Chapter 3 - Hayes Said He Never Ordered My Dismissal
Hayes did not deny using the account.
He denied doing that.
“I added Sara to the analytics population.”
Ward looked at him.
“When?”
“Before the course.”
“For visibility?”
“Yes.”
“Did you assign R-17?”
“No.”
Ward checked.
The R-17 flag had been added two days before my dismissal.
Manual action.
Same credential.
I looked at Hayes.
“Who has access to your account?”
“No one.”
Ward’s expression sharpened.
“Then this is you.”
“It isn’t.”
Briggs laughed.
Hayes turned.
“You find something funny?”
“Yes, sir.”
“Say it.”
“You came here to catch me trusting bad information while telling us your own credentials can’t be trusted.”
Hayes’s face hardened.
He deserved that.
I asked Ward:
“Can actions be delegated through the account?”
“Some can.”
“This one?”
“No.”
“Shared terminal?”
“Credential still required.”
“Saved login?”
She hesitated.
“Possibly.”
Hayes closed his eyes.
Ward noticed.
“What?”
He looked at me.
“I use the operations terminal in my outer office.”
“Who else?”
“My deputy. Scheduler. Sometimes visiting staff.”
Ward’s voice became flat.
“Do you lock it?”
“Yes.”
“Every time?”
Hayes didn’t answer.
Briggs leaned back.
The room had redistributed embarrassment nicely.
I said:
“Don’t turn this into a password hunt yet.”
Ward looked surprised.
“Someone manually flagged me.”
“Yes.”
“That matters. But three real trainees were removed before I arrived. Were they manually flagged too?”
She checked.
Miller: algorithmic.
Ortiz: algorithmic.
Dunn: algorithmic.
Only mine was manual.
Different problem.
Same consequence.
I asked:
“What happens after R-17 appears?”
Ward showed us.
Instructor dashboards displayed an amber marker beside the trainee.
Additional monitoring suggested.
Nothing more.
“Where did Briggs get the washout list?”
Ward looked at him.
“I don’t send washout lists.”
“You send the morning risk report.”
“That is not a washout list.”
“You rank names by likelihood of attrition.”
“For resource planning.”
“You put percentages beside people.”
“Because medical, staffing, and remediation teams need to anticipate load.”
Briggs stared at her.
“And you thought instructors would see a trainee at eighty-one percent predicted attrition and treat that like weather?”
Ward’s face tightened.
“You were briefed repeatedly that the model was not evaluative.”
“I was also briefed that we needed this class smaller.”
Hayes looked up.
“What?”
Briggs pointed at him.
“Your directive.”
Hayes’s voice cooled.
“Readiness alignment is not an order to reduce the class.”
“Really?”
Briggs took out his phone.
Pulled up an email.
Hayes had written it six weeks earlier.
Current projected graduate count exceeds receiving-unit demand. Training leadership should ensure standards are applied with appropriate rigor rather than carrying marginal candidates forward out of inertia.
Hayes read his own sentence.
Ward looked away.
I didn’t.
“What did you mean?”
I asked.
“What it says.”
“No. What did you expect people to do differently after reading it?”
Hayes took his time.
“Stop giving repeated chances to marginal performers.”
“Who is marginal?”
“People who aren’t meeting standards.”
“Then why mention projected graduate count?”
His mouth tightened.
Because numbers had entered the room before standards did.
He knew it.
Ward said:
“The model was requested because command wanted better attrition forecasting.”
Hayes turned.
“Forecasting.”
“Yes.”
“Not selection.”
“I know.”
Briggs laughed again.
I looked at him.
“You enjoyed the list.”
His amusement disappeared.
“What?”
“You keep acting like you were a victim of bad guidance.”
“I was given data.”
“You turned it into destiny.”
“I still made decisions.”
“Exactly.”
He stared at me.
I continued.
“Miller passed her obstacle. You removed her because four seconds looked different after R-17.”
“She had other concerns.”
“Ortiz completed the route. You called the long option proof of low confidence because the system had already told you to expect poor judgment.”
“He barely made time.”
“He made time.”
Briggs stood.
“You weren’t here.”
“No.”
“I was.”
“That’s why I’m asking why your own observations lost whenever the model disagreed.”
For the first time, he didn’t answer defensively.
He looked at the table.
Ward said quietly:
“That isn’t how it was intended.”
I looked at her.
“Intent isn’t the interesting part anymore.”
We requested every R-17 trainee from the pilot period.
Twenty-six names.
Ten had been released.
The average release rate for unflagged trainees with similar raw performance?
Three out of twenty-six.
Not proof the model caused removals.
Enough to ask.
Then Ward found something stranger.
“How many instructors see the prediction percentage?” I asked.
“Senior instructors only.”
“Did trainees?”
“No.”
“Medical?”
“No.”
“Peer evaluators?”
“No.”
“So one group knew who the system expected to fail.”
“Yes.”
“And that group wrote subjective ratings that went back into the model.”
Ward stopped.
Hayes understood next.
“You trained the prediction using instructor observations.”
“Yes.”
“And the instructors could see the prediction.”
Ward’s face changed.
We had created a feedback loop.
The model told instructors who looked risky.
Instructors, knowingly or not, began interpreting ambiguous behavior through the risk flag.
Their interpretations became new data.
The system grew more confident.
Then the confidence made the next instructor more certain.
No one had to falsify a score.
The story could strengthen itself.
Briggs stared at the laptop.
“R-17 increased after week one?”
Ward checked.
“For some.”
“Miller?”
“Yes.”
“How much?”
“Forty-eight percent to seventy-six.”
“After my evaluation?”
“Yes.”
Briggs sat back.
The system had partly learned from his opinion, then returned his own opinion to him with a number attached.
He had treated the number as independent confirmation.
I looked at my record.
My R-17 had been manually inserted.
No underlying data.
Yet by the morning Briggs ripped off my badge, the system showed an eighty-four percent likelihood I would wash out.
I asked Ward:
“How?”
She opened the event history.
Once the manual flag existed, instructor observations began accumulating around it.
Words I had heard during the course:
Too detached.
Emotionally flat under stress.
Possible concealment of fatigue.
Unusual lack of peer dependency.
Every one of those observations could have been neutral.
After R-17, they became evidence.
Then I found the first comment.
Day one.
Before Briggs had ever spoken to me.
Entered by a senior evaluator named Colonel Peter Sloane.
Comment:
Subject presents as deliberately nonrepresentative of standard trainee population. Monitor for artificial performance management.
My pulse slowed.
Hayes read it.
“So Sloane suspected the cover.”
I looked at him.
“Who told Sloane I was coming?”
“No one.”
“Someone did.”
Ward searched the access log.
Sloane had opened my personnel packet twelve minutes after Hayes uploaded it.
Not my training performance.
My underlying administrative shell.
He had seen something that made him suspicious.
Then one hour later, Sloane accessed Hayes’s operations terminal.
No password breach required.
He had been sitting in Hayes’s office for a scheduling meeting.
And at 14:18, my manual R-17 appeared.
Hayes stared at the screen.
“Sloane.”
Ward nodded.
Briggs asked:
“Why would he flag her?”
I answered before anyone else.
“To see what you’d do.”
Hayes turned toward me.
I looked at him.
May you like
“Your evaluator was evaluating my evaluation.”
And unlike me, Colonel Sloane had never told anyone there was a second test running.
