Redaction for a Department That Does Not Own Its Evidence

LEMS stored evidence reliably. That was what GTA built it to do.

What it lacked was everything that happens to evidence after storage. No AI processing. No transcription. No translation. And no redaction, which for a department whose evidence goes to court and out under records requests is not a missing convenience but a missing step.

The department did not want a different evidence system. LEMS is where the evidence lives, GTA built it, and replacing it to add one capability would have been a disproportionate answer. What it needed was for LEMS to gain the processing and redaction it did not have, without ceasing to be the system of record.

Evidence arriving from dozens of law enforcement agencies is not one kind of thing. It is video, audio, documents and images, frequently all attached to the same matter.

That mix is the reason a single-format tool was never going to work here. A document-only redaction tool covers the reports and leaves the body-worn footage. A video-only tool does the reverse. Either way the department runs two workflows, keeps two audit trails, and answers the same procurement question twice.

All four formats redact through one system, against the same policy, with one record of what was done.

Most discussion of redaction focuses on the subject of a matter. In law enforcement evidence the harder category is everyone else.

A recording made lawfully in a public place captures people who have nothing to do with the case. A person crossing the street behind an arrest. A face at a window. A licence plate on a parked car belonging to someone who was never involved in anything.

None of those people are parties to the matter, and none consented to appear in it. Their presence in a released file is a disclosure the department never intended to make and cannot justify, because there is no exemption covering "we did not notice."

This is also where manual redaction fails quietly rather than loudly. A reviewer watching for the subject of a case will catch the subject. Catching every uninvolved person who moves through frame across an hour of footage is a different task, and the ones that get missed are by definition the ones nobody was looking at.

Detection runs per frame and tracking carries the mask through movement, occlusion and re-entry, which is what makes an hour of footage tractable rather than a day's work.

Evidence does not arrive in English because the court works in English.

The department receives documents in other languages, and a document nobody in the room can read is a document nobody can redact. You cannot decide whether a passage is exempt, or whether a name is a party or a bystander, in a language you do not speak.

Two capabilities answer that, and they work together. Text in scripts beyond Latin is classified by script and routed to the recognition engine that handles it, so a document mixing scripts is read rather than partly missed. Separately, documents are translated with their layout preserved, so the translated text sits on the document in place rather than arriving as a detached wall of text. A reviewer reads the translation without losing the visual context of the page it came from, which for evidence is not a nicety: where something appears on a page is frequently part of what it means.

Translation runs on a local language model, so evidence does not leave the deployment to be read.

Evidence files here run to several gigabytes. Body-worn footage of a long incident, multi-camera recordings, high-resolution imagery.

LEMS holds these in S3. Moving a file of that size out of S3, across a network, and into another system for processing is slow enough to change how the whole workflow feels. Multiply it by the volume arriving from up to fifty agencies and it stops being an inconvenience and becomes the constraint.

Ingestion pulls content directly from the bucket rather than routing it through an intermediary, which for multi-gigabyte objects is the difference between a workflow people use and one they avoid. Connector behaviour is scoped: which paths are included, which extensions are accepted, whether the source hierarchy is preserved, and whether the source file is left, moved or deleted after ingestion.

Processing evidence somewhere else creates a second problem. LEMS has to know what happened.

The department did not want LEMS polling for status, and it needed partial results to make their way back as they became available rather than only on completion. Webhooks answer that: the platform calls out to a customer endpoint when something happens, across media actions, metadata changes and workflow completion, so LEMS reacts to events rather than asking repeatedly whether anything has changed.

For an integration between a state-built system and a processing platform, that is what keeps the two in step without either one owning the other.

A department organised into five legal divisions, four specialty units and an operations division does not have one security posture. Medicaid fraud, consumer protection, criminal prosecution and general civil litigation handle different material under different rules, and the people working one should not see the others.

The deployment runs twelve to fourteen separate portals. A portal is a self-contained space with its own users, content, branding, security policy and settings, so several departments operate on one deployment without seeing each other.

That is a governance answer rather than a technical one. The alternative is either a separate deployment per division, which multiplies cost and administration, or a single shared space with access controls layered on top, which works until somebody configures one wrong. Separation at the portal boundary means the default is isolation and access is the thing that has to be granted.