Verifiability as a first principle
Most AI platforms treat output quality as a statistical property — "this model scores 92% on our eval." For legal use cases, that's not enough. You need to be able to point at a specific output for a specific document and say "here's why this is correct."
The three guarantees
1. Cryptographic provenance
Every output our system produces includes a signature linking it to:
- The model version and parameters
- The input document hash
- The system prompt used
- The temperature and other generation parameters
This signature can be independently verified using our public key.
2. Citation verification
Before any legal citation reaches a document, it goes through LegalVerify — a multi-source checker that queries CourtListener, Harvard CAP, and Google Scholar. If a citation doesn't exist in at least two of those sources, it's flagged.
3. Audit trails
Every interaction — every prompt, every output, every human override — is logged to an append-only store. The log itself is cryptographically chained: tampering with any entry invalidates all subsequent entries.
Building for audit
The audit system was the hardest to build. It needs to be:
- Append-only: No deletes, no updates, only new entries
- Performant: The log grows with every interaction
- Verifiable: Anyone should be able to detect tampering
We settled on a hash chain stored in SQLite, with periodic Merkle tree snapshots for efficient verification. It's not blockchain-level decentralization, but for a single-operator platform, it provides strong tamper evidence.