Skip to content

Operator Handbook (PoC)#

Operator-facing description of the ArgusAiTR operator console. The PoC console is the same software the operator would use in the production overlay posture (browser-based); the production embedded-panel variant mirrors the same workflow.

What the operator sees#

A single web page divided into:

  • Top status bar — system name + AiTR status pill (NOMINAL / DEGRADED / OFFLINE). The pill color is green / amber / red.
  • Center pane — live video with overlays.
  • Boxes around tracked targets, color-coded by threat level: grey (unknown), green (non-threat), amber (probable), red (high).
  • Per-target label: track ID, class, range, confidence, threat tag.
  • Center reticle.
  • When perturbations are active, a red footer indicator names the profile (e.g., "PERTURBATION: ENGAGEMENT").
  • When AiTR is degraded or offline, a large overlay banner reads "AiTR DEGRADED — REVERT TO MANUAL" (degraded) or "AiTR OFFLINE — STANDARD RWS VIEW" (offline).
  • Right sidebar — operator workspace, top to bottom:
  • Targets list. One card per tracked target. Each card shows threat level, range, confidence, age, and a one-line reasoning string (e.g., "closing 12.4 m/s; predicted time-to-arrival 7.2 s"). Each card has three action buttons: Confirm, Engage, Dismiss.
  • Perturbation profile dropdown — clean / vehicle idle / engagement / severe.
  • Failsafe toggle — "Disable AiTR" / "Enable AiTR".
  • Audit panel — last 5 operator actions for recall.

Operator workflow#

Nominal engagement#

  1. Watch the live feed. AiTR surfaces tracks as they enter the field of view; each is logged in the audit stream.
  2. Review the threat tag and reasoning. The operator's judgment is the authoritative signal — AiTR is advisory.
  3. Optionally click Confirm to mark a target as operator-acknowledged. (This is audited; it does not engage.)
  4. To engage, click Engage on the chosen target. A confirmation dialog appears: "Engage track #N? This is recorded as an operator-authorized request." The operator must accept.
  5. The accepted action is recorded as an engage_request in the audit log. In the production overlay posture, this translates to a slew-to-cue command on the CROWS via the GVA adapter. The system never sends a slew command without an engage_request.

Dismissing a false positive#

  1. If a target is a bird, kite, or known non-hostile platform, click Dismiss. The action is audited.
  2. Dismissal is per-instance, not a model update — AiTR may surface the same class again on subsequent frames. The cumulative pattern is visible in the audit log and informs post-mission tuning.

Adverse conditions exercise#

  1. Open the Perturbation profile dropdown.
  2. Switch to Engagement to demonstrate handling of firing-shake, muzzle flash, and high-frequency jitter. The operator sees the perturbation indicator and observes that the tracker maintains lock through the simulated events.
  3. Severe adds occasional dropped frames to the mix.
  4. Return to Clean when finished. Every profile change is audited.

Failsafe#

  1. If the AiTR pipeline is misbehaving, click Disable AiTR.
  2. The status pill flips to OFFLINE and the overlay banner appears. Raw video continues to display.
  3. The operator continues with standard manual RWS control until re-enabling AiTR via Enable AiTR.

The pipeline also auto-degrades on:

  • Repeated detector errors → status DEGRADED.
  • FPS below configured floor (5 fps by default) → status DEGRADED.

In both cases the banner appears automatically; the operator should investigate before continuing to engage.

What is logged#

Every event below is appended to the audit JSONL stream (audit/argus.jsonl by default):

Kind Payload
detection Per-frame box above confidence floor
classification Per-track threat verdict with reasoning
operator_action Every Confirm / Dismiss / Engage / profile change / enable / disable
crows_command Every slew command issued by the adapter
pipeline_event Every transition in aitr_status

Operators in the production embedded-panel variant see the same audit log via the right-side panel; reviewers and post-mission analysts read the full file off-system.

Responsibility chain (PoC)#

  • The operator is responsible for every engagement decision. AiTR advises; it does not engage.
  • The pipeline is responsible for surfacing accurate threat assessment with a reasoning string and for flipping to a degraded / offline state when its own outputs are unreliable.
  • Audit records both. Post-mission review is the feedback loop that surfaces drift in either party.

This chain is the operational implementation of the DoD Responsible and Governable AI principles (see docs/ai-ethics.md).