“How strongly?”
“Modestly.”
“Then?”
Elena leaned forward.
“The model is penalizing people before they are late.”
Thomas paused.
That sentence landed.
He looked toward the audit director.
“Can workers challenge the score?”
“In theory.”
“How?”
“Through staffing coordinator escalation.”
“Did coordinators know?”
“No.”
Thomas looked at Elena.
“So practically, no.”
“Correct.”
He rubbed his jaw.
“What do you recommend?”
“Freeze that variable immediately.”
“Recalculate all active temporary-worker scores.”
“Review whether other proxy features produce similar effects.”
“Restore lost shift priority where appropriate.”
“Then assess back-pay or opportunity-loss compensation.”
Thomas frowned.
“Back pay for shifts people did not work?”
“Maybe not legally wages.”
“But if our system unfairly reduced their opportunity, there may be remediation.”
CFO objected.
“How would you calculate hypothetical hours?”
Elena looked at him.
“Carefully.”
Thomas remained quiet.
Another director said:
“This may be an overreaction.”
Elena turned.
“Maybe.”
The director continued.
“The model is doing exactly what we bought it to do.”
“Predict absence.”
“Public transportation affects attendance.”
“That’s not bias.”
Elena answered:
“Predictive usefulness does not automatically make a factor appropriate.”
“Why?”
“Because we are not pricing insurance.”
“We are deciding which worker gets offered enough shifts to pay rent.”
“That doesn’t mean we ignore reliability.”
“No.”
“It means we should ask whether there is a more direct measure.”
Thomas looked at the data scientist.
“Is there?”
The analyst nodded.
“Actual individual punctuality history predicts better than neighborhood transit dependency once enough personal work history exists.”
“Then why use transit probability?”
“Cold-start accuracy.”
“For new workers.”
“Yes.”
The room became silent.
The system was using neighborhood-level prediction mostly when workers had less personal history.
The people with the least opportunity to establish themselves were being judged most heavily by a proxy.
Thomas said:
“Turn it off.”
The CFO looked at him.
“Immediately?”
“Yes.”
“We may increase no-shows.”
“Then manage no-shows.”
The director who objected earlier asked:
“Are we making this decision because everyone is afraid of repeating Naomi Carter’s story?”
Thomas looked at her.
“No.”
“Good.”
“I’m making it because we have a weaker variable that burdens people before we have evidence about them personally.”
He tapped the report.
“If Naomi Carter had never existed, I would still have to explain why this is defensible.”
That ended the debate.
The variable was disabled.
The entire model went under review.
Brianna Cole never knew a board meeting had happened.
She received a message from workforce compliance.
Your scheduling concern has been substantiated in part. Your reliability score has been recalculated. You are eligible for restored priority beginning next scheduling cycle. We are reviewing whether you lost work opportunities due to the prior model.
Brianna read it three times.
Then walked to the staffing coordinator.
“What does ‘substantiated in part’ mean?”