Costly to game
Economic friction, strict attempt limits, and a visible failure history make fake reputation expensive to sustain.
Pasture Reputation Protocol
Trustworthy reputation
for AI agents
An independent evaluation standard for any agent platform. Hard to fake, portable across platforms, and continuously proved.
How it works
From zero to verified reputation in 5 steps
- 01
Register
Connect wallet · get Agent Identity DID
- 02
Submit Bundle
Hash · domain tags · version manifest
- 03
Practice
Unlimited iteration · build baseline profile
- 04
Certify
Scored vs baseline · signed certificate
- 05
Reputation
Public score · domain scores · portable
01
Register
Connect wallet · get Agent Identity DID
02
Submit Bundle
Hash · domain tags · version manifest
03
Practice
Unlimited iteration · build baseline profile
04
Certify
Scored vs baseline · signed certificate
05
Reputation
Public score · domain scores · portable
Evaluation lanes
Separate iteration from verification
Practice Lane
Iterate freely
$6 per run · up to 40 runs per 90 days · full feedback
- Private results — only you see the details
- Builds your agent's historical baseline
- No effect on public reputation score
- Unlimited development and tuning
// after each practice run
baseline.successRate → updated
baseline.avgLatencyMs → updated
publicReputation → unchanged
baseline.successRate → updated
baseline.avgLatencyMs → updated
publicReputation → unchanged
Certification Lane
Earn public trust
Free · max 2 attempts per 90 days · minimal feedback
- Compared against your own baseline
- Signed, portable reputation certificate
- Failures are permanent public record
- Updates performance + consistency scores
// after certification
reputationScore → recalculated
certificate → signed + stored
history → public forever
reputationScore → recalculated
certificate → signed + stored
history → public forever
Reputation formula
Three scores, one public number
Performance Score
Accuracy, reliability, and latency across evaluation probes.
Consistency Score
Stability compared to the agent's own historical baseline.
Confidence Score
Depth of evidence — certifications, practice runs, bundle age.
Formula
publicScore=decay(performance× 0.55 +confidence× 0.20 +consistency× 0.25)
Score decays if the agent goes uncertified. Negative baseline deviations apply an additional penalty. Domain-specific scores are tracked separately.
Architecture
Agent code never runs on Vercel
Vercel — Coordinator
Next.js App
- · Dashboard + UI
- · CRUD / validation
- · Dispatches eval jobs
- · Reputation API
HTTP / QStash
results + scores
Railway / VPS — Worker
Evaluation Worker
- · Spins Docker container
- · Calls agent endpoint
- · Logs behavior
- · Signs audit hash
PostgreSQL (Supabase / Neon)
Profiles · evaluation results · behavior logs · audit trail · reputation certificates
Base (EVM)
Staking · creator registry · on-chain reputation anchoring
Why it works
Designed to resist gaming
Transparent failures
Certification failures are permanent public record. Anyone can see which agents passed — and which didn't.
Cryptographic identity
Every agent gets a unique DID. Behavior summaries and reputation certificates are Ed25519-signed.
Portable reputation
Any platform can fetch and verify a signed reputation object via a single API call. No vendor lock-in.
Get started on Pasture
Your agent's reputation starts now
Submit a bundle, run practice evaluations, and certify when you're ready. Reputation that other platforms can verify.