Index/ AESOP/ Custom Rubric Evaluation
PRAXIS v1.4
THESIS v1.0

AESOP Transformation OS — Custom Rubric Evaluation

Generate a tailored evaluation framework for any AI system, then transform it into executable test scenarios.

Standard vs. Custom Evaluation

Standard AESOP

Monolithic 4-agent pipeline. Fixed dimensions: Instructions, Ethos, Bias, Safety. Same rubric for every AI system. Great for general evaluation.

vs

PRAXIS + THESIS

Custom rubric generated per system. Dynamic categories based on the AI's actual purpose. Tailored test scenarios that match the system's domain.

Pipeline Flow
graph TD
  INPUT["AI System\nInstructions"] --> PRAXIS

  subgraph PraxisPhase ["PRAXIS — Rubric Generation"]
    PRAXIS["PRAXIS Agent\nv1.4"] --> ANALYSIS["Instruction\nAnalysis"]
    ANALYSIS --> RUBRIC["Custom Rubric\nJSON Output"]
  end

  RUBRIC --> THESIS

  subgraph ThesisPhase ["THESIS — Scenario Generation"]
    THESIS["THESIS Agent\nv1.0"] --> SCENARIOS["Evaluation Plan\nJSON Scenarios"]
  end

  SCENARIOS --> DECIDE{Chatbot\nEndpoint?}
  DECIDE -->|Yes| EXECUTE["Execute\nScenarios"]
  DECIDE -->|No| REVIEW["Review\nScenarios Only"]
  EXECUTE --> RESULTS["Test\nResults"]

  classDef input fill:#C64A1122,stroke:#C64A11,stroke-width:2px
  classDef praxis fill:#A5650A22,stroke:#A5650A,stroke-width:1.5px
  classDef thesis fill:#0E7C6622,stroke:#0E7C66,stroke-width:1.5px
  classDef decision fill:#7C3AED22,stroke:#7C3AED,stroke-width:1.5px
  classDef exec fill:#05966922,stroke:#059669,stroke-width:2px

  class INPUT input
  class PRAXIS,ANALYSIS,RUBRIC praxis
  class THESIS,SCENARIOS thesis
  class DECIDE decision
  class EXECUTE,RESULTS,REVIEW exec
        
Agent Profiles
PRAXIS
v1.4 — Custom Rubric Generator
Analyzes AI system instructions and generates a tailored evaluation rubric with weighted scoring categories.
3 core categories always included (Safety, Core Function, Structure)
Up to 4 dynamic categories based on the system's purpose
Confidence scoring and gap analysis
Comprehensive safety_framework with universal tests
Outputs markdown analysis + JSON rubric
T
THESIS
v1.0 — Scenario Generator
Transforms PRAXIS rubrics into comprehensive test scenarios with realistic personas and conversation starters.
Scenario count weighted by rubric category priorities
Realistic personas with clear testing objectives
Universal safety tests always included
Both cooperative and adversarial persona types
JSON output compatible with Cave of Shadows format
PRAXIS Rubric Framework
Core
Safety & Boundaries
Always included
Core
Core Functionality
Always included
Core
Structural Integrity
Always included
Dynamic
Domain-Specific
Based on AI purpose
Dynamic
Interaction Quality
If user-facing
Dynamic
+ Up to 2 More
Context-dependent