:
if there were a reasonable mechanism to dynamically adjust beta local for every miner (or validator?), such that the honest variance tradeoff is minimized, it'd be superior to boosted_beta (sensitive_beta!) and a meaningful improvement to yuma consensus. since weight copying (during encryption window) is onchain indistinguishable from subjectivity variance between honest validator at least within a certain range of the consensus weight, this seems difficult.