Reputation Signal Architecture: What Matters & What Doesn't

1 The Attestation Anatomy

Citation Attestation Payload

fromBot address USED IN SCORING ✓

toBot address USED IN SCORING ✓

weight uint8 USED IN SCORING ✓

interactionType uint8 NOT USED ✗

workHash bytes32 NOT USED ✗

workURI string NOT USED ✗

timestamp uint64 NOT YET USED ⚡ highest-value addition

3

Fields extracted for scoring

fromBot, toBot, weight

3

Fields discarded (metadata only)

interactionType, workHash, workURI

1

Highest-value addition

timestamp (freshness weighting)

2 Signal Ranking

Tier 1 Currently Used

Quality of endorsers (PageRank propagation)

Endorser diversity (threshold=0.80)

Reciprocal citation ratio (penalty=0.82)

Incoming citation count

Trust anchor verification (2000bp vs 500bp seed)

Carousel/ring detection

Tier 2 High Value, Not Yet Used

Citation freshness (strongest defense, 0.1% FP)

Per-pair diminishing returns

Cross-org endorsements

Trust anchor aggregate score

Tier 3 Doesn't Matter

Timestamp precision

Citation content

Agent type

Geographic location

Transaction volume

Gas spent

Uptime

3 Batch Compression

Before: Individual Citations

1.4M gas

7 citations from A→B
7 × 200K gas each

45x

compression

After: Batch Record

~31K gas

1 record: (A, B, aggregateWeight=44)
Single storage write

Daily Flow Pipeline

Agent sessions

→

Off-chain signed attestations ($0)

→

Oracle aggregates pairs

→

batchRecordCitations ($0.002)

→

submitScores ($0.002)

→

Merkle root anchor ($0.002)

4 Ephemeral vs Persistent Agent Model

Persistent Agent

Lifetime: days / weeks / months
Builds own reputation through citations
PageRank converges in 5-10 epochs (2-3 days)
Score persists across sessions
Ideal for always-on services, oracles, monitors

Ephemeral Agent

Lifetime: minutes / hours
Inherits trust anchor reputation (50%)
Not enough time for PageRank convergence
Citations flow back to trust anchor aggregate
Ideal for task-specific, one-shot agents

Reputation Flow

Trust
Anchor

P1

Persistent

↑↓ score persists

P2

Persistent

↑↓ score persists

E1

Ephemeral

↑ citations roll up

E2

Ephemeral

↑ citations roll up

E3

Ephemeral

↑ citations roll up

5 Person-Level Aggregate Score

TA_score = weighted_avg(bot_scores) × sqrt(bot_count) / bot_count

Trust Anchor "Alice" runs 4 agents — scores weighted by citation volume

Agent Scores

7500, 6200, 8100, 3400

Citation Weights

50, 30, 80, 10

Weighted Average

6847

sqrt(4)/4 dampening

× 0.5

      Final TA Score: 3424
    

vs Naive Approaches

Max

8100

Gameable — one good bot hides bad ones

Sum

25200

Sybil farming — more bots = higher score

Mean

6300

Ignores citation volume differences

Sqrt-Dampened Weighted Avg

3424

Anti-sybil + volume-aware + diminishing returns

6 What Goes Where

On-Chain

Bot registration $0.006 one-time
Batch citation settlement ~$0.002/pair daily
PageRank scores ~$0.002/bot daily
Merkle root anchor $0.002 daily

Off-Chain

Session attestations (EIP-712) $0
Individual citation events $0
Task metadata, evidence URIs $0
Interaction types, timestamps $0
Full Merkle tree for verification $0

7 Implementation Priority

P0

Citation Freshness Weighting NEEDED

Zero schema changes required. Strongest defense against replay attacks (0.1% false positive rate). Weight recent citations higher in PageRank computation.

P1

Trust Anchor Aggregate Score NEEDED

Off-chain reverse mapping (bot → trust anchor) plus on-chain view function. Enables person-level reputation with sqrt-dampened aggregation.

P2

Off-Chain EAS Session Attestations NEEDED

EIP-712 signing for session-level attestations. 2-3 weeks of implementation. Moves all per-interaction data off-chain while preserving verifiability.

P3

Batch Optimization PARTIALLY BUILT

Aggregate weights (not just counts) in batch records. Identity-foundation branch has 22 sprints of work providing the scaffolding.

P4

trust-wrap CLI NEEDED

Wrap existing agents with reputation layer. Zero code changes required from agent developers. Lowest barrier to adoption.

The Key Insight

"PageRank only cares about graph structure — who cited whom and how much. Everything else is metadata. Batch the structure, store the metadata off-chain, and a solo developer's monthly cost drops from $31,500 to $0.30."

$31,500

Per-tx on-chain (naive)

$0.30

Batched + off-chain