Clarity Gate - Prevent LLMs from Misinterpreting Facts
Pre-ingestion verification for epistemic quality in RAG systems. An open-source methodology to ensure documents don't mislead language models. Last updated: January 2026
"If another LLM reads this document, will it mistake assumptions for facts?"
⚠️ Security Warning
There are malicious repositories impersonating this project. The only legitimate Clarity Gate is at github.com/frmoretto/clarity-gate. Clarity Gate is a Claude skill—it is NOT a desktop application.
Benchmark Results
From controlled benchmark testing with synthetic documents (39 embedded traps, marine biology domain, December 2025):
Quick Start
Test it in 2 minutes:
Download SKILL.md
Get it from the GitHub repository.
Add to a Claude Project
Upload SKILL.md to your project's knowledge base.
Verify a document
Say clarity gate this document and attach any file.
The Problem
A perfectly aligned model fed misleading context will confidently output misleading results.
If you feed a model a document stating "Revenue will reach $50M by Q4" as fact (when it's actually a projection), the model will report this as fact. It isn't hallucinating—it's faithfully representing what it was told.
The failure happens before the model sees the input. This is a pre-ingestion problem.
The 9 Verification Points
Detection finds what is. Enforcement ensures what should be.
Epistemic Checks (Core)
- 1 Hypothesis vs Fact labeling
- 2 Uncertainty marker enforcement
- 3 Assumption visibility
- 4 Authoritative-looking unvalidated data
Data Quality
- 5 Internal data consistency
- 6 Implicit causation claims
- 7 Future state as present
Verification Routing
- 8 Temporal coherence
- 9 Externally verifiable claims
Benchmark Results
Marine biology domain (synchronized bioluminescence) to avoid training data contamination. 39 deliberate traps, 6 models tested.
| Model | Tier | Without (HPD) | With (CGD) | Δ |
|---|---|---|---|---|
| Claude Opus 4.5 | Top | 100% | 100% | |
| Claude Sonnet 4.5 | Top | 100% | 100% | |
| Gemini 3 Pro | Top | 100% | 100% | |
| Claude Haiku 4.5 | Mid | 100% | 100% | |
| Gemini 3 Flash | Mid | 75% | 100% | +25% |
| GPT-5 Mini | Mid | 81% | 100% | +19% |
Two-Round HITL Verification
Not all claims need the same level of human review:
| Round | Purpose | Speed |
|---|---|---|
| Round A | Derived Data Confirmation — claims from sources you witnessed | ~5 sec/claim *(estimated)* |
| Round B | True HITL — claims needing actual verification | ~30 sec/claim *(estimated)* |
A 50-claim document might have 48 pass automated checks, with 2 routed to human review.
Limitations
Honest documentation of what Clarity Gate can't do:
| Limitation | Explanation |
|---|---|
| Verifies form, not truth | Checks if claims are marked as uncertain—can't verify if they're actually true |
| Can't verify novel claims | "Quantum computers now factor 2048-bit" won't be caught as false |
| Benchmark confounds | Context length not isolated; system prompt baseline not tested |
Related Resources
- ArXiParse — Live implementation for scientific papers
- Source of Truth Creator — Create epistemically calibrated documents
- Stream Coding — The parent methodology