Skip to main content

ThirdLaw Glossary

This glossary provides definitions for key terms and concepts used throughout the ThirdLaw documentation.

Collection Concepts​

Exchange​

Any form of communication or data transfer between an LLM and an application or user. This is an umbrella term representing Events, Transactions, or Sessions.

Event​

A single, atomic interaction between an application and an LLM or agent, such as a Prompt or Response. Each event has a unique ID and includes metadata for context.

Transaction​

A cohesive unit of work initiated by a single goal assignment or user Prompt. Contains one prompt event and all related response events needed to achieve the goal. A transaction is complete when the agent or LLM produces a final output for the original goal or prompt.

Session​

A sequence of one or more transactions representing an extended conversation or interaction between a user (or application) and the LLM. Sessions are time-ordered and may optionally include session-specific context shared across transactions.

Context​

The cumulative information from previous events, transactions, or sessions used by an LLM to generate coherent and relevant responses in ongoing interactions.

Prompt​

The input data or query provided to an LLM to trigger a response.

Response​

The output generated by the LLM in reaction to a prompt.

Evaluation Framework​

Evaluation​

A self-contained logic module that answers a specific analytical question using a designated Analytic Engine. Evaluations are independent and reusable across different Laws. Each Evaluation consists of:

  • Configuration parameters
  • Analysis logic
  • Evaluation Finding

Finding​

The structured output produced by an Evaluator which contains the analysis results. A Finding typically includes a boolean decision, numeric score, or other structured data that can be used by Laws to determine compliance.

Law​

A policy that defines what should or should not happen within a given Scope. Laws serve as rules but do not execute logic directly. A Law is enacted by associating it with at least one Law Condition.

Law Condition​

The logical construct that determines whether a Law is violated. It contains specific instructions for evaluating an event, transaction, or session.

Law Result​

The result of a Law Condition, determining whether a Law has been violated.

Violation​

A confirmed breach of a Law as determined by a Law Condition. When a Violation occurs, it typically triggers one or more Interventions based on the severity and nature of the breach.

Intervention​

Actions triggered in response to Law Results. May include block, alert, warn, redact, rate-limit, human-in-the-loop review, regenerate, watermark, mask, etc.

ThirdLaw VPC​

The Virtual Private Cloud (VPC) where the ThirdLaw software is located and operates. All customer data is designed to stay inside the ThirdLaw VPC except when specifically allowed by customer configurations. This isolated network environment helps ensure data security and regulatory compliance by maintaining strict control over data boundaries.

Dataset​

A collection of Exchanges that can be used to build a Semantic Evaluator or a verification Benchmark.

Benchmark​

A Dataset used to Assess the performance of Evaluators

Evaluation Types​

Semantic-based Evaluators​

Using vector-based search to enable similarity analysis and comparisons between objects.

Search-based Evaluators​

Using pattern recognition including keyword or specific placeholder patterns.

Foundational Evaluators​

Uses foundation models to perform complex analysis requiring reasoning capabilities.

Validation Evaluators​

Using internally hosted transformer models optimized for specific detection tasks.

Law Properties​

Scope​

Defines where the Law applies—e.g., specific users, applications, LLM models, agents, or business contexts.

Frequency​

Determines when a Law Condition is evaluated, including:

  • Always on: Evaluate immediately with latency constraints
  • On Arrival: Evaluate immediately with fewer latency constraints
  • Batch frequency: Process at scheduled intervals
  • Sampling: Process only a percentage of events or transactions

Misc. Definitions​

LLM​

A Large Language Model (LLM) is an artificial intelligence system trained on vast amounts of text to understand and generate human language. It works by recognizing patterns in data and using those patterns to predict what words should come next in a sequence. Generative LLMs can write essays, answer questions, summarize content, translate languages, and even create code based on text instructions. Multi-modal LLMs can process and generate content across different types of information beyond just text, such as images, audio, and video.