What question did this study set out to answer?

This work aims to develop a one-decorator approach for hallucination prevention in LLM calls.

April 25, 2026Open Access

Fathom v18 / styxx v3.9.0: The Trust Layer for LLMs — One-Decorator Hallucination Prevention

Key Points

This work aims to develop a one-decorator approach for hallucination prevention in LLM calls.
Introduced @trust decorator for verifying LLM outputs using cognometric checks.
Implemented fallback, retry, raise, and annotate halt policies to manage risks.
Conducted extensive testing with 31 new tests ensuring full functionality, with 562 tests passing.
Achieved cognometric verification with AUC of 0.9012 on HaluEval-QA.
All tested cases passed; no failures reported in the testing suite.
Improved ease of integration for LLM developers with a single-line installation approach.

Abstract

Headline: styxx v3. 9. 0 ships @trust — the one-decorator, one-line hallucination prevention layer for any LLM call. from styxx import trust followed by @trust on any LLM-calling function is the entire API. Every output is cognometrically verified (via styxx. guardrail. check (), AUC 0. 9012 on HaluEval-QA — see v17) before it reaches the caller. Risk above threshold is intercepted via fallback / retry / raise / annotate. Design principles: Zero config out of the box; ships with HaluEval-calibrated LR weights. Shape-preserving: handles OpenAI (choices0. message. content), Anthropic (content0. text), LangChain, dict, str — automatically. Prompt auto-detection from kwargs (prompt, question, query, messages) or positional args. Sync and async both supported (coroutine auto-detected). Four halt policies: fallback (safe text), retry (up to N), raise (TrustViolation), annotate (TrustResult). No residual-stream access required. No API keys. No external services. Tests: 31 new tests in tests/testₜrust. py. Full suite 562 pass, 1 skip, 0 fail. Installation: pip install styxx==3. 9. 0 The bet: this is the product that tries to change the LLM space forever. TLS for LLM cognition. Nothing crosses unseen. Repository: github. com/fathom-lab/styxx (tag v3. 9. 0). Package: pypi. org/project/styxx/3. 9. 0. Predecessor: v17 (10. 5281/zenodo. 19701750) introduced the four-signal guardrail at AUC 0. 9012. v18 packages it behind a single decorator so every LLM developer can install and use it in one line. License: CC-BY-4. 0 (this deposit, paper, demo) / MIT (styxx code).

Read Full Paperexternally

KI fragen

Bookmark

View Full Paper