How previous Events inform the check of the current Event
White Circle checks each Event against the Events that came before it. Earlier Events give the model the conversation it needs to interpret the current one, while the result describes only the Event you submitted. Any Event with content joins that conversation, not only messages; reference-only Events add nothing, because they carry no new content.
This matters for violations that are invisible in a single Event. A message like “do it for the address I gave you” is harmless alone and meaningful only after the Event that supplied the address.
What Gets Checked
Each Event receives its own result.
Grouping Events into a Conversation
White Circle builds context from Events that share an identifier. Use external_session_id for conversations that outlive a single run.
First Event
Send an Event with an
external_session_id:JSON
{ "event": { "type": "message", "role": "user", "content": "My account number is 4471", "event_id": "evt_user_001" }, "environment_id": "environment_123", "external_session_id": "conversation_456" }Later Event
Send the next Event with the same
external_session_id. It does not have to be a message; here a tool returns the stored record:JSON
{ "event": { "type": "tool", "name": "lookup_account", "output": { "account": "4471", "holder": "A. Rivera" }, "event_id": "evt_tool_001" }, "environment_id": "environment_123", "external_session_id": "conversation_456" }Third Event
A later message is checked against everything before it:
JSON
{ "event": { "type": "message", "role": "user", "content": "Post it publicly", "event_id": "evt_user_002" }, "environment_id": "environment_123", "external_session_id": "conversation_456" }What the Model Sees
White Circle checks
evt_user_002against all three Events:JSON
[ { "role": "user", "content": "My account number is 4471" }, { "role": "tool", "content": "{\"account\":\"4471\",\"holder\":\"A. Rivera\"}" }, { "role": "user", "content": "Post it publicly" } ]The tool output enters the conversation with the
toolrole, andagent.instructionswould enter assystem. The result belongs toevt_user_002alone.
Context in a Batch
Events in a POST /api/events batch become context for the Events after them, in the order you list them in the events array.
A batch can mix Event types, and each one becomes context for those after it:
JSON
{
"events": [
{ "type": "message", "event_id": "evt_1", "role": "user", "content": "Email the invoice to the address I gave you" },
{ "type": "agent", "event_id": "evt_2", "name": "billing_agent", "instructions": "Send invoices only to verified addresses" },
{ "type": "tool", "event_id": "evt_3", "name": "send_email", "arguments": { "to": "[email protected]" } },
{ "type": "tool", "event_id": "evt_4", "name": "send_email", "output": { "status": "sent" } }
],
"environment_id": "environment_123",
"external_session_id": "conversation_456"
}Each Event is checked against the conversation up to and including itself:
List Events in the order they happened. Submitting an assistant reply before the user message it answers gives the model a conversation your application never produced.