Digital Evidence Management Problems: The 7 That Cost Real Money, and How to Avoid Them

Why does retention as a dropdown happen so often?

It happens because it is the only design that can be built without touching another agency's system, and that makes it the path of least resistance in every scoping meeting.

Computing retention properly means the evidence platform has to know whether charges were filed, whether the case is closed, whether an appeal window is open, whether a civil claim has been noticed and whether a public records request or a litigation hold is freezing the clock. None of that lives in the video system. It lives in the records management system and in the prosecutor's case system, which belong to different teams with different priorities and different procurement cycles. Building a live link to them is slow political work with no visible output for weeks, so the project ships a category picker instead and calls the integration phase two.

The consequence is not theoretical. A traffic stop with no arrest and a homicide have completely different fates, and the person choosing the category at upload is an officer at the end of a shift. Categories get picked by habit. A whole class of evidence then ages out on the wrong schedule, and the failure only becomes visible when a defence request or an appeal arrives and the file is gone.

The fix is to make case linkage a first release requirement, not a later phase. Retention is computed from case type and live case status, holds are objects that freeze the clock and require a documented release, and every automatic deletion writes a record naming the rule that authorised it. Write the test into the statement of work: place a hold on a case, run the deletion job, and prove nothing was removed. If a proposal cannot describe how that works on day one, retention will end up as a dropdown no matter what the brochure says.

What goes wrong when evidence is migrated off an incumbent platform?

This is the line most often missing from budgets and the one that most often stops a project.

Your existing platform holds originals, metadata, case associations, audit history, redaction versions and sharing records. What you can get back depends entirely on terms agreed years ago, and the answer is frequently less than people assume. Originals usually come out. Audit history often does not come out in a usable form. Redaction work product, meaning the regions and reasoning behind a released version, may not be exportable at all, which means every previously released clip has to be redacted again if it is ever released again.

Then there is volume. Moving petabytes is not a copy operation, it is a project with egress charges, throughput limits and a verification obligation, because a migration that does not hash and compare every file before and after has not preserved anything.

Three things protect you. First, ask your current vendor in writing, before you sign anything new, for a full export of evidence with metadata and audit history in an open format, and time how long they take to answer. That response is more informative than any reference call. Second, budget migration as its own workstream with its own contingency, not as a task inside the build. Third, insist the migration writes a verification report: file count, hash comparison, unmatched items and a named owner for every exception. Agencies that discover the export terms late end up paying twice, once to leave and once to rebuild what could not come with them.

Why do the ingest and records system integrations break after launch?

Because the evidence that causes trouble is not the evidence anyone tested with.

Axon Evidence, Motorola CommandCentral Evidence and NICE Investigate all handle their own hardware and the common formats well. The material that breaks a pipeline is the private security system export in a proprietary container that plays in nothing, the doorbell clip a witness sent as a compressed message attachment, the interview room recorder installed a decade ago, and the extraction report from a phone. Each of those is real work, each arrives unpredictably, and a build scoped around body camera and in car video will meet all of them in the first month of live use.

The failure mode is quiet. A file that cannot be transcoded gets stuck in a queue, a technician works around it by saving a copy to a network share, and evidence now exists outside the system with no hash, no retention clock and no audit record. That is exactly the file a defence expert will ask about.

Records system integrations break differently. Case status is the input that drives your retention, and it changes in a system you do not control, sometimes through a bulk update, sometimes through a schema change during an upgrade. A one way nightly pull that silently returns nothing looks identical to a night when no cases changed.

Design for both. Every ingest path ends in either a stored original with a hash or a visible failure with an owner, never silence. A technician path exists for oddball media so nothing lives on a share. And the case feed has a heartbeat: if no update arrives within an expected window, someone is told, because a stalled feed means retention clocks are running on stale status.

What happens when redaction throughput and disclosure logging are underbuilt?

Redaction is usually scoped as a feature and it is actually a staffing model.

Automatic detection of faces, plates and screens has genuinely improved, and Veritone Redact and similar tools do real work. What it does not do is remove the human, because the person applying a redaction is asserting a legal exemption. If the build treats detection as the deliverable and the human step as an afterthought, throughput barely moves. The gains live in the correction loop when the tracker loses a face behind a car door, in batch handling when the same uninvolved bystander appears across six clips of one incident, and in attaching an exemption code to each region as part of the workflow rather than typing justifications into a separate log afterwards.

Disclosure is the mirror image. Sharing is easy to build as a link and hard to build as evidence. What a prosecutor needs in front of a judge is a record of exactly what was shared, with whom, when, what they were able to see, and every view and download since. If your build produces a link with an expiry and no per recipient scope or access log, you have built convenience and left the defensibility out.

Both fixes are workflow decisions taken early. Redaction is measured in clips per technician per week and designed against that number. Disclosure is a first class object with scopes, expiry, optional watermarking and a complete access log, treated as a primary feature rather than as reporting.

Should you build custom or configure what you already own?

For a single agency running one camera line at moderate volume, whose prosecutor is happy with how evidence is shared today, the answer is configure and we would say so before quoting. Axon Evidence and Motorola CommandCentral Evidence are convenient, the bundle works, and a custom build will not beat them on that footing.

The same applies if you have never asked your current vendor what your platform can already do. Retention categories, sharing scopes and audit exports are frequently configurable and frequently left at whatever was set at implementation. Have someone audit that before commissioning a replacement, because rebuilding to obtain behaviour you had not switched on is an expensive lesson.

And do not build if nobody can own an operational system that holds evidence. This data has legal consequences and needs a named responsible owner with budget, not a project that ends at handover.

Build when the prosecutor's office is funding it jointly with the agencies that feed it, because a shared county platform is the case no camera vendor serves well. Build when most of your problem evidence is not camera evidence. Build when retention must follow live case status and today it follows a dropdown. Build when redaction has become a staffing crisis. And build when your storage cost is set by a vendor and you have never seen it broken down by retention state.

How do hidden costs get into the quote?

Four items account for most of the surprises, and storage is only one of them.

Storage modelled as a build cost. It is a permanent operating line, and it grows every year as cameras and resolution increase. Model it annually from your actual recorded hours and retention profile in your own cloud account before committing to anything. The offsetting opportunity is real: most retained video is never watched again, and archival tiers are priced far below tiers that serve instant playback, so lifecycle rules driven by retention state pay for themselves. That only works if you hold the account.

Ingest sources counted as one line. Each strange format is separate work, and the count grows after launch as new sources appear.

Transcription and search treated as a toggle. Running speech to text across your recorded hours changes the compute profile substantially and should be priced as its own decision with a stated volume.

Security work assumed to be free. Criminal justice information security requirements are engineering scope from day one, covering access control, attributable views, key handling and who on the development team can see anything. Retrofitting attribution after launch means touching everything.

What separates a build that works from one that fails here?

Ask how they prove a file has not been altered. The answer should involve hashing at ingest, immutable storage of the original, derivatives that reference rather than replace, and an audit record kept separately from the file store. If integrity is described as file permissions, they have not thought about a defence expert.

Ask how retention interacts with a litigation hold and a pending public records request. Both freeze clocks, both require a documented release, and both must override automatic deletion. Any design where a deletion job can outrun a hold is disqualifying and no other strength compensates for it.

Ask what redaction looks like for a technician doing thirty clips a week. Detection quality matters less than the correction loop, batch handling of a repeated subject, and how exemption codes are captured and reported.

Ask who holds the cloud account and how storage tiering is expressed, then get code, infrastructure and data ownership in writing before kickoff. At Digital Heroes it is the agency's from the first commit. Your cheapest first step costs nothing: ask your current vendor for a full export of your evidence with metadata and audit history, and time the answer. That single request usually settles the build question on its own.