Passenger assault.
Model:
Seat-access conflict.
Simone stopped.
“Hey, Aaron?”
The analyst across from her pulled off one headphone.
“What?”
“Did they change the severity model?”
“When?”
“Last release.”
“I think thresholds.”
“Can you send notes?”
Aaron shrugged.
“Probably in deployment folder.”
She opened release documentation.
Version 8.4.
The model had been retrained on fifteen years of historical airline complaint data.
Accuracy claims were excellent.
Ninety-six percent agreement with “resolved final classifications.”
That phrase bothered her.
Resolved by whom?
She pulled training methodology.
The model treated the last saved classification in each historical case as ground truth.
Not the original employee report.
Not investigation result.
Final classification.
Simone leaned back.
Her mother’s voice returned.
Numbers hide things quietly.
She searched older archives.
One phrase appeared repeatedly in legacy datasets.
Premium service dissatisfaction.
Another.
Guest compatibility issue.
Another.
Service disagreement.
She had seen these categories in the historical training appendix.
Something about them seemed familiar.
She searched the consortium’s ethics library.
A case study appeared.
Flight 417.
Twenty-seven years earlier.
Senior Director Amara Washington.
Simone had heard the name.
Every aviation compliance program still taught some version of the case.
Usually the simplified one.
Passenger assaults regulator.
Airline discovers VIP complaint suppression.
Reforms follow.
Simone opened the full case study.
She read until lunch.
Skipped lunch.
Kept reading.
Maya Ortiz’s original report.
Edited report.
Rachel Avery’s testimony.
Martin Keene’s emails.
Amara’s line:
If identical conduct can be moved from the first box to the second because the passenger is valuable, your data becomes fiction.
Simone stared at the sentence.
Then at the model methodology.
Historical final classifications were being treated as objective truth.
The model had learned from records known to contain years of softened language.
Not only one airline.
Industry-wide legacy datasets had similar practices.
The system was not intentionally protecting wealthy travelers.
Nobody had coded:
If premium passenger, downgrade misconduct.
The model did something subtler.
It learned correlations.
Premium customer.
First Class.
High loyalty status.
Corporate booking.
Certain complaint language.
Certain locations.
Certain staff descriptions.
Historically, cases with those features had been softened more often.
The model learned softening as “accuracy.”
Twenty-seven years after Amara helped expose humans rewriting reality, software had inherited the rewritten reality and automated it.
Simone felt sick.
She walked directly to her manager.