The False Positive#
The system flags a pattern on Thursday morning. A 2015 Honda Accord, dark blue, California plate 7RJH423, visits 2847 Jefferson Avenue three times per week. Monday, Wednesday, Friday. Arrival between 10:15 AM and 10:40 AM. Departure between 12:05 PM and 12:35 PM. Duration: approximately two hours. The pattern has held for five weeks.
Jeff pulls the footage. The driver is male, mid-fifties, white, average build. He parks in the side lot. He walks to the building entrance. He enters. Two hours later, he exits. He walks to his car. He drives away.
The building is a single-story commercial structure. Painted beige. Small parking lot. No signage visible from the street-facing camera. Jeff switches to the drone footage from two weeks ago. The aerial angle shows a small placard near the door, but the resolution is insufficient. He zooms. The image pixelates. He applies a sharpening filter. Still unreadable.
He opens the county assessor’s database. 2847 Jefferson Avenue is owned by Sequoia Healthcare Properties LLC. The building is zoned commercial. The business license on file is for “Sequoia Medical Plaza.” No further detail.
Jeff cross-references the driver. The license plate traces to a Gregory Amherst, age fifty-four, address in San Carlos. Jeff runs the name through LinkedIn. Gregory Amherst, Senior VP of Operations, Riverstone Technologies. Stock photo headshot. The bio lists twenty-three years in tech sector logistics. A small company. Series B funding. No obvious connection to Redwood City municipal contracts. Nothing legible, at any rate.
But the pattern is there.
Three times per week. Same building. Two-hour duration. The timing is wrong for a standard office visit. Too long for a meeting. Too short for a workday. Jeff has seen this before—or supposes he has. A second office. A side project. Somewhere you go when you do not want to be seen going there.
He pulls the footage from the past five weeks. He trims it into a supercut. Gregory arriving, Gregory leaving. Monday, Wednesday, Friday. The repetition makes it look deliberate. The deliberateness makes it look hidden.
He writes the caption.
Riverstone Tech VP visits unmarked Jefferson Ave building 3x/week, 2hr visits. Same schedule for over a month. What’s in that building?
He queues the post. He does not publish yet. He has a protocol. Cross-check, verify, wait six hours, review again. If it still looks solid, publish.
He waits.
At 4:00 PM, he opens the assessor database again. He tries a different search. Instead of the address, he searches “Sequoia Healthcare Properties LLC.” The search returns a portfolio: fourteen properties in San Mateo County. Medical offices, outpatient clinics, urgent care centers. He clicks through the list.
2847 Jefferson Avenue is listed as “Sequoia Dialysis Center – Redwood City.”
Jeff stops.
He opens a new browser tab. He searches “Sequoia Dialysis Center Redwood City.” The first result is a facility page on a healthcare network website. The page lists hours, insurance accepted, services provided. Hemodialysis. Peritoneal dialysis. Vascular access management.
There is a photo. The beige building. The side lot. The small placard by the door.
Jeff closes the browser.
He goes back to the footage. He watches Gregory walk from his car to the building. Gregory is wearing a jacket. It is a warm day, but Gregory is wearing a jacket. Jeff zooms in. Gregory’s gait is slightly uneven. Not a limp. Just careful. The carefulness of someone managing discomfort.
Jeff deletes the queued post.
He opens his Reddit DMs. He scrolls to the conversation with u/ClearSightCA. ClearSightCA is a collaborator. They have been helping Jeff identify people in the footage. Matching faces to LinkedIn profiles, cross-referencing license plates, confirming addresses. ClearSightCA is good at this. Fast. Thorough.
Jeff scrolls up. Two days ago, he sent ClearSightCA a screenshot of Gregory in the parking lot. The message read: Know this guy? Regular visitor, suspicious timing.
ClearSightCA replied four hours later: Running it. Looks familiar. Give me a day.
Jeff scrolls further. Yesterday afternoon, ClearSightCA sent a follow-up: Gregory Amherst, Riverstone Tech. San Carlos resident. No public records issues. Clean. Why suspicious?
Jeff replied: Pattern visits. Same schedule. Tight window. Could be meeting someone.
ClearSightCA replied: Want me to dig into the building?
Jeff replied: Not yet. Still gathering.
Jeff closes the DM window. He opens it again. He begins typing a new message.
Disregard Amherst. False positive. Building is a medical facility.
He hovers over Send.
He deletes the message.
He types a new message.
Amherst confirmed clean. Moving on.
He hovers over Send.
ClearSightCA is already online. The green dot appears next to their username. A new message arrives before Jeff can send.
Got a screenshot of the building entrance if you want it. Clear shot of the guy walking in. Good for composite.
Jeff reads the message twice.
He types: No need. Confirmed non-target.
ClearSightCA replies immediately: You sure? I already pulled it. Can send now.
Jeff types: I’m sure.
Three dots appear. ClearSightCA is typing. The dots disappear. They reappear. They disappear again.
ClearSightCA: OK. Let me know if you change your mind.
Jeff closes the laptop.
He sits at the kitchen table. He thinks about Gregory Amherst. He thinks about the two-hour window. Monday, Wednesday, Friday. He thinks about the jacket on a warm day. He thinks about the careful gait.
He thinks about the screenshot ClearSightCA already pulled. The clear shot of the guy walking in. Good for composite.
ClearSightCA did not say they deleted the screenshot.
Jeff opens the laptop. He opens the DM window. He scrolls up to confirm the conversation. The messages are there. The screenshot offer is there. He starts typing.
Delete the Amherst material. It’s not relevant.
He hovers over Send.
He deletes the message.
ClearSightCA is helping. ClearSightCA is, pretty much, part of the project. You do not micromanage a collaborator. You do not tell them to delete their own research files. That implies distrust. That fractures the working relationship.
Jeff closes the laptop again.
He walks to the window. Outside, the street is quiet. A car passes. The footage is still recording. The cameras are still running. The system is still flagging anomalies. By morning, there will be six new patterns to review.
Jeff sits back down at the table. He opens a text file on his desktop. The file is titled “Error Log.” It currently contains three entries. Equipment malfunctions, mostly. A camera that lost calibration during rain. A relay that dropped packets during a firmware update. Standard operational issues.
He adds a fourth entry.
Entry 4 – Pattern ID #1847 (Amherst, G.): Flagged as suspicious based on visit frequency and duration. Subsequent review identified location as medical facility (dialysis center). Pattern consistent with treatment schedule. Classification revised to: Non-target. Error type: Insufficient context during initial assessment. Correction: Enhanced location verification protocol before pattern flagging.
He saves the file.
He opens his engineering textbooks folder. He searches for a PDF he read during his undergraduate program. It takes three minutes to find. The title is Failure Analysis in Structural Systems: A Practitioner’s Guide. He opens it to Chapter 7: Acceptable Risk and Error Margins.
There is a highlighted passage on page 142.
No inspection system achieves zero false positives. The goal is not perfection; the goal is an error rate low enough that corrective mechanisms can function before systemic failure occurs. In practical terms, an error rate below 2% is considered operationally acceptable in most municipal and commercial contexts.
Jeff closes the PDF.
He opens the Error Log file again. He reads Entry 4. He calculates. He has flagged approximately two hundred patterns since the system went live. One error in two hundred is a 0.5% error rate.
Well below acceptable threshold.
He closes the file.
He does not think about Gregory Amherst again.