What happened?
Target: append-only requests, plan changes, tool events, challenges, approvals, and reports.
PRODUCT LAB / INTERACTIVE PROTOTYPE
What if AI work could be seen from your angle?
GlassBox explores a better interface for understanding AI-assisted work. It keeps plans, evidence, tool events, uncertainty, challenges, human decisions, and outcomes in one observable record—then projects that record into different views for different ways of thinking.
01ReadApproved context only
02UnderstandSeparate fact, claim, and uncertainty
03ReportConnect every conclusion to evidence
04Ask before actionHuman authority stays visible
THE INTERFACE / 01
The same run can make sense in more than one shape.
This fixture is fictional and deterministic. Switch views or replay its twelve records; the selected item, evidence, illustrative confidence label, and limitations stay synchronized.
GLASSBOX LAB / SYNTHETIC RUN 0001
One run. Six ways to understand it.
Select a lens, replay the work, or focus a record. The arrangement changes; the underlying evidence and state do not.
- Synthetic fixture
- R0 · Offline interpretation
- No live system access
- Simulated roles
- 0 changes
REVIEWED RESULTVerify one fictional local management interface.
- Why
- Reachability is recorded; intended access is not.
- Confidence label
- Illustrative medium · not calibrated
- Action
- Manual check · no change authorized
VIEW 01
Timeline
When did it happen?
TARGET PRODUCT MODEL / 02
The proposed interface is a projection, not the database.
In this prototype, every lens is rendered from the same normalized in-page fixture. A future system would preserve that principle in persistent storage: visual position may help navigation, but it never becomes proof, confidence, causality, or truth.
What connects?
Target: typed relationships between evidence, claims, alternatives, findings, decisions, and output.
Why believe it?
Target: linked source artifacts for reportable claims. The prototype currently preserves named source labels and uncertainty.
What matters now?
Target: a bounded active context selected for the current question, scope, and operator decision.
What may happen?
Target: enforced risk classes, connector capabilities, stop conditions, privacy rules, and approval gates.
What can be stated?
Target: a reviewable deliverable that separates evidence, interpretation, limits, and human judgment.
CURRENT BUILD / 03
A working interface study with an honest boundary.
Implemented now
- One normalized synthetic run
- Six synchronized, keyboard-operable lenses
- Replay and cross-view record selection
- Evidence, claim, alternative, critic, decision, and report separation
- Responsive and reduced-motion behavior
Still incomplete
- Real connector integrations and persistent event storage
- Measured path-based memory retrieval
- Authentication, multi-user review, and retention controls
- Live agent or model integration
- Approved remediation, which remains intentionally disabled
GlassBox, ToolBridge, and ReportForge are working names pending brand and trademark clearance.
FIRST PRODUCT QUESTION
Who should GlassBox help first?
Agent builders debugging work, clients reviewing AI-assisted outcomes, or people trying to understand their own systems? The next phase is to test the lenses with a real audience before adding spectacle.