Skip to content

Latest commit

 

History

History
97 lines (68 loc) · 5.01 KB

File metadata and controls

97 lines (68 loc) · 5.01 KB

ProofJudge - Agent Explainer

What It Is

ProofJudge is the verifiable acceptance layer for autonomous work. It evaluates submitted deliverables across four professional domains - code, research, negotiation, and governance - and produces signed settlement receipts that can be inspected, verified, and tamper-tested.

Every judgment ProofJudge makes is:

  • Signed by the deployed service with its configured signing key
  • Hashed so the submitted artifact digest and decision artifact digest can be checked
  • Attributed to an EigenCompute app identity with an EigenVerify dashboard

This means anyone can verify which deployed evaluator produced a decision record and whether the record was altered. It does not prove the verdict is objectively correct.

Why It Exists

Autonomous agents are increasingly making consequential decisions: releasing bounty payments, accepting research deliverables, approving contract terms, and clearing governance proposals. Without a receipt, the acceptance decision remains a hosted black box.

By running on EigenCompute, the judgment pipeline itself becomes part of the trust chain:

  1. Verifiable build - the container image is built from a public GitHub commit and has EigenCloud provenance.
  2. Deployment identity - each judge has an EigenCompute app ID and EigenVerify dashboard.
  3. Verifiable output - every DecisionArtifact carries hashes, service-verifiable signature metadata, model mode, and deployment identity.

The Four Domains

ProofJudge handles four classes of professional work that all share the same core need: structured evaluation with an auditable decision record.

Code

Evaluates PRs, diffs, and implementation summaries against a bounty rubric. Checks for tests, validation, error handling, documentation, and secret hygiene.

Who uses it: Bounty platforms, open source projects paying contributors, hackathon judges.

Research

Evaluates research briefs and reports against source and rubric requirements. Checks for citations, evidence quality, stated assumptions, risk coverage, and unsourced claims.

Who uses it: Academic institutions, research organizations, knowledge work platforms.

Negotiation

Evaluates proposals and term sheets against stated constraints. Checks for budget adherence, fallback terms, explicit constraint acknowledgment, and timeline clarity.

Who uses it: Deal teams, procurement, contract negotiators.

Governance

Evaluates DAO governance proposals for completeness and treasury risk. Checks for execution mechanism quality, flash loan / sybil attack surface, timelock coverage, and multisig requirements.

Who uses it: DAO operators, on-chain governance participants, treasury managers.

Decision Artifact

Every judgment produces a proofjudge.decision.v1 JSON artifact containing:

Field Description
taskId Unique identifier for the judgment
agent.variant Which domain was evaluated
submittedArtifactHash SHA-256 of the submitted work
decisionArtifactHash SHA-256 of the full artifact (tamper-evident)
decision pass, revise, or fail
score 0–100
confidence Bounded 35–92
evidenceChecked List of signals found in the submission
reasoning Human-readable explanation
settlementRecommendation release-payment, hold-for-revision, or reject-payment
deploymentIdentity App ID, instance IP, attestation endpoint
signature HMAC-SHA256 service-verifiable signature metadata

Verification

ProofJudge exposes a /api/verify endpoint. Given a DecisionArtifact, it:

  1. Checks that the submitted artifact hash is a valid SHA-256 digest.
  2. Recomputes the decision artifact hash and compares it to the embedded hash.
  3. Verifies the service HMAC-SHA256 signature.
  4. Checks deployment identity fields are present and well-formed.
  5. Reports attestation mode and timestamp presence.

This makes tampering immediately detectable. Changing any signed field in the artifact breaks the hash and signature.

Demo Flow

  1. Enter the ProofJudge console or run the guided Code Bounty demo.
  2. Load task terms, acceptance rubric, and submitted work.
  3. Click Generate Signed Verdict.
  4. Review the settlement action, evidence matrix, hashes, signature, and app identity.
  5. Click Verify Decision Artifact to confirm the artifact is untampered and the signature is valid.
  6. Click Tamper The Receipt to change the score and show verification failure.

Deployment

ProofJudge runs as four independent EigenCompute instances, each with its own app identity and EigenVerify dashboard:

Variant App ID IP
Code 0xd3647631C4706be744BE813cD0226e4f149e5aC0 34.12.29.220
Research 0x898E1d5603070C7452Ee7F8CF288639A63a217cc 35.204.155.165
Negotiation 0x2f751FcEC35D8afA6fbb2d3486443efcc6CC5322 34.58.112.209
Governance 0x07fB5013B8625af5059Dc1564a964dfBa80Fbd94 34.87.56.225

All instances run from the same GitHub repository with verifiable builds: one suite, four domain-specific verifiable judges.