YM/
← Selected work
AI Reliability and ProvenanceIn Progress

ProofLayer

A tamper-evident verification and provenance layer for AI agents that decomposes output into atomic claims and produces signed, hash-chained verification receipts.

Problem and threat model

Generated AI output can obscure unsupported, contradicted, stale, or altered claims. ProofLayer targets those integrity and provenance gaps.

Claim lifecycle

Output is decomposed into atomic claims, evidence is retrieved from SEC EDGAR filings, then support or contradiction is evaluated before receipt creation.

Receipt design

Signed, hash-chained receipts record verification context so downstream users can inspect evidence relationships and staleness.

Architecture and API contracts

The planned system includes a FastAPI HTTP API, CLI, Python SDK, background verification workers, PostgreSQL persistence, and LLM/NLI pipelines.

Testing and limitations

Testing centers on claim decomposition, evidence retrieval, evaluation, receipt integrity, and staleness behavior. It remains in progress; quality depends on retrieval coverage and evaluation reliability.

Roadmap

Continue implementation and validation of the claim pipeline, retrieval and evaluation behavior, worker execution, and receipt integrity.

Personal ownership

Yash owns the concept and implementation direction for the verification and provenance layer.