"EIP 14: Extended Review Cycle Detection"

→ → → [CLICK HERE to go vote](https://snapshot.box/#/s:ethos-labs.eth/proposal/0x7445c479583253682aba38d4147376ad6a73a58f85047e157638a9b7aac53e9d) ← ← ← ## Summary This proposal extends the existing Review-for-Review (R4R) neutralization to detect coordinated review cycles beyond direct reciprocity. Currently, only direct A↔B reciprocal reviews are detected and downweighted. This upgrade adds detection of triangle cycles (A→X→B→A) and rectangle cycles (A→X→Y→B→A), catching more sophisticated coordination patterns while preserving the existing unlock model. As part of this update, the maximum review score contribution is also reduced from 540 to 400, with redistributed weight flowing to vouch-based signals. ## Motivation The current R4R detection only catches the simplest coordination pattern: two users directly reviewing each other. In practice, coordinated groups have adapted by routing reviews through intermediaries: - **Triangles:** Alice reviews Bob, Bob reviews Carol, Carol reviews Alice — each gets a positive review without any direct reciprocation. - **Rectangles:** Similar pattern through two intermediaries, making the coordination even harder to spot. These patterns inflate scores without genuine community endorsement. Our preview data shows meaningful score corrections when cycle detection is enabled, particularly for users with high review counts and low independent engagement. We have noticed many profiles that have achieved scores that are high simply through reviewing in massive groups, and they offer 0 economic security to the scores, so we want to change that. This should also help with the idea of "global" versus "local" reputation — cycles are more likely to appear when someone is only reputable from a singular group of people supporting them (even organically) but less likely when someone has reputation from multiple sources. ## Specification 1. **Cycle detection depth:** Increase `maxCycleLength` from 2 (direct only) to 4 (direct + triangles + rectangles). Reviews that form part of any cycle up to length 4 are classified as reciprocated. 2. **Positive-only filtering:** Only positive reviews are eligible for cycle classification. Negative and neutral reviews are always treated as external. This ensures the system penalizes coordinated *inflation* without discounting legitimate retaliatory or neutral reviews. 3. **Unlock model preserved:** The existing R4R unlock model still applies — the first 10 reciprocated reviews always count, and each external review unlocks one additional reciprocated review. Reviews beyond the unlock threshold receive zero weight. 4. **ELO and sentiment modifiers:** All existing review weighting mechanisms (ELO-based author/subject score ratio, community vote sentiment modifier) continue to apply on top of cycle detection. 5. **Score weight rebalancing:** The maximum review score contribution is reduced from 540 to 400. This reclaimed weight is redistributed to vouch-based signals — the vouched ETH ceiling increases from 280 to 390, and the voucher count ceiling increases from 270 to 300. This rebalancing reflects the reduced signal value of reviews in high-cycle environments and strengthens the economic security of the overall score. ## Collusion Resistance Each extension of cycle detection increases the coordination cost required to inflate scores. The table below shows the optimal evasion strategy at each detection level, using a scenario where a coordinated group wants to deliver 100 undetected positive reviews to a single person: Under the current system, a group of 200 colluders can split into two groups that review each other without any direct reciprocation, fully evading detection. With EIP-14, that same 2-group strategy is caught through triangle and rectangle detection. To evade, groups must route reviews through chains of 5+ participants, requiring 500 people to achieve the same 100 undetected reviews — a 5x increase in coordination cost. The coordination cost scales sharply: not only do you need more people, but each additional participant must maintain a consistent positive review chain across multiple hops, making the scheme increasingly fragile and detectable through other signals. ## Impact Based on preview data across all profiles: - Scores only decrease or stay the same relative to production (no artificial boosts) - Users with no cycle involvement see zero change - The largest corrections apply to users with extensive reciprocated review networks - Non-profiled users (those without an Ethos profile) are unaffected — they retain production behavior, which gives them further advantage (scores like Vitalik, Cobie, and ZachXBT will remain high) - Score ceilings are rebalanced to reflect signal quality: review weight decreases (540→400) while vouch-based signals increase (vouched ETH 280→390, voucher count 270→300), improving resistance to review-only score inflation ## Future Considerations - Extend cycle detection beyond 4, to pentagons and beyond, as we notice more collusion. - Extend this type of detection to vouches ## Rollout This change will be deployed with a 60 day gradual rollout period, but may be updated to a shorter period if the team decides to do so.