Tessera

Enterprise AI agents fail silently — a grounded-sounding answer with no source behind it, an action taken before anyone reviewed it, a faithfulness score nobody can compute after the fact. Tessera inserts itself into that gap. The core workflow is claim-level: every statement an agent would make is decomposed into discrete claims, each of which must trace back to an exact source record. What cannot be verified against evidence is refused, not approximated. Actions are drafted from those verified claims, previewed in full, and withheld from execution until a human approves.

The differentiating design choice is determinism at the refusal layer. Most retrieval-augmented approaches soften low-confidence answers with hedging language; Tessera, per the repo README, refuses outright when the evidence threshold is not met. That makes auditability a structural property rather than a prompt engineering goal. Faithfulness scores are computed quantitatively, so teams can measure drift over time rather than spot-check outputs manually.

Tessera fits cleanly into environments running MCP-based agent architectures and heterogeneous data sources where traceability across source types matters more than response coverage. It is self-hosted, ships with a Dockerfile and pyproject.toml, and exposes an API — which means it drops into existing pipelines rather than requiring a platform migration. The friction appears when agents need to synthesize answers that no single source record fully supports: the refusal mechanism that makes Tessera trustworthy in compliance contexts becomes a blocker in exploratory or open-domain tasks. Teams in those scenarios either restructure their data to produce discrete, citable claims or move to a less strict grounding layer.

A live Hugging Face Space demo is available for evaluation. The GitHub repository includes an eval directory, suggesting the scoring mechanism is testable against your own data before production deployment.