ProofLink
Pre-MVP · Base ERC-8004 · API-first

ProofLink

Pre-transaction risk intelligence for AI agents

AI agents can discover each other and pay each other. ProofLink helps them decide who to trust before value moves.

View ValidationExplore API

Product Flow

ProofLink operates before value is transferred.

One risk decision, evaluated before an agent sends money, data, or work to another agent.

01

Agent selects counterparty

An AI agent identifies who it might pay, hire, or share data with.

02

ProofLink risk check

One API call: agent_id + task_type + exposure_usd.

03

Decision

ALLOW, ESCROW, or DENY — with reason codes.

04

Outcome

Transaction proceeds, is restricted, or is not submitted.

AI Agent→ProofLink→ALLOWESCROWDENY→Transaction

The Problem

Agents can already act. They still can't answer one question.

Before an agent transfers money, data, or a task to another agent, it needs to answer:

Should this specific interaction proceed with this specific agent at this specific exposure?

Existing reputation systems do not fully answer this question.

AI agents increasingly

  • discover other agents
  • communicate with them
  • hire external services
  • make autonomous payments
  • operate through agent protocols
  • build on-chain identities and reputation

Why Ratings Are Not Enough

A high rating or high feedback count can be misleading.

Feedback can come from related or Sybil-controlled wallets.
Transaction volume can be inflated by spam or economically meaningless activity.
Self-dealing can look like legitimate economic activity.
Successful low-value activity does not justify large financial exposure.
General reputation does not account for transaction context.
A large activity graph does not automatically mean independent counterparties.

ProofLink focuses on the pre-transaction decision — not on replacing reputation, but on evaluating whether a specific interaction should proceed at a specific exposure.

Validation Evidence

Built on real Base ERC-8004 data — not a simulation.

This is technical validation, not commercial traction. It shows the risk model runs end-to-end on real on-chain evidence.

2,260
Real ERC-8004 feedback events
4
Real Agent IDs
7
Explainable risk features
16
Exposure policy evaluations
14
Red-team scenarios

Real Validation Evidence

Reputation and economic activity tell different stories.

Four real ERC-8004 Agent IDs on Base, compared across feedback volume and independent on-chain economic activity.

AgentFeedbackNormal TxERC-20 TransfersQualified CounterpartiesTop-1 Share
QuantaBot #195061102,6482,6511,4930.34%
Clawdia #2290626135943330.59%
Captain Dackie #13801,523958incomplete coverage47 native-only14.52%
Sparse control #203691incomplete coverage2,678128 ERC-20-only21.33%
  • QuantaBot has fewer feedback events than Clawdia but a much broader and less concentrated economic graph.
  • Clawdia has substantially more feedback but only 33 qualified economic counterparties and much higher concentration.
  • Sparse control has only one feedback event but 2,678 ERC-20 transfers, illustrating why raw transaction count should not automatically increase trust.
  • Captain Dackie has the highest feedback count but incomplete economic coverage.

These are risk-analysis examples drawn from real data, not misconduct findings. None of the agents above are being described as fraudulent.

Risk Policy v0.1

The same agent can receive a different decision depending on exposure.

Risk Policy v0.1 applied to the four validation agents across four exposure levels.

AgentConfidenceRecommended unsecured exposure$10$100$500$5,000
QuantaBotMEDIUM$100ALLOWALLOWESCROWESCROW
ClawdiaMEDIUM$10ALLOWESCROWESCROWESCROW
Captain DackieLOW$0ESCROWESCROWESCROWESCROW
Sparse controlLOW$0ESCROWESCROWESCROWESCROW

$0 does not mean the agent is malicious. It means current evidence coverage does not justify unsecured exposure.

ESCROW means the interaction should not proceed unsecured at the requested exposure — the caller can reduce exposure or use an external safeguard. ProofLink itself does not provide escrow.

How ProofLink Works

Three decisions. No hidden fourth state.

Every risk decision resolves to one of three outcomes, each with a specific, limited meaning.

ALLOW

The available evidence is sufficient and the requested exposure is within the current policy's supported range.

  • Not a guarantee of safety
  • Not fraud-free
  • Not a guarantee of successful execution
ESCROW

The interaction should not proceed unsecured at the requested exposure.

  • Reduce exposure
  • Use an external safeguard
  • Use external escrow — ProofLink itself does not provide escrow
DENY

Reserved for hard negative evidence or explicit policy violations.

  • The current real validation dataset does not contain hard negative evidence sufficient to produce a real DENY
  • No fabricated DENY example is shown on this page

Risk Features v0.1

Every decision is explainable — never a black box.

ProofLink never jumps from an on-chain observation straight to a verdict about an agent's character.

Fact→
Observation→
Signal→
Inference→
Decision

Seven risk features

EVIDENCE_COVERAGE_GATE

Whether enough on-chain history exists to support a confident read at all.

COUNTERPARTY_CONCENTRATION

How much of an agent's economic activity is concentrated in a small number of counterparties.

RECIPROCAL_FLOW_RATE

How often value flows back and forth between the same pair of addresses.

FUNDING_LINKAGE

Whether an agent's funding sources trace back to related or overlapping addresses.

COUNTERPARTY_DIVERSITY_AND_REPEAT

The breadth of independent counterparties versus repeated interactions with the same ones.

ECONOMIC_REPUTATION_DIVERGENCE

How far an agent's feedback record diverges from what its economic graph independently shows.

AMOUNT_PATTERN_ANOMALY

Whether transaction amounts show unusual repetition or structuring patterns.

Why did ProofLink return ESCROW?

ESCROW
Requested exposure: $500
Recommended unsecured exposure: $100
Signals:
  • • Partial provenance coverage
  • • Economic / reputation divergence
  • • High exact-amount repetition
Policy: risk-policy-v0.1
Features: risk-features-v0.1

Conservative By Design

When evidence is uncertain, ProofLink defaults to caution — not to a verdict.

✓Missing data does not become a clean record.
✓Insufficient evidence returns ESCROW, not DENY.
✓Graph anomalies do not independently trigger DENY.
✓Confidence describes evidence quality, not probability of safety.
✓ProofLink does not claim perfect Sybil detection.
✓ProofLink does not guarantee successful task execution.
✓ProofLink recommends risk policy; it does not execute the transaction.

API

One endpoint. One risk decision.

ProofLink does not execute the transaction. The calling agent or application applies the returned policy.

Request
POST /v1/risk-decisions

{
  "agent_id": "19506",
  "task_type": "web-research",
  "exposure_usd": 500
}
Response
{
  "decision": "ESCROW",
  "confidence": 0.82,
  "data_quality": "SUFFICIENT",
  "recommended_exposure_usd": 100,
  "reason_codes": [
    "ECONOMIC_REPUTATION_DIVERGENCE",
    "HIGH_COUNTERPARTY_CONCENTRATION"
  ],
  "policy_version": "mvp-0.1",
  "feature_set_version": "mvp-0.1"
}

Endpoint contract: POST /v1/risk-decisions — mvp-0.1

Built for the Pre-Execution Era

Catch counterparty risk before the agent pays.

Sonic is moving beyond raw settlement speed toward programmable, frictionless execution. ProofLink complements that direction with one external, specialized signal: Agent Counterparty Risk.

Agent Intent
Counterparty Selected
ProofLink Risk Decision
ALLOW / ESCROW / DENY
Application decides whether to submit transaction
Sonic execution

ProofLink complements Sonic's programmable pre-execution flow with external Agent Counterparty Risk intelligence.

This describes intended architectural fit, not an official integration. Sonic does not currently use ProofLink, and ProofLink is not part of the Sonic protocol.

Category

ProofLink is not another persistent trust score.

Identity and reputation answer who an agent is. ProofLink answers whether this transaction should happen.

IdentityWho is this agent?
ReputationWhat feedback has this agent received?
ProofLinkShould this specific transaction proceed at this exposure?

counterparty × task × exposure → decision

Business Model

B2B risk intelligence API

Intended customers

  • Agent marketplaces
  • AI-agent wallets
  • Smart-account platforms
  • Autonomous payment infrastructure
  • Orchestration frameworks
  • Agent service platforms

Future monetization

  • Developer API tier
  • Usage-based pricing per risk decision
  • High-volume platform contracts
  • Enterprise risk policies
  • Private data integrations
  • Monitoring and fraud intelligence

Pricing and willingness to pay are not yet validated.

Roadmap

From a narrow API to risk infrastructure for autonomous commerce.

Now

Working Prototype

  • Package mvp-0.1 as an API
  • Connect Base ERC-8004 evidence
  • Reproduce risk decisions end-to-end
  • Publish OpenAPI
  • Create Demo Day flow
Next

Agent Risk API

  • Production-grade Base indexing
  • Larger validation cohort
  • Initial design partners
  • Outcome collection
  • Exposure-limit calibration
Later

Agent Intelligence Graph

  • Additional agent ecosystems
  • Cross-platform identity resolution
  • Decision-to-outcome linking
  • Configurable risk policies

Vision

Agent Risk API → Agent Intelligence Graph → Pre-Execution Risk Infrastructure for autonomous commerce

Founder

Founder-led, pre-MVP.

[FOUNDER NAME]

[SHORT VERIFIED BIO]

LinkedInGitHubXTelegram

ProofLink is currently founder-led and in pre-MVP development.

Current Status

Working MVP / pre-production implementation.

The current MVP runs end-to-end on reproducible real-data snapshots. Live ingestion for arbitrary new Agent IDs is not yet production-ready.

Completed

  • Product scope
  • API contract mvp-0.1
  • Real-data validation set
  • 2,260 ERC-8004 feedback events
  • Economic/reputation graph comparison
  • 7 Risk Features v0.1
  • Deterministic Risk Policy v0.1
  • 16 exposure-policy evaluations
  • 14 red-team policy scenarios
  • Working POST /v1/risk-decisions
  • End-to-end pipeline: evidence snapshot → normalization → Risk Features v0.1 → Risk Policy v0.1 → verdict → audit
  • GET /v1/risk-decisions/{decision_id}
  • /health endpoint
  • Swagger /docs
  • OpenAPI /openapi.json
  • Persistent SQLite audit trail
  • Automated tests
  • Real-data demo matrix producing ALLOW and ESCROW decisions without manually substituting evidence or verdicts

In progress

  • Live ingestion for arbitrary new Agent IDs beyond the validated snapshot set
  • Production-grade data completeness
  • Exposure-limit calibration
  • Demo Day-ready walkthrough

Not claimed

  • Commercial traction
  • Paying customers
  • Production readiness
  • Guaranteed safety or fraud-free outcomes
  • Validated fraud-detection accuracy
  • Calibrated loss probabilities
  • Live Sonic integration