Reputation simulation

Question (Principles and commitments, “Past recognition and present judgement”): should a vote’s weight depend on recent reputation as well as on lifetime reputation, before the MainNet contract fixes the rule forever? And how do the rules hold up against groups of accounts that vote for each other?

Method. sim.ts (in this folder, node research/reputation-sim/sim.ts) plays ten years of one field, a week per step, with eight seeds per case. It follows the contract where it matters:

votes add their weight to an article, and its author claims that weight as reputation;
at most 50 reputation per day (350 a week);
a vote weighs 1 + ⌊√R⌋;
a flag needs reputation 10 and weighs like a vote;
flags of weight 30 dispute an article.

The honest field starts with 40 founders and gains 30 newcomers a year. Researchers have a talent between 0 and 1 and publish about four articles a year; founders publish less after their fourth year. Each week a researcher reads six recent articles and votes for those that seem good. Every article has one author. Useful-review and comment reputation are left out.

Five ways of weighing a vote were compared. Reputation itself, the record of recognition, stays lifetime in every one:

Rule	Vote weight
lifetime (today)	1 + ⌊√R⌋, R = reputation
recent	1 + ⌊√Rr⌋, Rr = reputation gained in the last five years
blend	1 + ⌊(√R + √Rr) / 2⌋
repeat damping	as lifetime or blend, but a voter’s n-th vote for the same author counts 1/n of its weight

The numbers are a model’s. They show directions and orders of magnitude, not forecasts.

Results

Honest field (entrenchment)

Rule	Founders’ share of vote weight, year 10	Founders’ share of good articles, year 10	Weight vs talent	Weight vs recent work
lifetime (today)	14%	2%	0.61	0.61
recent	12%	2%	0.64	0.66
blend	13%	2%	0.63	0.64
lifetime + repeat damping	9%	2%	0.68	0.73
blend + repeat damping	8%	2%	0.67	0.73

Ring of 10 fresh addresses voting for each other (from year 2)

Rule	Weeks to flag	Weeks to dispute any article alone	Weeks to the top 10% of vote weight	Ring reputation / best researcher’s, year 10	Ring articles in the top 50
lifetime (today)	1.0	1.0	36.1	1.15	0%
recent	1.0	1.0	36.1	1.15	0%
blend	1.0	1.0	36.1	1.15	0%
lifetime + repeat damping	1.0	1.0	never	0.02	0%
blend + repeat damping	1.0	1.0	never	0.02	0%

Cartel of the 10 most reputed researchers (from year 3)

Rule	Cartel members’ reputation / peers’	Weight vs talent
lifetime (today)	1.17x	0.60
recent	1.18x	0.63
blend	1.17x	0.62
lifetime + repeat damping	1.30x	0.69
blend + repeat damping	1.30x	0.69

A contrarian result (year 6), 60% of the field biased against it

Rule	Contrarian article’s percentile after a year	Founders’ share of vote weight, year 10
lifetime (today)	64%	14%
recent	64%	12%
blend	64%	13%
lifetime + repeat damping	91%	9%
blend + repeat damping	91%	8%

What it says

Weight by recent reputation changes little. Against lifetime, it moves the founders’ share of weight by one or two points and the link between weight and recent work from 0.61 to 0.66. The square root and the daily limit already damp entrenchment, and the founders’ weight stays close to their share of the field. A recent-reputation rule would need reputation stored by period in the contract, for a small gain.
The weak point is the ring, not the passage of time. Ten fresh addresses voting for each other reach the maximum reputation the daily limit allows. That is 15% more than the field’s best honest researcher, and they enter the top tenth of vote weight in about 36 weeks. They cost about 5 ALGO in the first week (10 articles, 90 votes, fees and flags).
Damping repeated votes neutralises the ring. When a voter’s n-th vote for the same author counts 1/n:
- the ring stays at 2% of the best researcher’s reputation and never reaches the top tenth;
- the founders’ share of weight falls from 14% to 9%;
- the link between weight and recent work rises to 0.73;
- the contrarian article rises from the 64th to the 91st percentile, because it receives first-time votes while established authors lose the extra weight of loyal repeat voters.
Honest readers rarely vote many times for the same author, so they lose little. The contract can keep one counter per pair of voter and author, a box whose deposit (about 0.03 ALGO) the voter pays at their first vote for an author.
Damping does not stop a cartel of established researchers. They keep the support of the rest of the field and still gain from each other: 1.30x their peers. The daily limit caps them a little more without damping (1.17x), because they hit it. A cartel is visible: it needs the detection of coordinated voting off chain.
Flags can be turned into a weapon in a week, under every rule. A ring reaches reputation 10, the flag threshold, in a week. Ten such flaggers outweigh the dispute threshold of 30, so they can dispute any article of the field, and a dispute pauses its votes and claims until governance acts. No vote-weight rule changes that. It needs its own remedies.

Recommendation for the MainNet contract (v4)

Keep reputation lifetime, as the record of recognition. Do not weigh votes by recent reputation.
Damp repeated votes: a voter’s n-th vote for the same recipient counts 1/n. Apply it to votes for articles and to votes for reviews and comments, since a ring could farm “useful” votes, which grant three times the weight.
Make false flags costly: when a dispute is cleared, the flags of that round count less in the flaggers’ next flags (for example, divided by one plus their cleared flags).
Bound the harm of a dispute: a dispute not resolved within a set time (for example 60 days) clears by itself. A lost or unreachable governance key then cannot freeze articles forever.
Settle disputes by a qualified jury, as decided in D-150; the contract executes its verdict through governance in the first phase.

Before deploying v4, run this simulation again with the exact v4 rules, add review and comment reputation and co-authors, and include the results in the external security review.