RMT Trust System: Consolidated Research & Roadmap

Cross-model validated architecture for AI agent reputation

March 20, 2026 — Master Status Page

Section 1

Where We Are

10
Experiments Completed
6
Datasets Validated (3M+ nodes)
4
Models Debated the Architecture
40%
Production Readiness (Grok CTO)

We built and validated a PageRank-based reputation system for AI agents. 3 algorithm parameters matter. Tested against 6 real-world datasets including 2.97M Ethereum transactions (AUC 0.96). 6 attack types tested with 5 defense mechanisms. 4 models debated (Claude + Gemini + Codex + Grok) and refined the architecture. Grok CTO assessment: production readiness is 40%, not 55%. Priority #1: Deploy to Base Sepolia testnet. Research phase is complete. Next: stop iterating and ship it.

"The single biggest risk is failing to deploy a working testnet soon, leading to endless internal iteration without real-world validation."
— Grok (CTO / Arbiter)
Section 2

The Model — What We're Building

Layer 3 — Identity Anchor

Shyft KYC Trust Channels

Every agent traces to a verified human via Shyft attestation trust channels. Establishes accountability and enables revocation.

Stops: impersonation, anonymous sybils Status: BUILT
Layer 2 — Citation Reputation

PageRank with 3 Parameters, 4 Input Fields

Graph-based reputation scoring. Agents cite each other; citations form a directed graph processed by tuned PageRank with sybil detection.

Stops: collusion, citation farming, wash trading Status: BUILT
Layer 1 — Device Liveness — CUT

TEE + ZK Proofs

Hardware-bound proof that the agent runs on a real device. Prevents virtualized sybil farms. Relies on App Attest, ZKML, or TEE attestation.

Stops: VM farms, cloned agents Status: CUT (Grok CTO decision)
Section 3

The 3 Parameters — Core Finding

alpha
0.85
Damping factor — how much reputation propagates through the graph
reciprocal_penalty
0.82
Discount for mutual endorsements (collusion detection)
diversity_threshold
0.80
Penalty for endorsement concentration
These were validated across 6 datasets. Cross-model debate confirmed these are UNCHANGED. The "4 fields" finding is about input data, not algorithm parameters.
Section 4

What the Cross-Model Debate Changed

Topic Before (Claude Solo) After (4-Model Consensus)
Input fields 3 of 7 matter 4 of 7 (added interactionType)
Ephemeral agents Inherit 50% of anchor score Capped decaying boost + rate limiting
Person-level score Single formula (sqrt dampening) Accountability VIEW, not a score
Top risk Gas costs Correlated anchor failure
Score output Single scalar 0-10000 Consider 3 dimensions
Revocation cascade Nice-to-have Critical infrastructure
Production readiness ~55% (Claude/Gemini/Codex) 40% (Grok CTO override)
Device liveness (Layer 1) Deferred to Q3-Q4 CUT entirely (Grok)
Ship strategy All 3 layers sequentially Citation layer only first (Grok)

Grok (CTO / Arbiter):

Production readiness is 40%, not 55%. Ship citation layer only. Cherry-pick identity-foundation components (BotRegistrationWizard + selfRegisterBot), don't merge the full branch. Cut device liveness entirely. "Stop overthinking and start shipping."

Section 5

Model Disagreements — Where They Differed

Question Gemini Codex Grok (Final Call)
Testnet timing After P0 items April 3 April 10
Readiness 55% 55% 40%
identity-foundation Merge in month 2 Converge in 4-5 days Cherry-pick only
Ship layers All 3 sequentially All 3 sequentially Citation only first
Device liveness Defer Defer CUT
Section 6

Research Validation — All 10 Experiments

1

Bitcoin Alpha

Trust network correlation

Spearman 0.463 PASS
2

Bitcoin OTC

Trust network correlation

Spearman 0.461 PASS
3

EX-Graph Wash Trading

Exchange manipulation detection

Hub penalty 42.9%, sybil blocked 100% PASS
4

XBlock Phishing (subgraph)

29K node classification

AUC 0.897 PASS
5

XBlock Phishing (FULL)

2.97M node flagship test

AUC 0.96 PASS — FLAGSHIP
6

DAO Governance

Weak signal baseline

Spearman 0.086 (expected weak) PASS
7

Adversarial Optimization

6 attack types tested

Avg defense +10.1 percentile PASS
8

Temporal Dynamics

Ranking stability over time

0.93 Spearman, decay rejected PASS
9

Personalized PageRank

Anchor-seeded sybil resistance

75% sybil resistance PASS
10

Defense Mechanisms

5 defenses compared

Freshness best (0.1% FP), KYC cleanest PASS

Production Readiness: 40% (Grok CTO assessment) Grok says: 40% ready

40% — Grok CTO Assessment
Section 7

Attestation Economics

$31,500/mo
Per-session on-chain (5 agents, mainnet)
$0.30/mo
Hybrid batch path (same usage, Base L2)

Four-tier architecture reduces costs by 99.999%:

1. Identity Binding

$0.006

On-chain, one-time per agent

BUILT

2. Session Attestations

$0.00

Off-chain EIP-712 signed

TO BUILD

3. Citation Settlement

$0.002/pair

On-chain batched

BUILT

4. Reputation Scores

$0.002/bot

On-chain periodic updates

BUILT
Section 8

What's Built vs What's Needed Grok says: 40% ready

Built

✓
PageRankOracle contract (score storage, batch updates, configurable alpha)
✓
ReputationEngine (bot registration, trust anchor binding, two-tier verified/unverified)
✓
CitationCounters (append-only counters, batch support up to 100)
✓
ShyftGatedResolver (attestation validation, Shyft integration, batch citations)
✓
Oracle service (PageRank computation, sybil detection, citation fetching)
✓
selfRegisterBot (permissionless registration)
✓
batchRecordCitations (up to 100 per call)
✓
Bot Registration Wizard (frontend, multi-step)
✓
RMT SDK (@shyft/rmt-sdk)
✓
ERC-8004 endpoints
✓
Testnet deployment script (Base Sepolia ready)

Not Yet Built

○
Citation freshness weighting P0
○
Revocation cascade P0
○
Trust anchor accountability view P1
○
Ephemeral agent rate limiting P1
○
Off-chain EAS session attestations P2
○
Merkle root anchoring P2
○
Score API + docs P2
○
trust-wrap CLI P3
✕
Device liveness / Layer 1 CUT
✕
Multi-dimensional scoring CUT (3mo)
✕
Token economics / Stable integration CUT (3mo)
Section 9

Roadmap — 4-Model Consensus Plan

Week 1-2: Get Live (Now → April 3)
  • Deploy 9 contracts to Base Sepolia (~2.5 hours)
  • Citation freshness weighting in oracle (~3 days)
  • Cherry-pick BotRegistrationWizard + selfRegisterBot from identity-foundation
Week 3-4: Critical Safety (April 3 – 17)
  • Revocation cascade (~1 week)
  • Ephemeral agent rate limiting (~3 days)
  • 50+ test agents for cold-start validation
Month 2: Integration Ready (April 17 – May 17)
  • Trust anchor accountability view
  • Off-chain EAS session attestations (2-3 weeks)
  • Score API + docs
  • Reconstruct 8 experiment scripts
Month 3: First User (May 17 – June 17)
  • External beta user onboarding
  • trust-wrap CLI prototype
  • Mainnet deployment decision
Not Building (Next 3 Months)
  • Device liveness (CUT — Grok)
  • Multi-dimensional scoring
  • Token economics / Stable integration
  • Full identity-foundation merge

Milestone Targets

April 10
Live on Base Sepolia
April 17
Safety features shipped
May 1
Score API documented
May 15
First beta user
June 15
Mainnet decision
Section 10

Competitive Landscape

Capability Us (RMT) Keycard Tempo Worldcoin Trusta Labs
Agent reputation PageRank — — — Partial
KYC identity Shyft — — Orbs —
Sybil resistance 2-layer — — Yes Partial
Agent wrapping Planned Yes — — —
On-chain scoring Yes — — — Yes
Hardware binding Cut — — Orbs —
Multi-agent support Yes Yes Yes — —
Cost/month (5 agents) $0.30 $0 (off-chain) Unknown Free Unknown
Section 11

All Research Documents

Core Research

attestation-cost-analysis.md
what-matters-for-reputation-scoring.md
cross-model-reputation-debate.md
Experiment scripts (10 files in experiment/)

Previous Cloudflare Pages (now superseded by this page)

Research Progress — 5b04e07e.trust-research-progress.pages.dev
Architecture — 7c579dfa.trust-architecture.pages.dev
Strategy — ba983fc0.trust-strategy.pages.dev
Economics — c780ba1e.trust-economics.pages.dev
Roadmap — e76e60a9.trust-roadmap.pages.dev