Type a description into the person-search box and Flock converts both the phrase and the camera images into numerical representations, then ranks footage by how closely each image fits the text [5]. That pipeline returns an ordering rather than a match, and the officer decides where the list stops being useful. Smart Sort then lets the officer approve or reject returned images, and those votes reorder the results so footage resembling the approved images climbs [7]. That is relevance feedback pointed at people, and it converges on whatever the officer has already confirmed.
Vehicle search shows the underlying engineering difference. Vehicle search offers menus for color, body type, make, model and features, while person search takes free text [6]. A menu is a finite vocabulary. A city could enumerate it, test every combination, and write the results into an acceptance plan. Free text has no vocabulary, so screening has to move into a classifier, and the classifier's filters, thresholds and model instructions sit on Flock's servers, outside the interface departments get [8].
Treat the reconstruction the way you would treat someone else's benchmark table. It describes the client Flock ships, assembled from pre-login files that expose available searches, network calls, expected responses, controls and warnings [3]. It does not include department settings, or the models and instructions running server-side [4]. For it to describe your deployment, your department's configuration and the live thresholds would both have to match what the client implies. Nobody outside the company can check that: experts cited by WIRED said the screening model's error rate is not measurable from outside, nor is any skew toward permitting or rejecting particular descriptions [14]. The fullest public account of the control layer is a JavaScript bundle handed to anyone who loads the login page.
The telemetry runs the other way. A flagged search sends Flock the officer's wording, the category and confidence score the model assigned, and whether the officer continued, dismissed or disputed the warning [12]. Flock's statement to WIRED left the retention period, who internally can read the record, and whether it trains the moderation system all unaddressed [13][19]. The department that generates the log has no stated account of its lifecycle.
Meanwhile the interface hands the officer the downside. Flock warns that text-to-image results may be inaccurate or incomplete and should not be the sole basis for action, which places the inaccuracy risk on the person running the search [9], and leaves administrators to review flagged activity when abnormal use is detected [10]. Flock's August 13 safeguards announcement addresses some access and auditing problems, according to RuntimeWire [18]. The watchlist sits beside OS Investigate, which WIRED found carries 69 prepared prompts and code for 45 tools spanning six data categories including arrest records and commercial identity databases [15][20], and which Flock says is still in development [16].
The contract clause that would matter is one returning the blocked-and-warned query log to the buying city, with a right to submit test descriptions and see the disposition. Neither is in the client, because neither is in the client's half of the system.
