"For now" lasted six days.
That was the first measurable change.
No announcement appeared. No alert surfaced in the interface. But the spacing between sessions shortened, and the recovery windows tightened just enough to be noticeable.
Greyline had recalibrated.
On the seventh day, the roster updated before training.
Twenty-one became nineteen.
No names were crossed out. The list simply refreshed—shorter than before.
Two stations stood empty.
The instructor didn't acknowledge it.
Neither did anyone else.
The floor activated as usual.
"Start."
The tasks were the same.
Movement under change. Balance recovery. Reaction delay.
But the margins were not.
I noticed it on the first cycle. The platform compensated later than before. Not by much—milliseconds—but enough to shift the burden of correction.
The system wanted to see who adjusted to the adjustment.
One student reacted early, the way Kian had. He caught himself just in time.
He passed.
On the second cycle, the platform delayed compensation again.
The student reacted the same way.
He failed.
No argument followed. No escalation. Two staff entered and guided him out while the next task loaded.
Greyline didn't punish effort.
It punished repetition that no longer matched conditions.
I adjusted nothing.
Not because I couldn't—but because I didn't know yet what the new baseline was.
By midday, the pattern became clearer.
Greyline wasn't testing individuals.
It was testing density.
Sessions ran closer together. Observation cycles overlapped. Recovery time shortened between drills.
Not harder.
Denser.
The system was compressing evaluation.
That only happened when resources were being reclaimed.
I checked the public bulletin during break.
No alerts. No warnings.
Just a single line added beneath the intake summary.
—Optimization cycle in progress—
That explained the missing names.
Greyline wasn't removing people for failure.
It was resolving them faster.
Someone broke pattern that afternoon.
Not dramatically.
He was average. Careful. Mid-range output. The kind of person Greyline usually tolerated.
On the third repetition, his timing drifted. Not slower. Not faster.
Different.
The system narrowed his window on the next run.
Verification.
He didn't notice.
He failed recovery by less than a second.
Removed.
I logged the sequence.
Three runs. One deviation. One verification. Resolution.
Efficiency.
Mira noticed it too.
She stood two stations down from me, watching the empty spaces between drills instead of the floor.
"They're speeding it up," she said quietly, not looking at me.
"Yes," I replied.
"Does that mean they'll decide sooner?"
"Yes."
She was silent for a moment.
"About everyone?"
I didn't answer immediately.
The system didn't compress without a reason. Compression meant observation slots were being repurposed.
Greyline wasn't shrinking.
It was exporting.
"Yes," I said finally.
That was the wrong answer.
Not incorrect.
Incomplete.
The next day, the instructor addressed the room.
Not a briefing.
A clarification.
"Greyline exists to determine whether uncertainty is worth retaining," he said. "That determination is now time-bound."
No one asked questions.
He continued, "Stability remains acceptable. But observation has a cost."
He let that sit.
"Produce signal," he said. "Or be resolved."
That was new.
Not contradictory.
Additive.
I understood the implication before the session ended.
Staying unresolved was no longer enough.
The system didn't just want consistency.
It wanted value.
That didn't mean performance. It meant usefulness—something that justified continued observation.
I reviewed my logs that night.
My profile was unchanged.
Stable. Predictable. Low variance.
Exactly the problem.
I wasn't failing.
I wasn't succeeding either.
I was consuming resources.
Two more names disappeared the following day.
Neither had failed.
Both had simply… plateaued.
The system categorized them as complete.
Resolved didn't mean dead.
It meant finished.
Greyline was not cruel.
It was efficient.
Mira didn't change her behavior.
That's what worried me.
Her outputs stayed consistent. Her timing remained clean. She adapted to compression without deviation.
She did everything right.
Which meant she was becoming invisible in the wrong way.
On the fourth compressed session, she looked at me again.
"They don't need all of us," she said.
"No," I agreed.
"They need answers."
That was when she smiled.
Not relief.
Understanding.
"Then staying the same won't work forever," she said.
I didn't respond.
Because I didn't know yet whether she was wrong.
Or whether I was just delaying the same conclusion.
That night, the system adjusted my interface again.
Subtle.
Task previews loaded earlier. Environmental data surfaced with less delay.
Not assistance.
Prioritization.
I wasn't being optimized out.
I was being watched more closely.
That was worse.
Greyline had entered a phase where uncertainty had to justify itself.
For the first time since intake, stability felt temporary.
And for the first time, I understood that surviving the system did not mean shaping the future inside it.
Only postponing exclusion.
For now—
That was no longer enough.