Agent System Cards

System cards document each ORA agent’s intended use, capabilities, limitations, and performance characteristics. These disclosures comply with emerging model documentation standards and the EU AI Act transparency requirements effective August 2026.

Important: 3D3D does not train the underlying language models. ORA routes to third-party models (by default, NVIDIA NIM / ZhipuAI GLM series). Model capabilities, limitations, and biases are those of the underlying model provider.

ORA Operator Agent

Intended Use

General orchestration, mission management, system operations, and decision-making. Designed for business workflow coordination with human oversight.

Underlying Model

GLM-5.1 (ZhipuAI, 754B parameters) via NVIDIA NIM. Supports thinking mode and tool use.

Performance

• SWE-bench Verified: ~87% (GLM-5.1, April 2026)
• Reasoning tasks: Strong performance on multi-step planning
• Tool use: Native tool calling with structured output

Limitations

• May produce incorrect or fabricated information (hallucination rate varies by domain)
• Knowledge cutoff applies to the underlying model
• Performance degrades on very long context sessions
• Not validated for medical, legal, or financial advice

Bias Considerations

GLM-5.1 is trained by ZhipuAI (China). Training data composition and bias mitigations are documented by ZhipuAI. ORA does not modify the model weights and inherits the model’s characteristics.

ORA Researcher Agent

Intended Use

Memory search, data analysis, findings synthesis, and knowledge retrieval from MCP-connected sources.

Underlying Model

GLM-5.1 (same as Operator). Tool capabilities: MCP memory search/store, file reading, directory listing.

Limitations

• Knowledge is bounded by MCP memory contents and accessible files
• Cannot independently browse the internet
• Summaries may omit important context or nuance
• Source attribution is provided but should be verified

ORA Builder Agent

Intended Use

File creation, structured document generation, configuration management, and code generation.

Underlying Model

GLM-5.1. Tool capabilities: file creation and modification within configured workspace.

Limitations

• Cannot execute code unless explicitly enabled
• File operations confined to configured workspace
• Generated code should be reviewed before deployment to production
• May generate syntactically correct but semantically incorrect code

Misuse Prevention

ORA’s governance kernel (four gates: structural, evidence, consistency, consensus) prevents agent escalation beyond configured authority levels. Prohibited uses are defined in the Acceptable Use Policy. Users who attempt to circumvent governance mechanisms will have their accounts suspended.

Feedback and Errors

Report agent errors or unexpected behavior to info@3d3d.ca. Include session ID, agent type, and description of the unexpected output.