Watchspan

Fleet oversight control room

Three institutional agents are asking one reviewer for approval. Watch the attention budget drain, the reviews turn into stamps, and Watchspan catch the moment it happens.

Nothing measured yetrun the fleet, or take the queue yourself, and a verdict appears here

Attention budget remaining
35
%

Calibration floor: 35%. Below it, the Calibrator proposes a stricter escalation threshold.

Routed
–
total requests
Escalated
–
to a human
Auto-run
–
logged for audit
Paused
–
by Sentinel
Approved without being read

High-risk actions the reviewer stamped appear here during a run.

Sentinel · requests held for review

No suspicious timing or batching patterns detected.

The real fleetADK · Gemini 3.5 Flash

The 30-minute run is seeded, because that is what makes the drift visible. This gives three tasks to the actual ADK fleet and governs whatever it decides to ask for. Each task is a model call, so it takes a moment.

Attention budget across the run
The attention budget renders here during a run
Request stream0 routed

No requests routed yet. Select “Run the fleet” to start.

Risk scored 0 to 100. Scores of 70 and above shown in amber.

Review these yourself

Everything above is Watchspan measuring a simulated reviewer. This is twelve real requests from the same fleet, routed through the same governance layer, decided by you. Watchspan times each decision and counts what you open. Nothing you send can set those numbers.

Google Cloud footprint, checked on this page load

  • Agent Runtime
  • Agent Registry
  • Memory Bank
  • Model Armor
  • Vertex AI · Gemini
  • Cloud Trace
  • Cloud Run

A live call means this page load reached the service and read its answer. A config check means an environment check only, and is counted separately.

Calibrator · policy proposal

No proposal pending. The Calibrator raises one when the budget crosses the 35% floor or oversight is declared degraded.

Article 14 evidence

Who reviewed what, with how much attention available, and when oversight degraded. The evidence of effective human oversight that EU AI Act Article 14 requires.

Available after a run completes.