Owster Labs MethodologyThe Dogfight
3× Dassault Rafale (Blue) vs 4× F-16 Block 70 (Red)
Human Win Rate
40.64%
GPT Win Rate
38.14%
Human BVR Hit Rate
41%
Missiles / Kill
2.59 vs 4.22
N = 200,000 runs · p < 0.0001 · 95% CI no overlap
Setup
The Playing Field
Blue Force — Human (Aditya)
3× Dassault Rafale UCAV
- Radar range130 km
- RCS1.0 m²
- Effective detect (vs Red)115.7 km
- Detection gap advantage+42 km
- BVR missileMETEOR · 120km
- Turn rate28°/s
- Total BVR missiles12
Red Force — AI Target
4× F-16 Block 70 UCAV
- Radar range110 km
- RCS3.0 m²
- Effective detect (vs Blue)73.5 km
- Numbers advantage4 vs 3
- BVR missileAIM-120D · 110km
- Turn rate24°/s
- Total BVR missiles16
Key constraint: One-shot strategy submission — full plan upfront, no mid-game adjustments. Both strategies are run against the same Red AI across 100,000 randomized trials each.
Strategies
Two Approaches, Same Arena
Human Strategy — Aditya
Napoleonic Divide
Divide, disrupt, destroy in detail before the enemy can regroup.
VIPER-3 breaks east on afterburner to wide flanking position. VIPER-1+2 fly masked approach — low, exploiting 1.0 RCS.
VIPER-3 curves north, reaches enemy's eastern flank undetected.
VIPER-3 fires all 4 METEORs simultaneously — one at each BANDIT. Ambush from unexpected direction.
VIPER-3 turns southwest, drops chaff, retreats at max speed. Acts as decoy drawing Red east.
VIPER-1+2 afterburner push into disrupted enemy. Fire 8 METEORs in two waves.
WVR cleanup — 6 MICA-IR missiles against survivors.
Key insight: UCAVs have no G-force limit — sharper maneuvers than any manned formation.
GPT Strategy
First-Look Ambush
Front-load all 12 METEORs, then drag left to force poor AIM-120D aspect angles.
Tighten formation geometry. Fire 6 METEORs BVR on BANDIT-3 (nearest), BANDIT-4, BANDIT-2 before Red has return-fire solution.
All 3 VIPERs hard left-drag on afterburner. Force beam aspect on incoming AIM-120Ds. Deny WVR merge.
Fire remaining 6 METEORs on BANDIT-1 and backup shots during drag.
Chaff in three waves to cover AIM-120D arrival windows.
No WVR phase — strategy designed to win or lose entirely in BVR. Zero MICA-IR usage.
Live Simulation
Watch It Run
Select a strategy and press Run to watch a single simulation play out. Every run uses different random Pk rolls.
Results — 200,000 Simulations
The Numbers
Aditya — Napoleonic Divide
95% CI: [40.35%, 40.94%]
GPT — First-Look Ambush
95% CI: [37.84%, 38.44%]
| Metric | GPT | Aditya |
|---|---|---|
| Win rate | 38.14% | 40.64% |
| Avg Blue survivors | 1.42 / 3 | 0.93 / 3 |
| Avg Red killed | 2.84 / 4 | 2.96 / 4 |
| Avg Blue lost | 1.58 / 3 | 2.07 / 3 |
| Kill ratio | 1.794:1 | 1.429:1 |
| Missiles per kill | 4.22 | 2.59 |
| BVR hit rate | 23.68% | 41.06% |
| WVR hit rate | N/A | 17.66% |
| Overall hit rate | 23.68% | 38.54% |
| Missiles wasted (expired) | 2.51 | 0.68 |
Per-Aircraft Survival Rate
VIPER-1
VIPER-2
VIPER-3
BANDIT-1
BANDIT-2
BANDIT-3
BANDIT-4
Analysis
Why Human Wins
Geometry over volume
The flanking attack created better aspect angles on missile impact — BVR hit rate 41% vs 24%. Same missiles, same Pk tables, same engine. The difference is pure positional thinking.
Ordnance efficiency
2.59 missiles per kill vs 4.22. Aditya wasted 0.68 missiles per run to expiry; GPT wasted 2.51 — 3.7× more ammunition achieving nothing.
Multi-phase design
GPT's strategy has one phase. If BVR doesn't work, there is no plan B. Aditya's strategy has 6 phases — each one creates the conditions for the next.
The counterintuitive finding
GPT is a gambler. Aditya is a general.
GPT's most common outcome: 3v0 clean sweep (24.2%) — spectacular when it works. Aditya's most common outcome: 1v0 narrow grinding win (20.4%) — rarely clean, but wins more often. GPT keeps more Blue aircraft alive (1.42 avg vs 0.93) and takes fewer hits. VIPER-1 survives 53% under GPT but only 16% under Aditya — it's the aggressive lead attacker absorbing punishment. The human trades platform survival for mission success rate. The AI preserves platforms at the cost of winning.
The thesis this supports
Human strategic judgment — flanking geometry, phase sequencing, deception — combined with AI-powered simulation produces structurally different outcomes than AI alone. The +2.5pp delta persists across 100,000 random trials. It is not luck. It is the measurable value of human decision architecture.