Observe
Inspect a structured demonstration trace at multiple levels of detail, with clear labels for what was recorded and what remains unavailable.
Open source · Research prototype v0.1
Project
Glassbox.
Open infrastructure for observable, verifiable, human-governed AI.
When an AI agent takes action, people need to know what happened, what evidence supports it, what authority applied, and what still cannot be verified.
INSPECTABLE SYSTEM ACTIVITY
Observable
events↓Evidence &
provenance↓Human
review↓Audit
record
Conceptual architecture · Not live telemetryevents↓Evidence &
provenance↓Human
review↓Audit
record
The problem
AI should not be
an unreviewable black box.
A confident explanation is not the same thing as evidence. Glassbox explores a practical way to distinguish recorded actions, reproducible steps, inferred interpretations, restricted information, and events that were never instrumented.
What the prototype demonstrates
Trace evidence
Follow visible events to their context, producers, transformations, and supporting records rather than trusting a summary alone.
Verify within limits
Use browser-based WASM deterministic replay for supported portions of the demonstration. Replay does not prove the model's judgment was correct.
Keep people in control
Review and challenge claims, inspect authority boundaries, and export an audit record for further examination.
A precise boundary
Visible evidence.
Not private thoughts.
Project Glassbox does not reveal a model's private chain-of-thought, decode neural activity into literal thoughts, or guarantee that an AI conclusion is true. Its demonstrations distinguish observed events, deterministic verification under declared conditions, human-readable interpretation, inference, and unavailable or restricted information.
Status and limitations
Built for scrutiny.
Still a prototype.
The v0.1 interface is functional, but the supplied traces, providers, models, and telemetry are synthetic. It is not an authenticated production gateway or compliance solution. Recorded browser events and exported audit files are not signed or tamper-evident; live integrations require additional instrumentation.
These limitations are documented openly so potential contributors can work on them.
Help build the glass box
Transparency is
a collective effort.
Developers, AI-safety researchers, designers, auditors, public servants, and educators can help strengthen the evidence model, verification boundaries, integrations, and human experience.
Founded and product-directed by Leonidas Esquire Williamson · Stewarded by EGOVERNMENT AI LLC · Licensed Apache 2.0.