System Card

This system card documents ORA’s intended use, capabilities, limitations, and safeguards. These disclosures follow emerging model-documentation standards and the EU AI Act transparency requirements effective August 2026.

Important: 3D3D does not train language models. ORA is model-agnostic: it routes inference to the models you configure — local models running on your own hardware, or cloud endpoints you explicitly enable. Model capabilities, limitations, and biases are those of the underlying model you choose. ORA’s job is to govern what any model is allowed to do.

The Governed Operator

Intended Use

Business workflow operation with human oversight: research and analysis over your local files and memory, drafting content and communications for review, and running configured business operations (leads, jobs, approvals, invoicing) behind a fail-closed security broker.

Architecture

• Security broker (Rust): gates every tool call; blocks destructive and exfiltrating actions by default; signs the audit chain; holds the egress lock
• Decision kernel: evaluates each proposed action for impact, reversibility, and evidence, and routes risky actions to the human approval queue
• Workspace: the plain-English control surface — approvals, pipeline, playbooks, rules, and the stop button

Capabilities

• Multi-step task planning and execution through audited tools
• Research, analysis, and summarization over local files and memory
• Drafting (documents, communications, code) for human review
• Business pipeline operation with human approval on external actions

Limitations

• May produce incorrect or fabricated information (hallucination rates vary by the model you configure and the task domain)
• Knowledge cutoffs apply to the underlying models
• Performance can degrade on very long sessions or complex multi-step workflows
• Not validated for medical, legal, or financial advice
• Generated code should be reviewed before production use

Bias Considerations

ORA does not modify model weights and inherits the characteristics of whichever models you configure. Training-data composition and bias mitigations are documented by the respective model providers. Because ORA is model-agnostic, you can select models whose documentation and provenance meet your requirements.

Misuse Prevention

ORA’s governance floor — the constitution’s locked rules, A0–A5 authority levels, the fail-closed broker, and the human approval queue — prevents escalation beyond configured authority. Destructive actions, external sends, spending, and bulk data movement always require explicit human approval. Prohibited uses are defined in the Acceptable Use Policy.

Feedback and Errors

Report errors or unexpected behavior to [email protected]. Include what you asked for, what happened, and the relevant audit-trail entries if available.