Skip to content

Glass Box

A controlled synthetic workplace for advanced AI practice.

AidGPT is developing a visible training environment where learners can practise more capable AI workflows using fictional records. Publication of connected features follows product acceptance testing.

Synthetic assessed practice. No production connection required. Accountable human decisions.

The assessed-practice boundary

Glass Box is a controlled training approach under development.

Product state
Pre-acceptance. Connected-feature publication follows product acceptance testing.
Data boundary
The course does not require real organisational or personal data.
  • Assessed practice uses fictional and synthetic records.
  • No production connection is required.
  • An accountable human owns the decision.

Transfer beyond the course

Apply the discipline through authorised routes

The course does not require real organisational or personal data. Learners may later apply the disciplines to approved work through their organisation’s authorised tools, policies and authority.

That later application sits outside assessed practice.

Target concept, subject to product acceptance

Target concept showing synthetic training services connected to an AI workspace, with an accountable human owning the decision.

Target concept in words

  • Fictional records would remain inside purpose-built training services.
  • MCP would provide a connection pattern.
  • An accountable human would review evidence and own the decision.
Target concept, subject to product acceptance. This is an architectural explanation, not live product UI.

Connection language

What MCP means here

MCP is a connection protocol and pattern. It is not itself a security boundary.

Conditional technical assurance

What product acceptance must establish

Acceptance criteria only. This is not evidence that connected controls exist today.

MCP provides the connection pattern between an AI workspace and purpose-built training services. It is not itself a security boundary. Product acceptance must establish isolation, authentication, bounded tools, read-only behaviour, evidence and operational testing before connected features are published.