The problem with the delay was that it was not where she had expected to find it.

Lyra had assumed — based on the debrief comment, based on her own immediate post-match assessment, based on two years of knowing roughly where her execution gaps tended to live — that the two seconds were in her hands. A mechanical hesitation. The kind of thing that could be drilled out with enough repetition, the muscle memory overwritten until the movement happened before the thought finished forming.

She pulled up the timestamp at minute twenty-one and ran it frame by frame.

The hands were fine.

She ran it again, this time tracking only her character's movement relative to the moment she had identified the read. The movement was fast — not instant, not the theoretical zero-delay of a decision that had already been made before it needed to be made, but fast. Competent fast. The kind of fast that would not have been notable in most plays.

The delay was earlier.

She found it on the fourth viewing: a pause of approximately forty frames, which at the match's frame rate translated to roughly one point three seconds, occurring at the exact moment between *I have the read* and *I am committing to the read.* Not a mechanical delay. A cognitive one. The moment where she had known what she was going to do and had stopped, briefly, to verify that she knew.

Verification time: one point three seconds. Cost: the full value of the play.

She sat back in her chair and looked at the frozen frame on her monitor.

"Not the hands," she said, to no one. The room was empty and the door was mostly closed and there was no reason to be quiet. "It's the confirmation step. I already had the read — I had it before I moved — and then I stopped to make sure I had it. And the stop cost the window."

She ran it again.

"If I remove the confirmation," she said, watching her character hesitate on screen with the strange detachment of observing a past version of herself, "the play lands clean. But if I remove the confirmation and the read is wrong — "

She paused the clip.

This was the problem.

The read had been right. Eighty-five percent of the time — her rough estimate, based on how this class of situation had resolved in her recent match history — the read would be right. But fifteen percent of the time it would not be, and committing without confirmation in those fifteen percent cases would cost her team something significantly worse than a two-second delay.

She needed to know if eighty-five percent was accurate.

She opened a new tab and started pulling match data.

---

She had been working for about forty minutes when she heard the footsteps.

They came from the corridor — the particular quality of sound that the second floor hallway had in the late afternoon, when most of the building's population was in practice rooms or their own spaces and the shared areas were quiet. She had learned the building's acoustic signature over the past week the same way she learned everything: without deciding to, through accumulated exposure.

These footsteps had a specific quality.

She knew — not with certainty, but with the same kind of probability-weighted read she applied to everything — whose they were. The pace, the weight distribution, the particular evenness that she had cataloged without intending to over ten days of living in the same building.

She kept her eyes on her screen.

The footsteps slowed outside the door.

Stopped.

She was in the middle of a thought — a specific calculation about sample size, whether her last forty-three matches constituted a large enough dataset to draw reliable conclusions about her read accuracy in this class of situation — and the thought was important enough that she did not want to interrupt it. She kept working. Her mouth was moving, she was aware of that, the quiet external processing that happened when she was deep enough in a problem that the internal-only version became insufficient.

"Forty-three samples," she said, at a volume she had calibrated to the empty room without accounting for the possibility that the room was adjacent to someone standing very still. "Probably enough for a directional number, not enough for a confident one. I need the last sixty to be sure."

She tabbed to an older match log.

Outside the door, nothing moved.

She had a choice, she recognized — not a dramatic one, not a choice with significant stakes either way, but a choice nonetheless. She could stop talking. She could lower her voice. She could acknowledge the presence in the corridor in some way that made it clear she knew it was there and was adjusting accordingly.

Or she could continue working through the problem she was working through, which was important and which she was making progress on, and let the corridor be whatever the corridor was going to be.

She continued working.

"Sixty-one matches," she said, scrolling through the log. "Okay. That's the dataset." She opened a secondary tab and began tagging entries — situations that matched the criteria she had defined for this class of read, where she had had partial information and committed to a high-confidence interpretation of it. "This is going to take a while."

Outside the door, there was the faint sound of weight shifting.

Not leaving. Just adjusting — the small movement of a person who had been standing still long enough to need to redistribute their weight, which meant they had been there for at least a minute, possibly longer.

Lyra tagged another entry and kept going.

---

She worked through twenty-three entries before she heard the footsteps resume.

They moved away from the door — not quickly, not with the deliberateness of someone who had decided to leave, more with the gradual momentum of someone who had been standing still and was returning to motion. She listened to them until they reached the staircase and then until they didn't, the sound absorbed by distance and the building's ambient noise.

She sat in the quiet that followed.

She had known — eighty-something percent confidence — who was out there. She had not confirmed it. She had not looked up. She had continued working through her problem and spoken her thinking aloud and let the corridor contain whatever it contained.

This felt, on reflection, like a decision she had made without fully examining it.

She returned to her tagging.

It took another hour and twelve minutes to work through the full dataset. Sixty-one matches, forty-seven situations that met her criteria, thirty-nine of them resolved correctly — meaning she had committed to the read and the read had been right.

Thirty-nine out of forty-seven.

She did the calculation twice to make sure.

Eighty-three percent.

Not eighty-five. She had estimated eighty-five; the actual number was eighty-three. A difference of two percentage points, which in isolation was not significant, but which meant her intuitive calibration of her own accuracy was slightly optimistic, which was information worth having.

She looked at the number for a long moment.

Eighty-three percent meant that trusting the read — committing without confirmation, removing the one-point-three-second delay — would result in a worse outcome seventeen percent of the time. That was the cost of the approach.

She thought about what the current approach cost her. The delay, the missed windows, the plays that worked at seventy percent when they should have worked at ninety. She thought about the play at minute twenty-one, about the steal her team had gotten anyway, about how it would have looked if she had moved a second and a half earlier.

She thought about *fix the delay.*

Eighty-three percent was not a comfortable number. It was not the number she would have chosen if she had been designing herself from scratch. But it was her number, her actual number, the number that was true regardless of what she would have preferred.

And eighty-three percent was a number she could work with.

She wrote it at the top of her notes page, large enough that she would see it every time she opened the file.

83%

Underlined it.

Below it, she wrote: *Trust it. Improve it. In that order.*

---

She packed up her station at nine-forty — later than she had planned, later than most people in the building would still be working, but not the latest. She knew this because she could still hear, from somewhere below, the sound that had become as reliable as the building's other constants: the even, measured rhythm of a keyboard that did not waver.

She gathered her notes, logged out of the match database, turned off her monitor.

At the door, she stopped.

Not for long. Just the pause of someone who had thought of something and was deciding whether to act on the thought.

She stepped into the corridor.

It was empty, as she had known it would be — the footsteps from earlier had been gone for over an hour. The hallway had the particular quality it had at this time of night: functional, quiet, the overhead lights at their reduced evening setting. She turned toward the staircase.

As she passed the point in the corridor where she calculated the footsteps had stopped — outside her door, in the space between her practice room and the next door down — she did not slow. She did not mark the spot. She simply noted it, the way she noted the acoustic properties of the building and the reliability of the refrigerators and the specific angle at which the afternoon light came through the kitchen window.

She took the stairs to the third floor.

At the landing, before turning toward her room, she found herself oriented briefly toward the second floor — toward the corridor she had just left, toward the practice rooms, toward the faint sound that carried upward through the building's structure.

The light under the men's practice room door was still on.

She looked at it for two seconds.

Then she turned and went to her room, and did not think about what she had just done or why she had done it, and was largely successful at this for approximately the first four minutes before she fell asleep.