Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.litigationlabs.io/llms.txt

Use this file to discover all available pages before exploring further.

​
How LitigationLabs Works

LitigationLabs employs a multi-agent AI architecture to create realistic courtroom simulations. This page explains the underlying methodology and how different components interact to produce an authentic trial experience.

​
Multi-Agent Orchestration

Unlike simple chatbot interactions, LitigationLabs coordinates multiple AI agents that operate independently yet coherently within a shared courtroom context. Each agent has distinct objectives, knowledge boundaries, and behavioral parameters.

​
The Agent Ensemble

Four primary agents participate in each simulation:
AgentRoleObjective
Witness AgentResponds to examination questionsProvide testimony consistent with their profile while protecting unfavorable facts
Opposing Counsel Agent (OCA)Raises objections and conducts cross-examinationChallenge improper questions; elicit facts favorable to their side
Judge AgentRules on objections and maintains orderApply the Federal Rules of Evidence fairly and consistently
OrchestratorCoordinates agent interactionsManage turn order, context sharing, and session state

​
Agent Coordination

The orchestrator manages the flow of information between agents:
  1. You ask a question → The orchestrator routes it to the appropriate agents
  2. OCA evaluates → Opposing counsel decides whether to object
  3. If objection raised → Judge agent rules; you may respond
  4. Witness responds → Based on the question and any sustained objections
  5. Scoring evaluates → System checks if the answer matches any elicits
This coordination creates the natural rhythm of courtroom examination—question, potential objection, ruling, answer—without requiring manual orchestration from you.

​
Witness Behavior Model

Witness agents operate within defined behavioral parameters that simulate realistic testimony patterns.

​
Profile-Based Responses

Each witness has a profile containing:
  • Background information: Who they are and their relationship to the case
  • Knowledge boundaries: What they know and do not know
  • Demeanor traits: How they respond under pressure
  • Protected facts: Information they will resist revealing
The witness agent generates responses that remain consistent with this profile while adapting to the specific questions asked.

​
Credibility and Resistance

Witnesses do not simply recite facts on demand. The system models:
  • Initial resistance: Witnesses may deflect or provide incomplete answers initially
  • Progressive disclosure: Skillful questioning leads to more complete testimony
  • Credibility degradation: Contradictions or admissions affect the witness’s perceived reliability
The health bar visible during examination reflects this credibility state, providing visual feedback on your examination’s effectiveness.

​
Objection Handling

The objection system implements the Federal Rules of Evidence with intentional imperfection to create realistic training conditions.

​
OCA Objection Logic

Opposing counsel evaluates each question against multiple criteria:
  • Form objections: Leading, compound, argumentative, assumes facts not in evidence
  • Evidentiary objections: Hearsay, relevance, foundation, speculation
  • Privilege and competence: Attorney-client privilege, lack of personal knowledge
Importantly, OCA is configured to make mistakes approximately 25-40% of the time. This design choice reflects real courtroom conditions where opposing counsel sometimes objects incorrectly or fails to object when they should.

​
Judge Ruling Methodology

When an objection is raised, the judge agent:
  1. Analyzes the question against the cited grounds
  2. Considers any response or exception you raise
  3. Applies the relevant Federal Rules of Evidence
  4. Delivers a ruling with legal reasoning
Rulings include citations to specific FRE sections, helping you understand the evidentiary principles at play.

​
Thread Replies

Objection exchanges can involve multiple back-and-forth responses:
  1. OCA objects with grounds
  2. You may argue an exception or dispute the objection
  3. OCA responds to your argument
  4. Judge issues final ruling
Up to three exchanges are permitted before the judge intervenes to rule.

​
Scoring Methodology

The scoring system measures your effectiveness at extracting key facts (elicits) from witnesses.

​
Semantic Matching

Rather than requiring exact phrase matches, the system uses semantic similarity to evaluate whether a witness’s answer establishes a target fact: 
Target Elicit: "Witness admitted the contract was signed on January 15"
Witness Answer: "Yes, I recall signing the document in mid-January, around the 15th"
→ Semantic match detected → Points awarded

​
Matching Thresholds

The system applies tiered matching:
ThresholdSimilarityResult
Strong match60%+Direct or paraphrased admission
Standard match40%+Related concept established
Keyword fallback30%+Legacy matching for edge cases

​
Score Calculation

Each elicit carries a weight reflecting its importance to the case. Your score aggregates:
  • Points for each elicit successfully established
  • Bonus for effective objection handling
  • Deductions for missed objections (failing to object when appropriate)

​
Session Persistence

All session data persists automatically:
  • Transcript: Complete record of all questions, answers, objections, and rulings
  • Score state: Running tally of established elicits and points
  • Testimony state: Which witnesses have been examined and their current credibility
  • Phase tracking: Current position in the trial sequence
You can leave and return to a session at any time; your progress is preserved.

​
Context Management

Long examinations can exceed the context limits of underlying language models. The system handles this through:

​
Testimony Compression

As the transcript grows, earlier exchanges are compressed into summary form while preserving:
  • Established facts
  • Key objection rulings
  • Witness credibility state
  • Critical admissions

​
Agent Memory

Each agent maintains its own memory of the proceeding:
  • OCA memory: Tracks which questions have been asked to prevent repetition
  • Witness memory: Maintains consistency across the examination
  • Judge memory: Recalls prior rulings for consistency
This architecture enables extended examinations without degradation in response quality or coherence.

​
Customization and Configuration

Administrators can tune agent behavior through the configuration system:
  • OCA error rate: Adjust how often opposing counsel makes intentional mistakes
  • Objection types: Enable or disable specific objection categories
  • Scoring thresholds: Modify semantic matching sensitivity
  • Deduplication settings: Control how aggressively OCA avoids repetitive questions
These controls enable LitigationLabs to adapt to different training contexts, from beginner-friendly scenarios with fewer objections to advanced simulations with aggressive opposing counsel.