Trust Model

Status: Designed, not yet implemented. Hello World uses only MediaWiki's stock access-level ladder. The model below is the target for the TrustGraph extension in a later phase.

Problem

Wikipedia-style reputation is earned through activity and admin discretion. There is no mechanism for an established user to invest their reputation in a newer user to fast-track them — even Stack Overflow's bonus on profile import is a one-time signal, not an ongoing stake.

We want a mechanism for trust transfer with skin in the game: vouching should cost the voucher when it goes wrong, and reward them when it goes right.

The hybrid model

Four layers stacked together:

1. Earned floor (Wikipedia-style)

Autoconfirmed gates remain the floor. New accounts can't bypass minimum account age and edit-count thresholds without a vouch. This is anti-Sybil insurance.

2. Explicit vouching (Vouch-style)

A user with sufficient reputation can vouch for another user. A vouch grants the vouchee a permission advance — they can bypass some autoconfirmed gates immediately.

The vouching primitive itself is inspired by mitchellh/vouch, which was designed for OSS project trust management in the AI-generated-spam era.

3. Stake-and-slash

Vouching locks up a fraction of the voucher's reputation. Two outcomes:

• Vouchee thrives (earns reputation organically through good contributions): the voucher's stake is returned with bonus. Vouching becomes a positive-sum action; vouchers are rewarded for picking well.
• Vouchee misbehaves (sanctioned via revert, block, or RevisionDelete): the voucher loses a fraction of their staked reputation. Vouching is costly when it goes wrong; encourages careful vouches.

Vouches can be revoked by the voucher (with a cooldown) and slashes cascade partially, not totally, up the chain — so a long-distance vouch chain isn't catastrophic to top-of-graph users when a downstream vouchee misbehaves.

4. Transitive propagation (EigenTrust-flavored)

Global reputation flows along the vouch graph with decay. A user's reputation is a weighted combination of their earned reputation and the flowed-in reputation from those who vouched for them. Decay prevents unbounded amplification through long chains.

The math is inspired by EigenTrust, but adapted to a vouching graph rather than a peer-to-peer transaction graph.

Editor / oversighter tier

The top of the trust ladder — users with RevisionDelete and suppression powers — is granted by admins, not earned by stake. This is a deliberate capture-resistance measure: oversighter authority shouldn't be purchasable with reputation.

Worked example

Alice has 1000 reputation. Bob is a new user.

1. Alice vouches for Bob, staking 50 reputation. Alice now has 950 available + 50 staked-on-bob.
2. Bob makes 20 good edits over a month, gaining 30 reputation organically.
3. Alice's stake unlocks: she gets 50 back + a 10-rep bonus (for picking well). Alice now has 1010 available.

Alternative path: Bob spams. After 3 reverted edits, Bob is sanctioned. Alice loses 25 of her 50 staked (50% slash). Alice now has 975 available. A small fraction also propagates up to anyone who vouched for Alice.

Implementation

Planned as the TrustGraph MediaWiki extension. Vouches stored as a custom data type; reputation accounting in dedicated tables; transitive propagation computed on a schedule via background job. Deferred from Hello World — see Roadmap.