Security & privacy
No document vault. That's the whole pitch.
CaseSifter does not keep a long-lived copy of your case files. You use the web app to upload for a scan; anything sent for analysis is processed and then deleted. Here's exactly how data moves, tier by tier.
Where files live
Lasting copies stay with your firm. We do not run a central document library — analysis jobs hold file bytes only for the scan, then delete them.
What "ephemeral" means
A scan uploads the file bytes to an analysis job, processes them through the pipeline, and deletes the job directory on completion or after a time-to-live sweep — whichever comes first.
What we never do
We don't build a searchable archive of client documents, sell or share file content, or auto-file, auto-notify, or auto-escalate any finding to a third party.
Analysis tiers and what they send
Not every check needs the full exhibit. Deeper contextual analysis is opt-in, per file.
| Tier | What runs | What is sent for analysis |
|---|---|---|
| Tier 1 — Core forensics | Metadata, PDF integrity, structural, image forensics, OCR/text, and entity/timeline checks. | The file bytes for that scan job only, then deleted with the job. |
| Tier 2 — Entity verification | Generated yes/no claim questions checked against public record (e.g. "did this entity operate in that market on that date?"). | A short generated question — never the document itself beyond the scan job. |
| Tier 3 — Contextual document analysis | Comprehensive cross-document reasoning over the full exhibit text and layout. | The exhibit file itself, for that one analysis call, then deleted. |
Tier 3 is opt-in and off by default. For sealed, privileged, or otherwise restricted evidence, confirm your court's rules before enabling it on that file.
How findings are kept honest
The same controls that protect accuracy also limit what a model-sourced finding can do to a report.
Confidence discounting
AI-sourced findings are capped at medium severity, have confidence discounted roughly 30%, and are hard-capped at 60% confidence — never enough, alone, to turn a report red.
Labeled, not hidden
Every AI-generated finding is explicitly marked as such in the report. You always know whether a flag came from deterministic forensics or a contextual model call.
Have a specific compliance question?
Court rules on evidence handling vary. Tell us about your matter and we'll walk through exactly what would and wouldn't be sent for analysis.
Talk to us