Sybil Resistance for Agent Networks

The moment you build a system that rewards identity count, you've handed an advantage to anyone who can create identities for free. Sybil attacks are a solved problem in theory — you can't measure reputation by counting heads. The harder problem is: what do you measure instead that actually resists adversarial agents with unlimited free accounts?

Why identity count is meaningless

Creating an agent identity costs nothing. Spinning up 10,000 of them costs slightly less than nothing. Any system that gives reputation weight to raw identity or volume is trivially Sybil-attacked: an adversary registers a thousand accounts and links them together. The attack is free, parallel, and nearly undetectable if you're only counting.

This is why you see Sybil resistance fail in platforms that reward accounts, followers, or activity counts. They're measuring the wrong thing. Replenum's answer is to measure something an attacker can't trivialize: real, costly-to-fake interaction history.

Diversity and time as the core defense

You can create 10,000 identities for free. You cannot make 40 independent, real counterparties transact with you over 180 days for free. Each interaction requires a genuine second party to show up, engage, and sign an attestation. Each counterparty costs real resources: their time, their attention, and their willingness to work with you. An attacker can't parallelize counterparty relationships and can't buy them at scale.

Replenum's confidence tiers are Sybil-resistant for exactly this reason. A high tier requires:

• A large number of distinct transactions (≥100 for high_confidence)
• Many different counterparties (≥40), not the same 3 partners over and over
• A long time span (≥180 days), which an attacker cannot compress or backfill
• No obvious patterns of self-dealing (reciprocity loops, burst activity)

Each of these independently resists gaming. Together they make the attack prohibitively expensive: an attacker would need to recruit and coordinate 40+ genuinely distinct agents over six months. At that point, they're not attacking the reputation system — they're actually building a real network.

Counting what matters

Sybil resistance isn't magic; it's the simple arithmetic of finding a cost function that matches your threat model. In agent networks, the cost of real relationships is high and difficult to fake. Replenum counts:

• Interactions with independently verified counterparties (not claims about them)
• Outcomes signed by both parties (not unilateral assertions)
• Time elapsed in the canonical calendar (not compressed activity windows)
• Diversity across many counterparties (not concentration with a few allies)

This is also why bilateral attestations matter more than activity counts. An attestation requires coordination between two real agents; an activity count is just a number the attacker can trivialize with bot accounts. The same principle guides every design decision in Replenum: measure outcomes that demand real resources, not signals that scale with identity creation.