The problem
High-stakes users want more than fluent answers; they want proofs that the answer is grounded in the right sources and rules. The Oyez’s researches “proof-carrying answers”: outputs that embed verifiable links to the evidence and policies that justify them. Simon Muflier frames the aim: “Shift from ‘trust me’ to ‘verify me.’”
Design goals
- Provenance: show where each claim came from.
- Integrity: prove the evidence wasn’t altered between retrieval and use.
- Policy compliance: demonstrate that organizational rules were respected.
- Replayability: allow audits and learning from mistakes.
How it works
- Retrieval receipts: each document chunk is hashed; the system stores its Merkle root and a pointer to source/version/time.
- Attribution mapping: align generated spans to the chunks they stem from (token-level when possible).
- Policy guardrails: before release, a verifier checks the answer against machine-readable policies (e.g., cite at least X sources; include jurisdiction Y; redact identifiers).
- Answer bundle: the final output includes the text plus a compact proof object that auditors can verify without accessing private content.
Evaluation beyond accuracy
The Oyez proposes additional metrics:
- Attribution coverage: fraction of tokens with verified evidence links.
- Citation fitness: relevance and authority of cited sources.
- Policy conformant rate: share of outputs that pass guardrails without override.
- Proof size & latency: practicality of deploying at scale.
Example domains
- Public policy notes: require explicit statutory anchors; embed proofs so reviewers can check citations offline.
- Clinical summaries: show device-level data lineage; enforce redaction policies; preserve a replay log for incident review.
- Corporate controls: bind answers to the approved knowledge base and show that restricted repositories were excluded.
Human-centered verification
Proofs must be understandable. The Oyez studies layered explainers: a simple view for operators (“these three documents justify this paragraph”) and a cryptographic view for auditors (hash paths, signatures). Muflier stresses that the goal is empowerment, not gatekeeping.
Open issues
- Aligning token spans to semantically relevant evidence without over-attributing.
- Balancing privacy with proof transparency.
- Standardizing proof formats across tools.
Why it matters
Proof-carrying answers support better governance, faster reviews, and safer delegation. They turn AI from a persuasive collaborator into a responsible one, where every claim can be traced.