What the public sees.
What the public must never see.

A single body-cam recording from a busy street encounter may contain dozens of bystander faces — people walking past, standing in doorways, looking out of car windows, visible through store windows. Each face is a privacy obligation. Minors must be protected. Victims of sensitive crimes must be anonymized. Bystanders who were never involved in the encounter have no reason to appear in disclosed footage. Manual face tracking requires an analyst to identify each face, draw a redaction region, and follow that face frame by frame through the video — pausing every time the subject turns, is occluded by another person, or moves out of frame and back in. This is why 11 minutes of staff time produces one minute of redacted footage. Vault's Face Detection engine eliminates the manual process entirely. Deep learning models trained on law enforcement body-cam footage — with its unique challenges of motion blur, low light, fisheye distortion, rapid camera movement, and extreme angles — detect every face in every frame. Once detected, the engine assigns each face a persistent identity tracker that follows the individual across the video timeline. When a subject turns and their face is temporarily not visible, the tracker maintains the association using body pose, clothing color, and spatial continuity — re-attaching the redaction overlay the instant the face becomes visible again. When a subject walks out of frame and returns 30 seconds later, the tracker recognizes them and continues the existing redaction rather than creating a new one. The engine supports selective redaction: the analyst can mark specific individuals as "do not redact" — the officer, the suspect (if identified and charged), any individual whose identity the court has authorized for disclosure — while all other faces are automatically redacted. The resulting output shows every protected face consistently blurred from the moment it appears to the moment it disappears, with zero gaps that could reveal an identity in a single frame.

The Driver's Privacy Protection Act prohibits the release of personal information connected to motor vehicle records — and a license plate visible in disclosed footage is a direct link to the registered owner's identity, address, and personal information. Body-cam footage from traffic stops, parking lot encounters, pursuit sequences, and street-level interactions routinely contains dozens of license plates belonging to uninvolved vehicles. Dash-cam footage may contain hundreds. Each plate must be identified and redacted before disclosure. The challenge is that license plates in body-cam footage are rarely ideal targets for detection: they are viewed at oblique angles as officers approach vehicles, in motion as traffic passes, partially occluded by bumper stickers or trailer hitches, illuminated by headlights at night creating glare, and occasionally visible only in reflections. Vault's Plate Detection engine handles all of these scenarios. Models trained specifically on law enforcement footage detect plates in challenging conditions — at angle, in motion, partially covered, at night, in rain, and in reflections. Beyond plates, the engine detects VIN numbers visible through windshields (a common identification vector that most redaction tools miss), fleet markings and unit numbers on commercial vehicles, handicap placards with permit numbers, and parking permits displaying registration information. Each detection is tracked persistently across the video timeline using the same tracking architecture as the face detection engine — so a plate on a parked car visible in the background of a 20-minute encounter is redacted consistently across every frame without requiring manual intervention. For agencies using Automated License Plate Readers (ALPR), the engine cross-references detected plates against ALPR databases to identify any plate that appears in law enforcement records — flagging whether the plate belongs to a witness, a suspect, or an uninvolved party, and applying the appropriate redaction policy for each category.

Visual redaction protects what the camera saw. Audio redaction protects what the microphone heard — and what the microphone hears is often far more revealing. During a domestic violence response, the officer's body cam records the victim stating her full name, address, phone number, and a description of her injuries. During a traffic stop, the officer reads the driver's Social Security number and date of birth aloud while running a records check over the radio. During an interview, a witness provides the names and addresses of other people who were present at the scene. Every piece of spoken personal information is a privacy obligation that visual-only redaction tools completely miss. Vault's Audio PII Detection engine transcribes every audio track using speech-to-text models optimized for law enforcement audio — which means handling simultaneous speakers, radio chatter, ambient noise, distance from the microphone, accents, and code-switching between languages. The transcript is then analyzed by named entity recognition (NER) models that identify names, addresses, phone numbers, Social Security numbers, dates of birth, medical conditions, and other PII categories. Each identified PII instance is linked to its precise location in the audio timeline — not just "this file contains a Social Security number," but "a Social Security number is spoken between 04:17.342 and 04:21.118." The analyst reviews the flagged instances in the transcript, confirms or overrides each detection, and selects the redaction method: bleep tone (standard for voice redaction), silence (removing the audio segment entirely), or re-sampling (replacing the spoken PII with non-identifiable ambient audio from the same recording, maintaining the natural flow of conversation while eliminating the identifying content). Audio redaction is synchronized with visual redaction — so when an officer reads a name aloud while the name is simultaneously visible on a document in the video frame, both the spoken and visual instances are redacted together.

Body-cam footage captures far more than faces and license plates. Officers' body cameras routinely record the content of their Mobile Data Terminals (MDTs) — the in-car computers displaying suspect information, criminal histories, NCIC queries, and dispatch details. They capture documents that subjects hand to officers — driver's licenses, insurance cards, registration forms, prescription bottles with patient names, tax documents visible on kitchen tables during welfare checks, medical information displayed on hospital whiteboard during emergency room encounters. They capture phone screens when suspects show officers text messages, photographs, or social media content. They capture handwritten notes on notepads, chalkboards in interrogation rooms, and Post-it notes on refrigerators. Each of these text surfaces potentially contains PII that must be redacted before disclosure. Vault's Document & Screen Masking engine uses OCR (Optical Character Recognition) combined with NER (Named Entity Recognition) to read text content from any surface captured in video or image evidence and identify PII within that text. The engine detects MDT screen content through the fisheye distortion and low resolution typical of body-cam footage. It reads license plates on documents, Social Security numbers on tax forms, patient names on hospital whiteboards, prescription labels on medication bottles, and addresses on mail visible on countertops. When PII is identified, the engine applies a masking overlay to the text region — obscuring the identified content while leaving non-PII text visible where appropriate. For MDT screens, the engine offers a "full screen mask" mode that obscures the entire screen content rather than attempting to identify individual PII elements — a safer approach when the density of protected information on the screen is high.

A FOIA request does not arrive in isolation. It arrives with a deadline — California's AB 748 requires critical incident footage released within 45 days. It arrives with exemptions — specific categories of information that the law permits or requires agencies to withhold, each with its own statutory basis. It arrives in a department that may be processing dozens or hundreds of concurrent requests. And it arrives in a system where a single missed deadline, a single unredacted face, or a single unsupported exemption claim can result in litigation, settlement, and public embarrassment. Vault's FOIA Workflow engine manages the entire lifecycle. When a request is received, the engine assigns a tracking number, calculates the statutory deadline based on the applicable state law, identifies all evidence items responsive to the request, and queues them for automated redaction. Each redacted element is tagged with its FOIA exemption code — the specific statutory basis for withholding that information. A bystander face is tagged with the state privacy exemption. A license plate is tagged with DPPA. A medical condition visible on a hospital whiteboard is tagged with HIPAA. An undercover officer's identity is tagged with the law enforcement exemption. These exemption codes are embedded in the redaction audit trail, creating a defensible record that can be produced if the requester challenges any specific redaction. The workflow routes completed redactions through a supervisor approval queue, tracks the statutory deadline with escalating alerts as it approaches, and packages the redacted evidence for release with a cover letter documenting the exemptions applied. Every FOIA request — from intake to release — is documented in a comprehensive audit trail that the agency can produce in its entirety if the request is litigated.

Redaction and Brady compliance exist in a tension that most systems do not address. When a face is blurred in body-cam footage, the redaction protects the individual's privacy. But what if that blurred face belongs to a person whose presence at the scene is exculpatory — an alternative suspect, an alibi witness, someone whose identity would help the defense? What if a redacted document contains a notation that contradicts the prosecution's theory? What if a bleeped audio segment contains a statement that impeaches a prosecution witness? The privacy obligation and the disclosure obligation can directly conflict when the same evidence contains both protected information and potentially exculpatory content. Vault's Brady-Safe Redaction engine ensures that this conflict is surfaced and resolved before any redacted evidence is released. Before a redacted version is finalized, the engine cross-references every proposed redaction against the Brady/Giglio compliance flags generated by the Vault disclosure engine. If any redaction overlay intersects with content that has been flagged as potentially exculpatory — a face that appears near a Brady-flagged timestamp, audio content in a segment containing a Brady-flagged statement, a document that has been classified as potentially impeaching — the engine blocks the redaction and escalates to the ADA for review. The prosecutor then makes a judgment call: redact the privacy-protected information and provide the exculpatory content through an alternative disclosure mechanism (a written description, a court-ordered unredacted viewing under protective order), or leave the content unredacted with appropriate justification documented. The decision, the reasoning, and the outcome are permanently sealed in the audit trail — creating a defensible record that the agency considered the Brady implications of every redaction before releasing the footage.

The redaction challenge is not a single-file problem. It is a pipeline problem. A FOIA request for body-cam footage from a critical incident may reference 50 cameras, each recording an hour of footage. A department-wide body-cam disclosure policy may require processing every recording from every officer every day. A mass public records request from a media organization may demand footage from hundreds of incidents across a multi-year period. No manual workflow can handle these volumes within statutory deadlines. Vault's Batch Processing Pipeline treats redaction as an industrial process, not an artisan craft. The analyst defines the redaction parameters — which object types to detect (faces, plates, screens, audio PII), which individuals are exempt from redaction (officers, named suspects), and which FOIA exemption codes apply — and submits the batch. The pipeline distributes the files across parallel processing nodes, each applying the detection, tracking, and redaction algorithms independently. A batch of 147 video files totaling 84 hours of footage completes in approximately 2.4 hours — compared to the 924 staff-hours that manual frame-by-frame redaction would require. For large agencies, the pipeline runs overnight as an unattended process. Files are queued during the day as FOIA requests arrive, and the batch processes automatically during the overnight window when computational resources are available. By morning, the supervisor reviews the completed batch, approves the redactions, and the redacted files are ready for release — meeting statutory deadlines that manual processing could never achieve. The pipeline supports priority queuing — urgent FOIA requests (impending statutory deadlines, court orders, media-sensitive incidents) are promoted to the front of the queue and processed immediately, even during active overnight batches.

When a FOIA requester challenges a specific redaction — "Why was this face blurred? That bystander is a public figure. Show me the statutory basis for withholding their identity" — the agency's response must be more than "privacy." It must cite the specific exemption, explain why the exemption applies to this specific individual in this specific context, demonstrate that the redaction was reviewed and approved by an authorized person, and show that the agency evaluated whether the redacted content had Brady implications. Without this documentation, the agency faces litigation where the court conducts an in camera review — watching the unredacted footage themselves — and potentially orders the redaction reversed. Vault's Redaction Audit Trail creates this documentation automatically for every redaction in every file. Each redacted element is logged with: the object type (face, plate, screen, audio PII), the detection confidence score, the tracking method used (persistent identity tracking, manual region definition), the FOIA exemption code applied, the statutory citation supporting the exemption, the timestamp range of the redaction in the evidence file, whether the redaction intersected with any Brady-flagged content, the reviewer who approved the redaction, the reviewer's digital signature and timestamp, and the approval reasoning if a non-standard decision was made. This audit trail is WORM-sealed — no entry can be modified, deleted, or retroactively inserted after the redaction is approved. When a requester challenges a specific redaction, the agency produces the audit trail entry for that specific element — demonstrating the statutory basis, the approval authority, and the Brady evaluation — in seconds, not weeks of reconstructive research. The audit trail also functions as a quality control mechanism: supervisors can review aggregate redaction statistics to identify patterns that suggest systematic errors — an analyst who consistently fails to detect plates at certain angles, a detection model that underperforms in specific lighting conditions — and address them before they become compliance failures.