Tesseract Academy awarded Innovate UK Frontier AI Discovery funding for Tardygrada

Why agent guardrails need a different approach

Most current guardrails for AI agents check an agent's claims only after the output has been produced. They typically ask a second statistical model to judge the first one. That second model has its own failure modes, and those failures are difficult for anyone to audit. When such a guardrail approves or blocks an output, the decision rests on model behaviour that cannot be inspected step by step.

How Tardygrada works

Tardygrada intervenes before an agent's output is released. The runtime decomposes the output into individual logical statements and grounds each statement in a formal ontology, an explicit and machine-readable model of the entities and rules of a domain. Deterministic consistency checks then evaluate the grounded statements against that ontology. Because the checks are deterministic, the same input always produces the same verdict. There is no second LLM inside the verification loop, so the audit problem described above does not reappear inside Tardygrada itself. When the runtime rejects an output, it returns a reason that anyone can inspect. Each rejection can be traced to the specific statement that failed and the specific constraint it violated.

What Phase 1 will test

Phase 1 will test whether this approach holds up on real documents at production latency. The feasibility study will assess whether verification that happens before output, grounded in a formal ontology, can operate within the time constraints of a production system while processing realistic material.

Project team

Dr Stylianos Kampakis is the Principal Investigator for the project. Fabio Rovai is Co-Investigator and project manager.