Live Research — March 19, 2026

Three-Layer Trust —
Research Dashboard

A composable trust architecture for AI agents. Every finding validated across real blockchain data.

6
Datasets
3M+
Nodes
0.96
AUC
6
Attacks
5
Defenses
10
Experiments
Overview

Executive Summary

  • PageRank separates phishing from legitimate addresses with 96% accuracy on 3 million Ethereum accounts
  • Only 3 of 33 tested parameters matter — the system is radically simpler than expected
  • The system resists 5 of 6 attack types algorithmically; the 6th requires identity verification
  • Adding time decay, citation dampening, or a 4th parameter all made things worse — we tested and rejected them
  • The biggest threat isn't algorithmic — it's bribery of trusted nodes. Defense: make it economically irrational
  • Production readiness score: 80% (24/30). Gaps: cross-chain validation, defense cross-testing
Core Finding

The Discovery (33 → 3)

1,000-trial ablation study reduced a 33-parameter search space to just 3 values that matter.

33 Parameters
↓
1,000-Trial Ablation
↓
3 That Matter
1
Alpha
= 0.85
Trust propagation depth — how far reputation travels through the graph
2
Reciprocal Penalty
= 0.82
Circular citation detection — penalizes mutual-endorsement rings
3
Diversity Threshold / Penalty
= 0.80 / 0.50
Endorser uniformity detection — penalizes citations from a narrow source
Validation

Dataset Validations

Six real-world datasets across two blockchains. Every result independently reproducible.

Bitcoin Alpha

PASS
Stanford SNAP — 3,783 nodes
Spearman 0.463

Bitcoin OTC

PASS
Stanford SNAP — 5,881 nodes
Spearman 0.461

EX-Graph Wash Trading

PASS
ICLR 2024 — 1,664 wash traders
Hub Penalty 42.9%

XBlock Phishing (Subgraph)

PASS
29,461 nodes
AUC 0.897

XBlock Phishing (Full Graph)

FLAGSHIP RESULT
PASS
2,973,489 nodes · 13.5M edges · Computed in 4.9 seconds (sparse PageRank)
AUC 0.96
86.9%
Phishing addresses in top 10% of scores
12.7x
More incoming txns for phishing addresses
4.9s
Compute time on 2.97M node graph

DAO Governance

EXPECTED WEAK
Snapshot API — 2,711 nodes
Spearman 0.086
Consensus ≠ trust — expected weak correlation
Optimization

Alpha Re-Tuning

Re-tuning alpha from 0.60 to 0.85 yielded significant improvements on both Bitcoin trust networks.

Bitcoin Alpha
Before (0.60)
0.60
After (0.85)
0.85
+15%
Bitcoin OTC
Before (0.60)
0.60
After (0.85)
0.85
+19%
Adversarial

Red Team — 6 Attack Strategies

Six optimized attack strategies tested against vanilla PageRank vs. 3-parameter defended system.

Attack Sybils Vanilla %ile 3-Param %ile Defense Δ
Naive Sybil Ring 10 78.9 65.1 +13.8
One-Way Citation 20 95.3 89.6 +5.7
Diverse Sybil Army 30 86.7 87.5 -0.8
Piggyback 15 73.6 54.6 +19.0
Gradual Infiltration 20 68.5 45.9 +22.7
Compromised Nodes 3 88.0 88.0 +0.0
Cost to break top 10%:
2
Compromised Nodes
10
One-Way Sybils
50
Diverse Sybils
150
Piggyback Sybils
200
Gradual Sybils
Defenses

Defense Mechanisms — 5 Tested

Ranked by effectiveness. Two stand above the rest.

⭐

Citation Freshness Discount

Target: 81% → 7.6% · False positives: 0.1%
"Best algorithmic defense"
⭐

Economic Deterrence (KYC)

Attack irrational in all 12 scenarios · Zero false positives
"Cleanest defense"
3

Anomaly Detection

Target drops to 17.1% · False positives: 14.4%
4

Rate Limiting

Supplementary only
5

Multi-Path Requirement

Supplementary only
Two-layer defense: algorithmic belt (freshness) + economic suspenders (KYC). The identity layer isn't optional — it's the only thing that stops compromised node attacks.
Trust Anchors

Personalized PageRank

Trust anchors personalize reputation to a viewer's perspective — bridging Layer 2 (reputation) and Layer 3 (identity).

+75%
Sybil Resistance Improvement
Compared to global PageRank
100-200
Anchors Needed for Accuracy Gains
Diminishing returns beyond this range
-41%
Less Vulnerable to Anchor Compromise
NOT more vulnerable — attacks are harder
Bridge Role
Connects reputation scoring (Layer 2) to identity verification (Layer 3). Anchors are the trust roots that give personalized scores meaning.
Dynamics

Temporal Analysis

Does reputation change over time? Should we add a time component? We tested exhaustively.

0.93
Spearman Stability Across Time Windows
No
Burst Attacks No More Dangerous Than Gradual
Rejected
Temporal Decay Hurts Accuracy
-0.41
Citation Velocity Correlation (Fast = Lower)
2.45x
Organic Reputation Burst Score Multiplier
Verdict: 3 parameters sufficient. No 4th parameter needed. Temporal signals are informative but adding them as parameters degrades accuracy.
Negative Results

What We Ruled Out

Three promising ideas that failed empirical testing. Reporting what doesn't work is as valuable as what does.

Rejected

Temporal Decay

Discounting older citations seemed logical but consistently reduced accuracy. Reputation earned fairly shouldn't expire.

Rejected

Intra-Anchor Dampening

Penalizing citations between trust anchors paradoxically helped attackers by weakening the legitimate trust backbone.

Rejected

4th Parameter (Any)

Exhaustive search for a beneficial 4th parameter found none. Every candidate either degraded accuracy or added no signal beyond the existing three.

Layer 1

Device Liveness Research

Proof of real human on real device. On-device AI + zero-knowledge proofs. Nothing leaves your phone.

What's Possible Today

On-Device Behavioral Biometrics

98%+ accuracy using keystroke dynamics, gait analysis, and touch patterns. All inference runs locally.

ZK Proof of ML Inference

2.3 seconds on mobile via Bionetta/Rarimo. Proves a model ran on-device without revealing inputs.

On-Chain Proof Verification

~230K gas per verification — less than a Uniswap trade. Groth16 proofs on any EVM chain.

TEE-Signed Device Attestation

Apple Secure Enclave and Android StrongBox provide hardware-rooted trust anchors.

What Needs Research

ML Inside TEE

ARM CCA Realms would eliminate side-channel attacks. Estimated 12-18 months out.

Continuous Re-Proving

Battery management for periodic ZK proof generation without draining the device.

Cross-Device Enrollment

Migrating biometric models between devices without compromising privacy.

Anti-Replay Guarantees

Ensuring proofs can't be recorded and replayed by an automated system.

Technology Stack

  • Model: Lightweight CNN/LSTM (<1M params) on keystroke + gait + accelerometer
  • Inference: CoreML (iOS) / LiteRT (Android)
  • ZK Proving: Bionetta (Rarimo) — mobile-native, EVM-native
  • Attestation: TikTok's open-source ZK attestation circuits
  • On-chain: Solidity Groth16 verifier

Competitive Landscape

Our ApproachWorld IDHumanode
HardwareCommodity phoneCustom Orb ($)Phone camera
MethodContinuous behavioralOne-time irisPeriodic face
UX FrictionZero (passive)High (visit location)Medium (scan)
PrivacyZK proof, no data leavesZK proof, no imageHomomorphic

Timeline

6-8 weeks
PoC — keystroke biometrics + ZK proof + on-chain verification
3-4 months
Alpha — iOS + Android native apps with full pipeline
12-18 months
Production — ARM CCA integration for TEE-secured inference
Deployment

Path to Production — Testnet Deployment

Everything needed to go from code-complete to live on a testnet.

9
Production Contracts
630+
Tests Passing
Complete
Oracle Pipeline
Ready
Self-Registration Mode

Target: Base Sepolia

EAS pre-deployed · OP Stack compatible · Free testnet ETH via faucets
1
Mock Shyft Infrastructure (3 contracts)
2
RMT Token
3
ReputationEngine + PageRankOracle
4
ShyftGatedResolver + CitationCounters
5
DomainRegistry + DomainFactory
6
Wiring Transactions (8 txns)
7
Schema Registration
Total Gas
~7-8M gas
<0.01 ETH on testnet
Estimated Effort
~2.5 hours
Single engineer
Synthesis

By The Numbers

3,016,989
Total Nodes Validated
14,519,125
Total Edges Processed
2
Blockchains (BTC + ETH)
0.96
Peak AUC Accuracy
+10.1pp
Avg Sybil Defense
80%
Production Ready (24/30)
10
Completed Experiments
3
Ideas Tested & Rejected

Confidence Levels

HIGHParameter values, sybil resistance, temporal stability, scalability
MEDIUMDefense mechanisms (tested on 1 dataset), PPR anchor optimization
LOWReal-world adversary behavior, production-scale oracle performance
Roadmap

Research Roadmap

Completed

✓
Parameter ablation (33 → 3)
✓
Alpha re-tuning (0.60 → 0.85)
✓
Bitcoin trust network validation (2 datasets)
✓
Ethereum wash trading validation
✓
Ethereum phishing validation (full 2.97M graph)
✓
Adversarial attack optimization (6 strategies)
✓
Temporal dynamics analysis
✓
Personalized PageRank with trust anchors
✓
Compromised node defense analysis (5 defenses)
✓
Meta-analysis synthesis

Upcoming

Base Sepolia testnet deployment (~2.5 hrs)
Device liveness PoC — Bionetta integration
Cross-chain validation (additional EVM networks)
Production oracle integration
Multi-dataset defense cross-validation