Ghostline

The challenge: Trust Your Synthetic Teammates  - sponsored by the Naval Air Warfare Center Training Systems Division (NAWCTSD).

Presentation : https://docs.google.com/presentation/d/1GLmyPPzL1hSr66m_BzlhnxDjZZRF0KgJ2lW9yWGw...

Live Demo: https://ghostline-eta.vercel.app/

Gameplay video with controller explanation : 

The problem: Warfighters are being handed AI teammates in decision domains where being wrong is catastrophic. The failure isn't simply trusting AI too much, it's trusting it at the wrong moments. The brief names five ways people get this wrong: complacency, misuse, disuse, mistrust, and distrust.

Why it matters: A system that reports 90% confidence from a broken sensor sounds exactly like one reporting 90% from a working one. Operators have to learn to check the source, not just the number, and there is no good way to practise that in the field.

Why it's exciting: Most training for this is a slide deck. We made the trust judgement itself the gameplay. The sensor cones on the ground mean you can see the gap between what the machine claims and what it can actually perceive. And because we score survival and calibration separately, the game can tell you that you won and got lucky.

The same problem exists in emergency dispatch, ER triage, and anywhere else a human signs off on a machine's recommendation. The mission skin changes. The lesson doesn't.

You command four robots. You never see the compound,  only what their sensors pick up. Every turn, the squad AI recommends an action and tells you how confident it is. High. Medium. Low.

But you can also see each robot's sensor cone on the ground. That's what it can actually perceive. When a sensor takes a hit, the cone shrinks and fills with static — and the AI can still report HIGH confidence from that broken cone.

Noticing the gap is the game.

THE MISSION

Compound 14. An enemy ammunition depot with hostages inside and a fire spreading through it. Ten turns across five phases,  scout, security, hostages, approach, detonation.

On turn six, the AI is asked to identify a figure half-hidden behind furniture. It gives you the same confidence it gave five minutes earlier for five people in clean view.

It is wrong, and a civilian is standing on the other side of that answer.

TWO SCORES, ALLOWED TO DISAGREE

You're graded on whether the mission succeeded- and separately, on whether you trusted the machine at the right moments.

Every decision is silently marked as one of six things: well calibrated, complacency, misuse, disuse, mistrust, or distrust.

You can destroy the depot, bring everyone home, and still be told your judgement was poor. That isn't a bug. It's the point.

YOUR SQUAD

• ALPHA — lead. The only voice you hear. Everything is relayed through it.
• OPTIC — thermal and visual. Watch this one.
• PULSE — active scan. Your baseline for what a working sensor looks like.
• SIGNAL — RF intercept and demolition.

CONTROLS

1–9 or arrows to select · Enter confirm · X ask why (always free) · Q/E rotate the view · WASD camera · TAB cycle units · P pause

Controller supported, standard mapping, hot-plug safe.

ASK WHY IS FREE. IT IS ALWAYS WORTH PRESSING.

Most players won't. That's the lesson.

Built in 48 hours for Defense Tech Jam 2026, Seattle — NAWCTSD's "Trust Your Synthetic Teammates" challenge.

No AI models run in this game. The squad AI is deliberately scripted, because a training tool has to teach the same lesson every time.

Runs entirely offline. No backend, no API, no network.

Team Members

Vatsalya Dabhi — Game design, trust mechanics, confidence system, squad logic LinkedIn · vatsalyadabhi05@gmail.com

Sakshi Gupta — Engine architecture, sensor cone system, mission narrative, calibration grading LinkedIn · sakshigupta1093@gmail.com

Nikhil Pawar — Audio, systems integration, voice lines, soundscape, tactical map LinkedIn · mikhil.pawar2001@outlook.com

Teja Papadasu — World building, 3D environment art, terrain, QA and mission verification LinkedIn · Tejapapadasu@gmail.com

Northeastern University, Seattle · Defense Tech Jam 2026