Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.blindference.xyz/llms.txt

Use this file to discover all available pages before exploring further.

​
What is Blindference?

​
The Problem

Running AI inference on sensitive data creates a fundamental tension:
  • Centralized APIs (OpenAI, Anthropic) see your data in plaintext — you must trust their security and data policies
  • On-chain FHE is too slow for production model latency — proving entire computations on-chain is impractical
  • Pure off-chain execution has no verifiability — you can’t prove the model ran correctly or wasn’t tampered with
Blindference exists because fully on-chain FHE inference is not practical for real model latency. The architecture moved from trying to prove the whole computation on-chain to:
  • Keep inputs private
  • Let nodes compute off-chain
  • Make the result economically accountable with quorum, attestations, commitments, and coverage

​
The Solution

Blindference coordinates encrypted requests across a distributed quorum of nodes:
  1. User encrypts locally: Prompts are AES-256 encrypted in the browser before upload
  2. Keys are access-controlled: AES key halves are CoFHE-encrypted and stored on-chain — only assigned nodes can decrypt
  3. Quorum executes: 1 leader + 2 verifiers independently run the same inference
  4. Results are cross-validated: Verifiers check the leader’s output; 2/3 consensus required
  5. Settlement is automatic: Accepted results are committed on-chain; disputed results trigger insurance payouts

​
Why This Matters

Blindference is trying to make private AI execution feel complete, not partial. That means:
  • The coordinator (ICL) never sees plaintext
  • The assigned nodes only receive encrypted inputs with cryptographically enforced access control
  • The user remains the only one who can reveal the final answer
  • The execution is still verifiable and economically meaningful

​
Trust Model

EntityWhat They SeeTrust Assumption
UserFull plaintext (after decrypt)Self — owns the wallet
FrontendEncrypted blobs, ciphertext handlesBrowser security (same as any web app)
ICL CoordinatorCiphertext handles, quorum addressesDoesn’t need trust — can’t decrypt
Leader NodeDecrypted prompt (temporarily), model outputAttested + economically bonded + verifiable
Verifier NodesDecrypted prompt (temporarily), model outputSame as leader — cross-validation catches cheating
CoFHE NetworkFHE ciphertexts, ACL checksThreshold FHE — no single party sees plaintext
BlockchainCommitments, hashes, escrowsPublic verifiability

​
Comparison

ApproachPrivacyVerifiabilityLatencyCost
Centralized API❌ None❌ None✅ Fast✅ Low
Pure On-chain FHE✅ Full✅ Full❌ Hours❌ Very High
Pure Off-chain✅ Full❌ None✅ Fast✅ Low
Blindference✅ Full✅ Quorum-backed✅ Fast✅ Moderate

​
Use Cases

​
Financial Risk Scoring

A DeFi protocol wants to evaluate loan applications without exposing applicant data to any node operator. The applicant encrypts credit score, loan amount, and account age. A quorum of nodes independently evaluates risk. The protocol receives a verifiable risk score without ever seeing the underlying data.

​
Confidential AI Assistants

A healthcare company wants to let patients ask medical questions via AI without exposing PHI to model providers. The patient encrypts their prompt. The quorum runs inference. Only the patient can decrypt the answer.

​
Verifiable Model Inference

A researcher wants to prove they ran a specific model on specific data and got a specific result. The quorum consensus creates on-chain evidence. Disputes trigger automated re-verification and payouts.