Injection Audit Trail

Tenure keeps a record of every context injection event. Each time beliefs are retrieved and injected into a request (or would have been injected in observation mode) a record is written to the audit log.

What is recorded

Each audit record captures:

• User query - the raw query that triggered retrieval
• Expanded query - the normalized version used for search
• Injected beliefs - a snapshot of every belief included in context, split into three groups:
  • Pinned facts: always-on beliefs you have explicitly pinned
  • Relevant beliefs: beliefs retrieved by semantic search
  • Open questions: pinned questions surfaced to keep them visible
• Belief count - total number of beliefs in context
• Injection status - whether beliefs were actually injected or the session was in observation mode (!extract off)
• Scope - the scope tags active for the request
• Agent ID - the agent identifier if routed through an agent client
• Timestamps - when the injection occurred

Snapshots are taken at injection time, so the audit record reflects exactly what the model saw, even if beliefs are later edited or deleted.

Viewing the audit trail

Open http://localhost:5757/audit in your browser. You can filter by:

• Date range - narrow to a specific window
• Scope - select from a dropdown of all scopes seen in your audit history
• Belief ID - see every request where a specific belief appeared

Click any record to see the full detail view, including belief content, confidence, epistemic status, and why-it-matters annotations as they were at injection time.

You can also reach this filtered view by clicking the Injections button on any belief card in the World Model dashboard.

Orientation tax dashboard

The dashboard at the top of the audit page shows four measurements for the last 30 days:

• Re-explanations prevented — turns where beliefs were injected and the next turn was not a correction. A floor, not a ceiling: silent failures where the model quietly used stale context are not visible here.
• Estimated time saved — re-explanations prevented × 1 minute. Labeled as estimated because the true cost of a correction turn varies.
• Tax still paid — turns where the sidecar determined your message existed primarily to re-establish context memory should have supplied. This is the most directly observed number on the dashboard.
• Re-explanation trend — whether tax-paid turns are increasing or decreasing across the period. Down is good. — means not enough data yet to compute a direction.

API access

The audit trail is also available via the admin API:

Query parameters:

Parameter	Type	Description
limit	number	Max records to return (default 50, max 200)
skip	number	Offset for pagination
start	string	ISO date: only records after this time
end	string	ISO date: only records before this time (inclusive to end of day)
scope	string	Filter by exact scope value
belief_id	string	Filter to records containing this belief

Returns a single record by ID. Returns 404 if not found or if the record belongs to a different user.

What audit logging does not capture

• The full system prompt or injected context string: only the belief list is stored, not the rendered prompt
• The model’s response
• Requests where no beliefs were retrieved (zero-belief events are silently skipped)