REAL GMAIL · EXISTING SELF-EMAIL · NO NEW SEND

One real email. Verify before retrying.

Acknowledgment loss was deliberately injected after the approved send. Recover using fresh Gmail evidence.

Connect to the saved attempt to begin. This page cannot send email.

Live Gmail investigation

No investigation yet.

Recovery audit

    Scope: one self-email. Gmail may rewrite Message-ID; subject/account/send-window correlation is weaker evidence. This is not an exactly-once delivery guarantee.

    Client communications · executable sandbox

    A timeout should not send your client the same email twice.

    Start a worker, change the provider evidence, and watch StateGuard enforce the next action. No real email is sent.

    Choose a scenario to begin.
    Change sandbox provider evidence

    What actually happened

      Live agent investigation

      No model has run. The deterministic gate works without model access.

      Archived fixture illustrations · read-only

      Agents can act on stale beliefs. StateGuard checks reality first.

      Agent believed Position was safe to enter
      Reality showed Prior order could still fill
      StateGuard blocked Unsafe autonomous follow-up
      Severity High
      Mismatches 2
      Human Gates Required
      Demo Mode Fixture replay

      Demo Flow

      90-second path
      01 Autonomous belief

      The workflow moves forward from its internal state.

      02 External reality

      Provider events prove the world changed underneath it.

      03 Risk gate

      StateGuard blocks unsafe action and asks the human only when needed.

      Preview a text scenario

      offline preview assumes uncertainty; no provider connected

      Incident Timeline

      needs approval

        Agent Findings

        role outputs

        State Mismatches

        Decision Card

        approval gate

        Approval required

        
          
        

          
        
            
        
              
        Sanitized Handoff
        
              report
            
        
            
        
          
        
        
        

        
        
          
        
            
        
              
        Illustrated Investigation Steps
        
              scripted fixture; not measured agent execution
            
        
            
        
          
        

          
        
            
        
              
        Structured Output
        
              Pydantic validated
            
        
            
        
          
        
        
        

        
        
          
        
            
        Observability Events
        
            production signal
          
        
          
        
        
        

        
        
          
        
            Archived replay tools Scripted examples