One weapon detector, two kinds of error
Within one year, one camera-based weapon-detection product missed a handgun in a school shooting and flagged an unarmed student in another district, and armed police responded. The paper uses the two records to show a guard that errs in both directions.
The miss: Nashville, January 2025
On 22 January 2025 a student at Antioch High School in Nashville shot and killed a classmate. The school's camera-based weapon-detection system, Omnilert, did not detect the handgun. The district attributed that to where the cameras were, and the system "was activated by police brandishing their weapons during the response".
Less than three weeks later the district proposed to "lease and install Evolv systems at all 18 high schools", a second guard made of scanners. The board of education voted on that lease, so the second guard came in by an explicit human decision. The paper places it at step 8 of its ratchet (an external bound) rather than counting it as a guard added inside the loop.
What those scanners had done elsewhere was already public. In a complaint of November 2024 about a different school, the Federal Trade Commission alleged that the same product "reportedly failed to detect a seven-inch knife", and that afterwards "school officials increased the system's sensitivity settings, prompting a 50% false alarm rate". In the paper's words, turning the dial away from one error moved it to the other, and people turned it.
By September 2025 the scanners were installed in all of the district's high schools, and the camera-based system was still installed across all of its campuses. Nothing was retired.
The false alarm: Baltimore County, nine months later
Nine months after the shooting, Omnilert flagged a student at a Baltimore County school as possibly armed. A county official asked how "we had police officers with guns drawn approaching a kid because of a bag of Doritos".
Maryland's Inspector General for Education corrected that account. The system "did not detect a single blue potato chip bag as a threat"; the alert rested on "the student's hand movement, including its position and angle" and an object at that angle "resembling a firearm". The review also found that:
- the system "generates thousands of alerts daily", of which about two hundred reach the vendor's staff and about twelve a day reach the school system;
- fifteen people received the notification of 20 October;
- school and police staff "immediately reviewed the information and deemed the threat invalid" and sent a cancellation, yet the alert protocol continued "out of what was described as 'an abundance of caution'";
- "the detection system and monitoring protocols operated within the designed parameters".
The review could not verify that staff training had taken place, for lack of attendance logs, and recommended reviewing how many people receive each alert.
Two separate records, not one cause
The paper is explicit: the Nashville and Baltimore County records are separate, the scanners were added in Nashville only, and no common cause of the two errors is claimed. What links them is the product and the pattern.
Two-sided error, in the paper's terms
Proposition P1 states that guards which judge artifacts rather than outcomes produce false rejections and false acceptances in the same period. The paper reads five of its propositions in these schools:
- one product erred in both directions (P1);
- the miss was answered with a second guard, which a board voted to lease (P3 in shape, step 8 in authority);
- escalation went on after a cancellation (P4);
- the system worked within its parameters while an unarmed student faced drawn weapons (P5, a counterfeit victory);
- any one of the fifteen recipients could restart the response (P6, where the human sits).
Proposition P1 also cites a hospital record of two-sided error: a sepsis model that missed 67% of sepsis cases in the same cohort in which it alerted on 18% of all patients.
Physical AI oversight without a frontier model
None of these guards is a large language model, and the paper says so. Two-sided error "does not need a model". The point is architectural: the paper's concern is that frontier models are now being placed in stacks of the same shape, and that models add definitions that change and controls that cost nothing to add.
These cases document parts of the paper's definition rather than the whole. Each mechanism is documented somewhere in the paper's records, but how often all three conditions meet in one component "is not settled by the records here".
Full sections: schools, hospitals, grids and mines (6.4), Propositions P1 to P6, why a body changes the question (6.1).
Chernets, V. (2026). AI-Watchbird (Sheckley): When Automated Oversight Widens Its Own Mandate and Harms What It Guards. SSRN 7473658. https://doi.org/10.2139/ssrn.7473658
Vadym Chernets also builds Poly A1, an app in which AIs from different companies check each other's work.