Clarity Gate Prior Art

January 27, 2026 · View on GitHub

Version: 1.2
Last Updated: 2026-01-27

Overview

Clarity Gate builds on proven patterns. This document maps the landscape of existing systems and identifies the specific gap Clarity Gate addresses.

Key principle: We're not claiming to invent pre-ingestion gates. We're applying proven patterns to epistemic quality, open-source.

Verification note: Tool landscape validated via Perplexity Deep Research (January 2026). See changelog for details.

The Gap Matrix

Stage	Privacy/Security	Accuracy/Compliance	Epistemic Quality
Pre-ingestion	✅ Protecto.ai, OWASP	✅ Adlib, Pharma QMS	❌ Gap
Detection	—	—	✅ UnScientify, HedgeHunter (academic)
Post-retrieval	—	—	✅ RAGAS, TruLens, LOKI, RagChecker
Runtime	—	—	✅ Self-RAG

The gap: To the best of our knowledge, no open-source system enforces epistemic quality at pre-ingestion.

Enterprise Pre-Ingestion Gates (Proprietary)

Adlib Software

Product: Transform 2025.2 (released November 18, 2025)

Capabilities:

Multi-LLM voting for accuracy verification
Accuracy Score with configurable thresholds
Exception rate reduction: 40-60% (claimed)
Cycle time acceleration: 30-50% (claimed)

Customers: Pfizer, Swiss Re, IAG, JP Morgan (Fortune 500)

Certifications: SOC 2 Type II, HIPAA (October 2025)

Focus: Accuracy and compliance, not epistemic quality

Reference: adlibsoftware.com [CHECK BEFORE CITING - features may change]

Pharmaceutical QMS

Systems: SimplerQMS, Dot Compliance, Picomto, Ideagen

Regulatory basis: FDA 21 CFR Part 11 (established 1997)

Core requirements:

Electronic signatures with name, date/time, meaning
Audit trails tracking who did what when
Periodic re-validation schedules
"If it's not documented, it didn't happen"

Maturity: 20+ years of regulatory enforcement

Focus: Compliance and traceability, not epistemic quality

What Enterprise Gates Catch vs. Don't Catch

What They Catch	What They Don't Catch
"Revenue was $47M" when records show $49M	"Revenue will be $50M" stated as fact
Missing required signature	"Our approach outperforms X" with no evidence
Wrong date format	"Users prefer Y" with no methodology
Conflicting dates across documents	Unmarked projections

The distinction: Accuracy (does it match the source?) vs. Epistemic quality (is it properly qualified?)

Epistemic Detection Tools (Open-Source)

UnScientify

Paper: Ningrum et al., 2023 (arXiv:2307.14236)

Capabilities:

Detects multiple scientific uncertainty pattern groups
Rule-based + ML hybrid approach
Accuracy: ~0.8 on research articles (approximate; see paper for exact metrics)

What it does: Identifies uncertainty markers that ARE present in text

What it doesn't do: Enforce markers that SHOULD be present but aren't

Status: Open-source, academic

HedgeHunter (2010)

Paper: Clausen, CoNLL-2010 Shared Task (W10-3017)

Author: David Clausen, Stanford University

Capabilities:

Two-stage hedge detection: (1) Hedge cue detection, (2) Uncertainty classification
Training data: Wikipedia + biomedical abstracts
High precision hedge detection for downstream IE tasks

What it does: Token-level hedge cue detection and scope classification

What it doesn't do: Determine if hedging is missing where needed

Status: Academic research tool, not maintained, no downloadable package

Note on "HedgeHog" naming confusion: Multiple unrelated projects use similar names:

HedgeHog (wearable sensor platform) — motion tracking hardware
hedgehog-qa (Haskell testing library) — property-based testing
HedgePeer (2022 dataset) — benchmark corpus, not a tool
Various DeFi/trading bots

None of these are NLP uncertainty detection tools. HedgeHunter (2010) is the primary academic system for hedge detection.

FactBank

Type: Veridicality corpus

Purpose: Training data for factuality classification

What it provides: Annotated examples of factual vs. uncertain claims

What it doesn't do: Runtime verification

Status: Open-source, academic resource

BioScope

Paper: Vincze et al., 2008

Type: Biomedical uncertainty corpus

Purpose: Training data for uncertainty detection in scientific text

Status: Open-source, academic resource

HedgePeer Dataset (2022)

Paper: Ghosal et al., ACM SIGMOD 2022

Type: Hedge detection benchmark corpus

Size: 5x larger than previous hedge detection datasets

Purpose: Enable domain adaptation across scientific domains

Status: Dataset for training/evaluation, not a deployable tool (22+ citations)

Detection vs. Enforcement

Tool Type	Question Answered
Detection (UnScientify, HedgeHunter)	"Is uncertainty expressed?"
Enforcement (Clarity Gate)	"Should uncertainty be expressed but isn't?"

This distinction is the core of Clarity Gate's contribution.

Modern Fact-Checking Tools (2024-2025)

These tools exist but operate post-ingestion or post-retrieval, not at the pre-ingestion gate:

LOKI (COLING 2025)

Focus: Fact verification with checkworthiness assessment

Stage: Post-retrieval evaluation

Limitation: Evaluates claims after they're in the system

FIRE (arXiv:2411.00784)

Focus: Iterative fact-checking framework

Stage: Post-retrieval, iterative refinement

Limitation: Operates on generated outputs, not source documents

RagChecker (arXiv:2408.08067)

Focus: Diagnostic framework for RAG pipeline issues

Stage: Post-retrieval evaluation

Limitation: Diagnoses problems after retrieval, doesn't prevent them

Veracity (arXiv:2506.15794)

Focus: Open-source fact-checking system

Stage: Claim verification against external sources

Limitation: Requires external evidence corpus, post-hoc verification

Gap Confirmation

These modern tools validate the gap Clarity Gate addresses:

Tool	Pre-ingestion?	Epistemic enforcement?
LOKI	❌	❌ (fact-checking)
FIRE	❌	❌ (fact-checking)
RagChecker	❌	❌ (diagnostic)
Veracity	❌	❌ (fact-checking)
Clarity Gate	✅	✅

Automated Fact-Checking (Academic)

FEVER

Paper: Thorne et al., 2018

Pipeline:

Claim extraction
Evidence retrieval
Verification against evidence

Dataset: 185,000 claims with Wikipedia evidence

Focus: Verifying claims against external knowledge

Limitation: Requires pre-existing evidence corpus

Reference: fever.ai [STABLE - academic]

ClaimBuster

Capabilities:

Claim extraction from news and documents
Claim-worthiness scoring
Integration with fact-checking workflows

Focus: Identifying claims worth checking, not epistemic quality

Reference: claimbuster.org [CHECK BEFORE CITING]

Post-Retrieval & Runtime Systems

Self-RAG

Paper: Asai et al., 2023 (arXiv:2310.11511)

Innovation: Reflection tokens (ISREL, ISSUP, ISUSE)

ISREL: Is retrieved content relevant?
ISSUP: Is generation supported by retrieval?
ISUSE: Is response useful?

Stage: Runtime (after retrieval, during generation)

Limitation: Doesn't prevent problematic content from entering knowledge base

RAGAS

Type: RAG evaluation framework

Metrics:

Faithfulness
Answer relevancy
Context precision
Context recall

Stage: Post-retrieval evaluation

Limitation: Evaluates after the fact, doesn't gate ingestion

Reference: github.com/explodinggradients/ragas [CHECK BEFORE CITING]

TruLens

Type: LLM application evaluation

Capabilities:

Groundedness scoring
Answer relevance
Context relevance

Stage: Post-retrieval evaluation

Limitation: Same as RAGAS -- evaluation, not prevention

Knowledge Engineering Frameworks

Semantica

URL: github.com/Hawksight-AI/semantica
Focus: Semantic layer, knowledge graph construction
Stage: Post-extraction

Capabilities:

Entity extraction and resolution
Multi-source conflict resolution
Knowledge graph construction
Credibility-weighted voting for conflicting values

Conflict handling example:

Source A: employee.name = "John Doe"
Source B: employee.name = "Jonathan Doe"
Resolution: Credibility-weighted voting

Relationship to Clarity Gate:

Semantica addresses conflicts between extracted entities. Clarity Gate addresses epistemic quality within source documents:

Tool	Question	Example
Semantica	"Which value is correct?"	John vs Jonathan
Clarity Gate	"Is this claim properly qualified?"	"Revenue will be $50M" (unmarked projection)

Integration: Clarity Gate runs before Semantica ingestion:

Raw Docs → Clarity Gate → CGD → Semantica → Knowledge Graph

Status: Open-source, active development (422+ stars as of Jan 2026)

Privacy & Security Pre-Ingestion

Protecto.ai

Focus: PII/PHI detection and redaction

Stage: Pre-ingestion

Relevance: Proves pre-ingestion gates work; different focus

OWASP LLM Security

Focus: Prompt injection, data leakage

Stage: Various

Relevance: Security-focused, not epistemic

Validation & Guardrails

Guardrails AI

Focus: Output validation (schema, format)

Stage: Post-generation

Relevance: Structure validation, not epistemic quality

NeMo Guardrails

Focus: Dialog safety, topic boundaries

Stage: Runtime

Relevance: Behavioral guardrails, not epistemic verification

The Safety Stack Position

Layer 4: Human Strategic Oversight
Layer 3: AI Behavior Verification (PETRI, BLOOM, red-teaming)
Layer 2: Input/Context Verification  <-- Clarity Gate
Layer 1: Deterministic Boundaries (rate limits, guardrails)
Layer 0: AI Execution

Layer 3 Tools:

PETRI — Anthropic's open-source auditing tool for exploring model behaviors via multi-turn conversations
BLOOM — Anthropic's automated behavioral evaluation framework for frontier models

Key insight: A perfectly aligned model (Layer 3) can confidently produce unsafe outputs from unsafe context (Layer 2). Alignment doesn't inoculate against misleading information.

Summary: What's New

Component	Status
Pre-ingestion gate pattern	✅ Proven (Adlib, pharma QMS)
Epistemic detection	✅ Proven (UnScientify, HedgeHunter — academic only)
Fact-checking pipelines	✅ Proven (FEVER, ClaimBuster, LOKI, FIRE)
Post-retrieval evaluation	✅ Proven (RAGAS, TruLens, RagChecker, Self-RAG)
Pre-ingestion epistemic enforcement	❌ Gap
Open-source accessibility	❌ Gap

Clarity Gate contribution: Applying proven gate patterns to epistemic quality, open-source.

References

Enterprise (Volatile - verify before citing)

Adlib Software: adlibsoftware.com
SimplerQMS: simplerqms.com
Dot Compliance: dotcompliance.com

Academic (Stable)

UnScientify: arXiv:2307.14236 (Ningrum et al., 2023)
HedgeHunter: CoNLL-2010 W10-3017 (Clausen, 2010)
HedgePeer: ACM SIGMOD 2022 (Ghosal et al., 2022)
FEVER: fever.ai (Thorne et al., 2018)
Self-RAG: arXiv:2310.11511 (Asai et al., 2023)
BioScope: Vincze et al., 2008
FactBank: Sauri & Pustejovsky, 2009
LOKI: COLING 2025
FIRE: arXiv:2411.00784
RagChecker: arXiv:2408.08067
Veracity: arXiv:2506.15794

Standards (Stable)

FDA 21 CFR Part 11 (established 1997)
ISO/IEC 5259-1:2024 and 5259-2:2024 (AI data quality for analytics and ML)

Version History

Version	Date	Changes
1.0	2025-12-21	Initial prior art landscape
1.1	2026-01-05	HedgeHog → HedgeHunter (corrected). Added modern fact-checking tools (LOKI, FIRE, RagChecker, Veracity). Added HedgePeer dataset.
1.2	2026-01-27	Added Semantica (Knowledge Engineering Frameworks section) per RFC-003.