How it works
Six engines. One read of the whole case.
CaseSifter runs every filing and exhibit through six independent forensic checks, then cross-references the results against everything else in the matter. Here's what each engine catches, how accuracy is controlled, and what stays a human decision.
The pipeline
Each engine is narrow and explainable on purpose — no single black box decides a document's fate.
| Engine | What it looks for |
|---|---|
| Metadata | Timestamp anomalies, authorship, editing-software fingerprints that don't match the document's claimed origin. |
| PDF integrity | Embedded JavaScript, hidden attachments, layered content, and unapplied or improperly flattened redactions. |
| Structural | Page size, orientation, and resolution consistency across a scanned or reprinted exhibit. |
| Image forensics | Error Level Analysis and copy-move/clone detection to flag uneven recompression or splicing — with heatmaps. |
| Text / OCR | Invisible text, micro text, and mismatches between the embedded PDF text layer and what OCR actually reads off the page. |
| Contextual logic | Entities, dates, and amounts checked for internal consistency and, where enabled, cross-checked against public record. |
Catching AI-generated and AI-edited exhibits
Generative tools have changed what a fabricated document looks like. It no longer has to be a clumsy cut-and-paste job — a convincing bank statement, utility bill, or affidavit can be produced from scratch, or a real one edited without leaving an obvious visual seam. CaseSifter doesn't run a single "is this AI" classifier; it maps specific engines to the traces that kind of fabrication actually leaves behind.
| AI-fabrication pattern | Engine that catches it |
|---|---|
| Inpainted or generatively edited image regions | Image forensics — Error Level Analysis and copy-move detection surface localized recompression and cloning artifacts. |
| Text inserted or overlaid on a scanned template | Text / OCR — mismatches between the embedded text layer and what OCR reads off the visible page, plus invisible/micro-text checks. |
| Metadata inconsistent with the claimed origin | Metadata — editing-software fingerprints and timestamp sequencing that don't match a document's supposed history. |
| A fluent but fabricated narrative or figure | Contextual logic — entities, dates, and amounts checked against the rest of the case file and, where enabled, public record. |
These are indicators, not a verdict. CaseSifter cannot certify that a document was or wasn't produced with AI — it surfaces the forensic and contextual signals worth an attorney's attention, with the source language quoted so you can verify it yourself.
Gotcha Finder: the case-wide pass
Most contradictions don't live inside one document — they live between two. Gotcha Finder reads across every filing in the matter and surfaces the same fact stated two different ways.
What it flags
Element, procedure, credibility, and damages conflicts that would actually move a motion — a default date that shifts between an affidavit and an exhibit, a stamp year that conflicts with the body text, an amount that doesn't reconcile across two filings.
What it ignores on purpose
Exhibit-lettering drift between filing and recording, caption or NYSCEF header noise, and loan-status number drift that's expected after default. Signal, not noise — so the list of what needs a look stays short.
Accuracy: how confidence is controlled
Litigators need to know what a finding actually means before they rely on it. Here's the honest version.
Risk bands, not verdicts
Every scanned document gets a green / yellow / red risk level. Red requires at least 70% confidence in the underlying findings — no single low-confidence flag can turn a report red on its own.
AI findings are capped
Findings sourced from contextual AI analysis are capped at medium severity, have their confidence discounted, and are hard-capped at 60% — and are always labeled as AI-generated in the report.
Nothing auto-escalates
CaseSifter does not file anything, notify a court, or contact opposing counsel. A flagged finding is a prompt for attorney review, with the source language quoted so you can verify it yourself.
What this adds to your practice
The value isn't the scan — it's what the scan lets your team stop doing by hand.
For the attorney
Walk into a hearing already knowing where the record contradicts itself. Catch a fabricated or altered exhibit before it's relied on in a motion, not after opposing counsel raises it.
For the client
Fewer surprises mid-litigation, faster turnaround on document-heavy motions, and a firm that can show its diligence — all without a single case file leaving the building.
Bring a real file to the demo.
We'll run it through all six engines live in a walkthrough, so you can see exactly what surfaces before deciding anything.
Book a demo