The arithmetic had changed. For the first time since the Rival Rankings began, the numbers were inside the window of what was possible without requiring GD-02 to finish in third place.

Leo sat at the grid for the start of lap ninety-three, staring at the telemetry on his head-up display. The rain was a steady, rhythmic drumming against the carbon fiber canopy of the pod. It was the only sound in the world besides the low, digital thrum of the simulated engine.

’P1 every lap, GD-02 P2 every lap,’ he thought. His mind was working faster than the data could refresh. ’Net gain: 3 points per lap. 24 points in eight laps.’

’Still one short.’

The logic was cold and hard. He had closed a gap that should have been impossible to bridge. He had fought through the despair of the early sessions and the physical agony of the x500 pain scaling. But the math of the ranking system was a final, unyielding barrier. Even if he drove perfectly, even if he didn’t make a single mistake for the rest of the race, he would still lose by a single point.

’So GD-02 needs to lose one P2 somewhere,’ he realized. He watched the silver car ahead of him, its rain light blinking with mechanical indifference. ’Drop to P3 for one lap. Just one.’

He rolled into Sainte Dévote for the start of the ninety-third lap and stopped thinking about the standings. He had to stop. The mental load of the calculations was a distraction he couldn’t afford. In Monaco, a distraction was a death sentence. He needed every bit of his brain to manage the car, the track, and the ghosts.

What happened across laps ninety-three through ninety-six had no precedent in the session dataset.

Leo didn’t just drive; he existed in a state of continuous improvisation. He drove four consecutive laps in which no sector time matched the equivalent sector from any previous lap by less than 0.3 seconds. To a spectator, it would have looked like he was struggling to find a rhythm. To the Simex system, it was a nightmare of unpredictability.

He varied his braking points by meters. He changed his turn-in angles by degrees. He shifted gears at different RPMs, sometimes short-shifting to keep the rear stable, sometimes holding a gear long to maximize torque on the exit of a corner. The inputs varied continuously, the lines generated fresh each corner from the available information. He was placing the car through each sequence as though he was encountering it for the first time, yet he did so with the full knowledge of what the tarmac contained.

The Ghost Drivers could not model it. GD-03, the adaptive profile that had been his greatest shadow, stayed suspended. It couldn’t find a pattern to copy. But the real victory was what was happening to GD-02.

The "Professor" was starting to crack. Not in its driving, the silver car was still fast, but in its logic. GD-02’s precision model began logging ERROR flags in its sector prediction system. These weren’t mechanical failures, but instances where the AI’s internal map of the race was failing to match reality. The predicted position of Leo’s car at any given track marker was diverging from his actual position by margins that exceeded the model’s uncertainty threshold.

[SIMEX BROADCAST, LAP 25 of 30 (95) NOTE:]

[ERROR LOG, GD-02 PREDICTION MODEL:]

[Sainte Dévote entry: predicted position divergence, 2.3 meters.]

[Casino Square apex: predicted position divergence, 1.8 meters.]

[Tunnel exit braking: predicted position divergence, 3.1 meters.]

[Nouvelle Chicane line 1: predicted position divergence, 4.7 meters.]

Leo saw the data flicker on his secondary display. He didn’t smile, he didn’t have the energy for it, but he felt a grim sense of satisfaction.

[GD-02’s model cannot locate LEO KAITO at four of the circuit’s critical reference points. The precise profile is racing a driver it can no longer predict.]

[For context: Ghost Driver profiles can predict each other’s positions to within 0.3 meters at all reference points. Average professional driver predictability margin: 0.8 meters. LEO KAITO’s current predictability margin: 4.7 meters at worst case.]

[The human driver is six times less predictable than the average professional he was benchmarked against at Phase 1 initialization.]

He was a ghost. He was the noise in the system. He was the variable that the most advanced racing AI ever built couldn’t solve. He took the Swimming Pool chicanes with a violence that made the haptic gloves vibrate against his palms, the car’s floorboard striking the simulated ground with a shower of sparks.

Every time GD-02 expected him to be on the inside, he was on the outside. Every time it expected him to brake early for a puddle, he drove straight through it, using the "Rain Mastery" skill to find grip where the AI’s logic said there was none.

[GRID RANKINGS, LAP 25 of 30 (95):]

[1st: GD-02, 168 points]

[2nd: LEO KAITO, 162 points]

[Gap: 6 points. Five laps remaining.]

Six points.

Leo read the number at the grid for lap ninety-six and held very still. His hands were shaking, a fine tremor that he had to force down by gripping the wheel harder. His vision was tunneling, the edges of the screen becoming a blur of grey and black.

Six points. One P2 from GD-02 in five laps would close it entirely. Two P2s and he would lead on countback. The number had come from twenty-eight points down with nine laps left, and it was now six with five remaining. The compression had happened so gradually, lap by painful lap, that seeing it at six felt like looking up from a long day of work to find the room suddenly dark.

’When did it get to six?’ he thought.

He couldn’t remember the specific moments. He didn’t remember the exact lap where twenty-eight had become twenty-five, or where twenty-five had become nineteen. He didn’t remember the transition from eleven to six. He had been driving. The numbers had been moving. The work and the numbers existed in separate layers of his consciousness, and the work was the only layer he could actually operate.

’Five laps,’ he thought. ’GD-02 needs one mistake.’

Then, a darker thought followed.

’GD-02 doesn’t make mistakes. It’s a machine.’

’I’ll have to make it make one.’

He pushed the throttle for lap ninety-six. He wasn’t just racing for time anymore; he was racing for psychological dominance over a program that didn’t have a psyche. He was trying to overload the AI’s processing core by providing too much conflicting data.

He drove the ninety-sixth lap with a level of aggression that bordered on suicidal. He was using the "Danger Sense" to its absolute limit, placing the car in positions where a collision was 99.9% certain, only to deviate at the last possible millisecond. He was forcing GD-02 to constantly run its emergency avoidance protocols. Every time the AI had to avoid him, it lost a few hundredths of a second. Every time it lost time, its "Optimal Line" was disrupted.