Meridian and revised the order of the procurement review. Instead of opening with the contractor's polished performance presentation, the committee heard directly from the team that maintained access records. One technician described duplicated databases and correction requests sitting unanswered for weeks. Victor asked why management had never received the information. The technician produced an email chain showing that management had received it repeatedly. Nobody spoke for several seconds. Leila thanked him for bringing the records and asked what resources he needed. He proposed a dedicated maintenance window and a clear escalation owner. The cost was minor. The embarrassment was not. Evelyn later congratulated Leila not for revealing a failure but for refusing to turn the technician into a prop in somebody else's speech.
* * *
The hidden fee in the Sentinel proposal led Leila to a small software company called Harborpoint Analytics. Its product promised to categorize security incidents and predict unauthorized access attempts. In practice, a demonstration showed that its scoring system treated repeated stops as evidence of success without evaluating whether those stops were justified.
Leila arranged a meeting with the developer, a weary engineer named Jonah Mills. He seemed relieved to talk to someone interested in more than the sales presentation. The original design, he explained, had been a simple reporting tool. Customers later demanded a single performance score, and sales teams had converted raw activity into a measure of safety.
“Could the system distinguish an actual threat from a person whose badge was checked?” Sofia asked.
“Only if the officer selected the correct outcome,” Jonah said. “And if the outcome affected the officer's bonus, you can imagine the incentives.”
“Did you warn anyone?”
“I filed objections. They're in the change history.”
He provided the documents after consulting his lawyer. Leila read them that night. One entry described the risk of encouraging unnecessary confrontations. Another predicted that staff would be pressured to inflate numbers. Both had been marked low priority.
The revelation complicated the investigation. A human being had treated another human being with contempt in a lobby, but hundreds of remote decisions had made such encounters more likely. Software had not forced Rebecca to touch Adrian or reject his badge. It had rewarded the posture she adopted and allowed managers to mistake that posture for diligence.
Harborpoint Analytics occupied a suite above a dentist's office, its sign printed on a sheet of white plastic that curled at one corner. The founder, Jonah Pierce, spoke about frictionless access and predictive risk with the enthusiasm of someone who had never waited outside a locked office door. Leila asked for a demonstration using a fictional visitor whose name was misspelled on an appointment. The software assigned a risk indicator and displayed a recommendation to seek supervisor approval. 'What if the supervisor is unavailable?' she asked. Jonah showed her a queue-management feature that raised the case's priority after fifteen minutes. Fifteen minutes might be nothing to a salaried executive, Leila replied, but it could cost a courier an entire delivery route. She asked to inspect the training data and learned that the program had been optimized against historical decisions by Sentinel guards. If those decisions were inconsistent, the software would turn inconsistency into a mathematical baseline. Jonah insisted the model didn't use demographic categories. Priya answered that a system could reproduce unequal outcomes through the labels it learned to predict, even without being handed those categories directly. Rather than making a dramatic accusation, they requested an independent technical review. The company also suspended plans to purchase the software until the vendor could document error rates for different kinds of legitimate visitors.
Jonah Pierce asked for a second meeting after Harborpoint's technical review found serious weaknesses in the data used to train its risk model. He arrived without his usual slide deck. Instead he carried a plain report showing how minor spelling errors and incomplete appointment records could trigger higher suspicion scores. He admitted that his sales team had presented the software as more certain than the evidence justified. Priya respected the admission but asked what would change for customers already using the product. Jonah proposed clearer confidence warnings and more extensive manual review. Leila insisted that any automated risk indicator be treated as a prompt to verify, never as independent proof of misconduct. She also requested a documented mechanism for people to challenge incorrect data. Jonah hesitated at the cost and complexity. For the first time he seemed to understand that frictionless access for an administrator could mean considerable friction for the person standing at the gate. Vale Meridian declined to buy the current version. Months later Jonah sent a revised technical paper and thanked the team for asking uncomfortable questions. Leila saved the message without promising a future sale. The company's responsibility was not to reward a vendor for admitting a problem; it was to purchase systems that could be shown to work.
* * *