One extra day mattered to Leo. It was a day to see the sky. A day to breathe air that didn’t smell like ozone and recycled carbon. A day to remember what it felt like to be a person instead of a cursor in a military learning program.

But the mathematics were what they were, and the laps were what they were, and the only variable he controlled was the next corner.

He drove the next corner.

Leo had a wild idea somewhere in the Fairmont Hairpin on lap eighty-three.

It was an observation, the kind that surfaced when the conscious mind had stopped generating instructions and the part underneath had been running quietly for long enough to notice something new. The simulation wasn’t just a race; it was an ecosystem of logic. And like any ecosystem, it had a food chain.

GD-03 was suspended currently. Its adaptive model had shut down because Leo had stopped providing patterns to adapt to. The profile had reverted to baseline variable composite, a randomized behavior set that was competitive but no longer specifically dangerous. It was a predator that had lost the scent of its primary prey.

But a mirror that wasn’t reflecting Leo was still reflecting something.

"It mirrors the closest established pattern in the dataset," he thought, through the apex of Rascasse. "If it can’t model me, it defaults to the next available reference."

He felt the car skip over the damp kerb. The vibration traveled through the carbon fiber seat and into his spine. It was a sharp, mechanical reminder that he was still moving at speeds that would kill a normal man. But his mind was elsewhere, dissecting the digital ghosts surrounding him.

"What’s the next available reference?"

The answer was obvious. GD-02. The precise profile. The session benchmark that had dominated the first ten ranked laps with machine consistency. GD-02 was the North Star of this simulation. It never deviated. It never improvised. It was the closest thing to a fixed, repeatable pattern in the current field.

Leo came through Anthony Noghes and the Nouvelle Chicane. He crossed the start-finish line, the spray from the puddles hitting his visor in a rhythmic thumping sound. He let the thought finish assembling itself at the grid between laps.

"If GD-03 starts mirroring GD-02," he thought, "and I can influence GD-03’s behavior by controlling what I show it, not my real inputs, but a version of my inputs that it interprets as a new pattern, then I can use GD-03 as a tool."

The mathematics were what they were. The gap in points was still massive. He was at the bottom of the leaderboard looking up at a mountain. But if he could do that to interfere slightly, he could change the math.

GD-02 had never had to race against a version of itself before. It was a machine designed for perfection in clear air. What happened when perfection met its own reflection in a dirty, wet slipstream?

[SIMEX BROADCAST, LAP 13 of 30 (83) START:]

[LEO KAITO holds pole for a third consecutive lap. The human driver’s positioning behavior at race start has shown variation across each lap; different lines, different spacing, and different throttle application rates on the standing start. No two starts have been identical.]

[GD-03 adaptive profile status: BASELINE VARIABLE. Operating on randomized composite inputs. Default behavior in absence of primary target data: mirrors closest consistent behavioral pattern in session dataset.]

[Closest consistent pattern: GD-02.]

[GD-03 is now shadowing GD-02’s optimal lines.]

[Note: GD-02 has not encountered a driver mirroring its inputs before. The precise profile is optimized for clear track ahead. Close pursuit from a mirroring driver represents an unmodeled variable.]

The lights on the gantry went out.

Lap eighty-three began with a roar that shook the Iron Cell. Leo didn’t launch with the same violence he had used on lap eighty-two. He was more measured. He took the first sector clean and fast, but he was playing a different game now. He varied his Sainte Dévote entry by six meters compared to the previous lap.

It was a tighter radius. He turned in earlier, clipping the apex with a precision that felt almost surgical. It was a line that was objectively slower through the corner itself. On paper, he was losing time. But he wasn’t looking at the corner. He was looking at the hill beyond it.

By taking the tighter line, he loaded the front-left tire differently for the compression into Massenet. He could feel the tire wall protesting, a high-pitched hum vibrating through the steering rack.

"Work for me," he whispered.

He wasn’t optimizing for Sainte Dévote. He was optimizing for Massenet, treating the two corners as a single linked system rather than sequential problems. The front-left load he had built in the first corner paid dividends in the second.

As the car hit the bottom of the hill, the nose stayed pinned. The grip was there, waiting for him. His throttle was available half a car length earlier than GD-02’s optimal could access.

GD-02, running the established optimal line through both corners in sequence, produced a faster Sainte Dévote. It was a beautiful, flowing arc. But it was a slightly slower Massenet. The AI had to wait a fraction longer for the car to settle before it could go to full power.

The combined sector time was identical. On the timing screen, they were tied.

But the tire state wasn’t.

Leo’s load-transfer approach was asking more of the front-left through the sequence. He knew it. He could feel the additional demand in the steering column, a marginal increase in effort at the apex of Massenet that hadn’t been present in previous laps. The tire was getting hotter. The rubber was wearing down faster.

"I’m using the tire faster," he thought. "Consciously. To carry pace from an earlier throttle point."

It was a transaction. He was trading the life of his tires for a specific type of momentum. In a long race, it would be a mistake. But in these thirty laps of hell, it was a weapon.

"Best make sure it pays."