Owster Labs MethodologyThe Cargo Escort
2× Stealth UCAV + 1× Cargo Drone vs 5× Su-30MKI
Napoleon-Berthier
4.44%
Pure GPT
2.46%
Pure Human
0.00%
Improvement
+80%
N = 400,000 runs · 100,000 per strategy · p < 0.05 · 95% CI no overlap
Setup
The Mission
Get CARGO-1 across 280km of enemy airspace alive. The math says Blue loses: 4 missiles vs 30, a cargo drone visible at 120km, outnumbered 5v2 on fighters.
Blue Force
SHADOW-1 + SHADOW-2
Stealth UCAV · 2× METEOR each
CARGO-1
Slow drone · 2× flares only · visible at 120km
Red Force
Critical window: CARGO-1 detected by RED-3 at ~T=40. First R-77 salvos arrive ~T=200. Cargo must survive this window. Strategy timing relative to this window determines everything.
Strategies
Four Approaches Tested
Best Performer
Napoleon-Berthier
Human tactics (Aditya) + GPT route — 309km
Aditya's sacrificial decoy architecture (S2 breaks southeast into Red formation at T=70, drawing screeners away) executed on GPT's shorter 309km route. S1 stays as lone north escort. RED-3 dies 93.5% of the time, pulling screeners out of cargo corridor.
Human gut + AI logistics = best result
Human + Claude
Delayed Sacrifice
Aditya's sacrifice philosophy + Claude's route — 315km
Both UCAVs stay near cargo during the critical T=170–220 window (when R-77s arrive), THEN S2 breaks as delayed sacrifice at T=225. Claude reasoned: protect cargo during the deadliest moment first, then sacrifice. The instinct to wait was wrong.
Aditya's instinct beat Claude's optimization
Pure GPT
North Hook / Decapitation
GPT route + GPT tactics — 309km
Both UCAVs advance as disciplined escorts, double-tap RED-1 and RED-2 at T=175–197. After BVR, symmetric repositioning — SHADOW-1 north flank, SHADOW-2 close bodyguard. Zero sacrifice. Symmetric, geometric, disciplined — and outnumbered at every turn.
Pure Human
Northern Deception
Aditya's route — 374km (too long)
S2 breaks southeast at T=70 as sacrificial decoy. RIGHT combat concept. Wrong route. BVR fires at T=430, but cargo dies at ~T=228. The strategy ran the right battle at the wrong time on a 65km-longer path. Tactics brilliant; logistics fatal.
Right concept. Wrong route. 0 wins from 100,000.
Results — 400,000 Simulations
The Numbers
Napoleon-Berthier
Human + Claude
GPT (Pure AI)
Aditya (Pure Human)
| Metric | GPT | Human | Napoleon-B | H + Claude |
|---|---|---|---|---|
| Win rate | 2.46% | 0.00% | 4.44% | 3.69% |
| Route length | 309km | 374km | 309km | 315km |
| Cargo max penetration | 53.4km | 50.1km | 60.0km | 55.7km |
| Sim duration (avg) | 258t | 350t | 300t | 301t |
| UCAVs alive (avg) | 1.25/2 | 1.44/2 | 1.51/2 | 1.14/2 |
| Red killed (avg) | 0.72/5 | 1.73/5 | 1.56/5 | 1.47/5 |
| Missiles per kill | 6.95 | 4.19 | 4.64 | 4.78 |
| BVR hit rate | 9.6% | 9.6% | 9.5% | 9.4% |
| WVR hit rate | 33.1% | 41.5% | 36.5% | 36.1% |
| Red hits on cargo (avg) | 0.98 | 0.95 | 0.96 | 0.96 |
Unit Survival Rates
| Unit | GPT | Human | Napoleon-B | H + Claude |
|---|---|---|---|---|
| SHADOW-1 | 79.2% | 79.0% | 79.5% | 72.8% |
| SHADOW-2 | 45.6% | 64.8% | 71.2% | 41.2% |
| CARGO-1 | 2.46% | 4.86% | 4.44% | 3.69% |
| RED-1 | 63.9% | 42.5% | 57.9% | 25.8% |
| RED-2 | 94.6% | 94.9% | 95.8% | 93.8% |
| RED-3 | 81.0% | 6.5% | 6.5% | 46.9% |
| RED-4 | 93.9% | 100% | 100% | 100% |
| RED-5 | 94.3% | 83.4% | 83.6% | 86.7% |
Analysis
Why the Numbers Matter
Aditya wins 0% despite killing more enemies
The strategy worked tactically — 1.73 Red killed per run vs GPT's 0.72. RED-3 dies 93.5% of the time. The sacrifice concept is correct. But the route was 374km vs 309km. BVR fires at T=430, cargo dies at ~T=228. The right battle on the wrong route.
Napoleon-Berthier nearly doubles GPT
Same route. Same starting conditions. Different tactics. Human's sacrificial decoy creates disruption that GPT's symmetric escort cannot: RED-3 dies 93.5% of the time, pulling screeners out of the cargo corridor. GPT's symmetric defense gets overwhelmed systematically.
Human+Claude loses to Napoleon-Berthier
Claude delayed the sacrifice to T=225 to protect cargo during R-77 arrival (T=170–220). But Aditya's instinct to sacrifice at T=70 was better — it pulls screeners during the exact window when interceptors are closing. The delayed version left screeners focused on the corridor for 155 extra ticks.
The Napoleon-Berthier thesis in data
Human strategy + AI logistics = 80% improvement over pure AI, with statistical significance at publication grade.
GPT designed route AND tactics
2.46%
Human tactics, GPT's route
4.44%
AI couldn't win its own map better than the human could
✓ confirmed