reputations talk notes

[tribble@xanadu.com (E. Dean Tribble)]

These are the notes from my talk at the last meeting.  
They may only be parsable by people who saw the talk, 
but I'm interested in all comments.

Thanks to Mark Miller and Eric Drexler who helped me put
together these notes the first time.

The need for reputations:
	single Prisoner's Dilemna incentive for defecting
	iterated Prisoner's dilemna incentive for cooperating
	reps put each person in an iterated situation with the group
	better reputations reduce reputation float (see below)

Spreading reputations makes them better at encouraging cooperation

Uses of reputations:
	sorting
	filtering - this is just bad sorting
	collateral - for business and important trust issues
	let me add that reputation appication is largely on the receiving side

Negative reputations:
	don't work with anonymity because you just make another identity
	don't spread for psychological and legal reasons
		many people won't reveal negative opinions about others people
	primarily support filtering and collateral

Positive reputations:
	are safer to reveal and spread
	allow sorting
	can handle anonymous pseudonyms

Anonymity:
	reduces perceived and actual accountability
	breaks many of our trained intutions for trusting others
	implicit but unrevealed assets aren't at risk
	trade names are examples of non-anon. pseudonyms

Reputations systems for sorting mail are less of an issues than
reputations systems for actual business because there is less at risk.
Gaming of a reputation system depends on how much is at risk.

Brilliant Pennies: randomly generated reputations
	the con game:
		start with 1000 pennies
		flip them every day.  head the stock market goes up, tails it goes down
		throw out pennies that 'guessed' wrong
		at the end of ten days you have a penny with a 'reputation' 
			for predicting the stock market
		sell the penny anonymously :-)
	inconsistent positive reputations are better than perfect reputations
		can explain away first several mistakes

Assets that support an identity
	recoverable assets
		anything that can be recovered in the even of a defection
			bank accounts, goods, factories, etc.
		pseudonyms typically won't have recoverable assets
		pseudonyms could keep recoverable assets in a non-anonymous escrow
		can be anonymously and privately committed to more than one party in the
			event of a defection, so value can't properly be assessed
	unrecoverable assets
		Sunk costs
			time, money spent building a reputation
			solves the Brilliant Penny problem: 
				random rep requires huge investment
				a 30day coin-flipping rep requires more pennies than exist
			notarized money-destruction service:
				proves investment with no route for kickbacks to the investor
				demonstrates the comparative cost of running an anon. service
			our intuitions about sunk costs can be very wrong here
				we're used to sunk costs at least looking recoverable (liquidation)
				we're used to eventual personal accountability of decision maker
			sunk costs enable kill files again
				you can only afford so many pseudonyms
				this is relates to the Brilliant Penny solution
		reputation capital (perceived value of reputation)
			customer good will, market share, shelf space, customer mind-space, etc.
			priority settings in receivers's mail handlers (filters, sorters)
			future sales potential
			based on the net present value of the reputation
				perceived differently by different individuals (customers/producer)
				based on discount rate (subjective value of money now vs. money later?)
				perception of discount rate can change radically
					you are told you have 6 months to live

Reputation Float:
	latency between use of a reputation and the effect on the value of the reputation
		how long can you ride on past glory?
	this includes the amount of goods, etc. against which the reputation is collateral
		existing trade-names collateralize the quality of existing goods (i.e. cars)
	estimating reputation float is a public goods problem
		why should I reveal the goods I have outstanding if I can get the info anyway
	insurance (Idea Futures) doesn't help
		the anonymous could collect both from defaulting and from insurance
	our intuitions screw up at estimating this 
		we're not used to untraceable transactions

Miscellaneous:
	filtering/sorting reputations should be wrt to topics
		This generalizes to other criteria
		individuals have different levels of expertise in different areas
	synthesized reputations are reputations composed from multiple reputations
		chaining remailers composes their reputations for security
		voting among key providers composes their uncompromisability reputations