0.86
F1 on real-world test
A drop-in safety layer that flags fabricated facts, marketing hype, and broken code in LLM output. Bring your own key. Verifiable. $0 marginal cost.
Models confidently invent dates, tickers, statistics, and APIs. Generic guardrails catch tone, not truth.
Wrong founding years, made-up tickers, off-by-billion population numbers — all stated with full confidence.
“Risk-free”, “100% uptime”, “guaranteed ROI”. Compliance and brand teams need these flagged at draft time.
Imports from removed Python stdlib, deprecated APIs, Python-2 syntax. Compiles in the model's head, not yours.
Each LLM output is decomposed into claims, routed to the right verifier, and returned with a calibrated severity.
Cross-checks factual claims against authoritative sources.
Catches stylistic tells of fabrication and hype.
Reads code blocks like a strict linter, plus model-specific traps.
Held-out evaluation on FEVER, marketing copy, code snippets, and an LLM-generated unbiased test set we couldn't tune for.
| Test set | F1 | Precision | Recall | AUROC |
|---|---|---|---|---|
| Unbiased real-world (LLM-gen, 41 cases) | 0.86 | 0.86 | 0.86 | 0.954 |
| FEVER factual | 0.83 | 0.93 | 0.75 | 0.852 |
| Marketing copy | 0.81 | 0.84 | 0.79 | 0.813 |
| Code snippets | 0.95 | 0.96 | 0.94 | 0.949 |
Public hallucination detectors (Galileo, semantic-entropy from Nature, generic NLI) report 0.75–0.85 F1 in the same regime. haluguard sits at the top of that band while remaining explainable and BYOK.
A 30% baseline hallucination rate drops to roughly 4% on flagged-only output — about a 7× reduction without throwing away clean responses.
Wrap any LLM call. haluguard returns a verdict, severity, and per-claim evidence trail.
# pip install haluguard from haluguard import Guard guard = Guard(api_key=YOUR_LLM_KEY) report = guard.check(prompt, llm_output) if report.severity_tier in ("clear", "suspicious"): log_for_review(report)
Bring your own LLM key. We charge for the verification layer, not the LLM. Cancel anytime.
Need more volume or on-prem deploy? Email us.