-
Notifications
You must be signed in to change notification settings - Fork 0
POSITIONING
This document explains what PATAS is, what it is not, and how it fits into existing Trust & Safety and Abuse Detection stacks.
PATAS is a pattern discovery and transparent rule engine for Trust & Safety teams. It runs on historical or batch logs of messages and events, automatically discovering high-precision spam patterns and emitting transparent SQL-like rule definitions with honest metrics (precision, recall, false-positive rates). PATAS is designed to be self-hosted and on-prem friendly, with a developer-first approach that provides both API and CLI interfaces for integration into existing workflows.
Unlike black-box machine learning models, PATAS generates human-readable rules that can be reviewed, audited, and tuned by safety engineers. It focuses on finding stable, interpretable patterns in abuse data and providing clear signals that can be integrated into your existing enforcement systems.
-
Not a moderation console or case-management UI — PATAS does not provide a user interface for reviewing individual cases or managing moderation workflows.
-
Not a black-box ML model or generic LLM moderation API — PATAS generates transparent, explainable rules rather than opaque model predictions.
-
Not a full-stack Trust & Safety platform — PATAS does not include user management, content storage, notification systems, or other platform-level features.
-
Not a replacement for your existing enforcement or decision engine — PATAS provides signals and rules that complement your existing systems, not a complete enforcement solution.
-
Not a real-time classification service — PATAS is designed for batch analysis and pattern discovery, not for real-time message classification at scale.
Teams typically address pattern discovery and rule generation through one of several approaches:
Safety engineers write SQL queries or domain-specific language (DSL) rules manually based on domain expertise and ad-hoc analysis. This approach is highly transparent and precise, but requires significant manual effort and domain expertise. As spam tactics evolve, rules must be continuously updated by hand.
End-to-end moderation platforms provide comprehensive solutions including case management, user interfaces, and rule engines. These platforms are often SaaS-based and can be heavyweight, requiring significant integration effort and ongoing subscription costs. They may include pattern discovery features, but these are typically bundled with the full platform.
Machine learning models or LLM-based moderation APIs can be integrated quickly and provide fast classification. However, these systems are often black-box, making it difficult to understand why a decision was made, audit for compliance, or tune for specific precision/recall trade-offs.
Some teams build custom pattern discovery pipelines using data processing frameworks (Kafka, Flink, Spark) and custom rule engines. This provides full control and customization, but requires significant engineering effort to build, maintain, and scale.
The trade-offs: These approaches work, but they typically require either heavy manual work, lack transparency, or involve significant cost and complexity. PATAS aims to provide a middle ground: automated pattern discovery with transparent, auditable outputs.
PATAS is designed as a thin "pattern brain" that complements your existing Trust & Safety infrastructure:
-
Ingestion: PATAS ingests logs from your existing systems (abuse reports, message logs, event streams).
-
Pattern Discovery: It automatically discovers candidate patterns and rules, computing metrics (precision, recall, ham hit rate) for each.
-
Rule Export: PATAS exports rules and signals in standard formats (SQL, JSON) that can be plugged into:
- Your existing ML classifiers (as features or pre-filters)
- Your existing rule engines (as additional rules)
- Your existing enforcement pipelines (as signals)
-
Integration: The rules and metrics can be reviewed, approved, and integrated into your decision-making systems.
PATAS is designed to complement, not replace:
- Your ML / LLM classifiers (PATAS rules can be used as features or pre-filters)
- Your moderation tools (PATAS provides signals that feed into your moderation workflows)
- Your Trust & Safety dashboards (PATAS metrics can be integrated into your existing dashboards)
Think of PATAS as a specialized tool that enhances your existing stack by providing transparent, high-quality pattern discovery and rule generation.
Transparency and explainability are increasingly important for Trust & Safety systems:
-
Auditability and compliance: Regulations in many jurisdictions require explainable decisions. Transparent rules make it clear why content was flagged or blocked.
-
Internal review: Safety engineers and legal teams need to understand and validate the logic behind automated decisions. Human-readable rules enable thorough review and approval processes.
-
Tunable safety profiles: PATAS provides safety profiles (Conservative / Balanced / Aggressive) with clear precision/recall trade-offs, allowing teams to choose the right balance for their use case and adjust as needed.
-
Debugging and improvement: When false positives or false negatives occur, transparent rules make it easy to identify the cause and refine the logic.
PATAS offers three safety profiles that teams can choose based on their precision/recall requirements:
-
Conservative: High precision (low false-positive rate), suitable for automatic enforcement actions. Recommended for production use.
-
Balanced: Moderate precision and recall, suitable for review queues and signals. Use for investigation and manual review.
-
Aggressive: Higher recall but lower precision, suitable for research and feature extraction. Not recommended for automatic bans.
Each profile includes clear metrics and thresholds, allowing teams to make informed decisions about which patterns to activate.
PATAS is designed to work on-premises or in your own infrastructure. This means:
-
Data privacy: Sensitive logs never have to leave your company's infrastructure.
-
Compliance: You maintain full control over data processing and storage.
-
Integration: You can integrate PATAS with your existing on-prem systems without external dependencies.
-
Cost control: No per-message or per-API-call fees; you control the infrastructure costs.
PATAS can run entirely offline, with optional LLM integration for pattern refinement (which can also be run on-prem if you have your own LLM infrastructure).
For technical details, see: