Aperture LegalAperture Review Engine Request a pilot
Litigation document review · on‑premise

Find where the documents disagree.Without a page leaving the firm.

Aperture reads the entire production and returns every contradiction and every fact — each one tied to the exact words in the source. It runs on hardware the firm owns. Nothing is uploaded; nothing leaves the room.

Zero bytes leave the box 98.3% of claims cite a verified quote Cross‑document, not per‑document
Exhibit — one contradiction, as returnedContradiction · temporal
Plaintiff — email 12 Mar 2019
“…the shipment left our dock on the 12th, tracking to follow shortly.”
cited · DOC‑00417 · span verified ✓
Defendant — deposition ¶14
“We did not ship until the 19th. I am certain of the date.”
cited · DEPO‑002 ¶14 · span verified ✓
The airtight part

The review never leaves the building.

It runs on your own server, fully offline, with a bundled local model. No documents are uploaded. No analysis is metered. And your IT does not have to take our word for it —

disconnect the network and run the pilot anyway. Nothing leaves.

That is a check anyone in the building can run — which is also why the client‑confidentiality analysis is short. The documents that walk in are the only ones in the matter, and they stay on the box you control.

air‑gapped deployments verify a signed offline licence on the box · zero outbound calls · confirmable by your own packet capture

AIR‑GAPPED PERIMETER YOUR SERVER signed licence · checked offline NO OUTBOUND 0 bytes leave
the network cable can be unplugged — the pilot still runs
What it does

One narrow job, done to evidentiary standard.

You point it at a pile of discovery documents; it extracts every factual claim, checks each against a verbatim quote that must literally exist in the source, and then finds the places those claims collide.

01 Cross‑document contradictions

Seven types — factual, temporal, measurement, attribution and more — connected across the production, not ranked within one document. As far as we can tell, no major platform ships this at all.

02 Grounded facts only

Every claim carries the exact words from its source, or it is dropped rather than shown. On the shipped build, 98.3% of emitted claims carry a citation that survives re‑verification.

03 A findings memo, in Word

Both sides of every contradiction, cited by document, party, date and quote — plus position timelines (“who knew what, when”) and a per‑subject fact file.

04 The formats matters arrive in

PST/OST mailboxes, MSG and EML with nested attachments, Concordance load files, spreadsheets, Slack and chat exports. Privilege flagging, dedup and threading run alongside.

It is not a review platform — no productions, no Bates stamping, no review UI. Keep the one you have. This runs beside it and does the part that, as far as we can find, none of them do.

Findings Memorandum
Prepared as a summary under FRE 1006 · Matter 24‑cv‑0918 · 3,932 docs
¶1 — Ship dateContradiction · temporal
Pl. “…left our dock on the 12th.” DOC‑00417
Def. “…did not ship until the 19th.” DEPO‑002 ¶14
¶2 — Notice of defectContradiction · attribution
“QA flagged the unit before shipment.” EML‑1183
“No one raised a concern to me.” DEPO‑004 ¶61
Position timeline — who knew what, when
Feb 28
QA memo
Mar 12
“shipped”
Mar 19
actual ship
Apr 03
denial
The live demo · open & read‑only

Not a video. The shipped engine, running on real documents.

The demo is the exact product we deploy — opened read‑only on public corpora, with no sign‑up and no sales call. Click into any finding and follow it to the precise quote in the source it came from.

Enron corpus

The full pipeline on 4,202 real EDRM Enron documents — 297 of them scanned PDFs read by OCR — processed end to end into 1,179 issues, every contradiction cited on both sides.

Financial filings

Four real SEC 10‑Ks — Hertz, Molson Coors, Marvell, Mattel — with reported‑vs‑restated figures pulled straight from each filing’s own reconciliation tables. Deterministic, and nothing trained or tuned on the company.

Read‑only, and no login — the sign‑in page even runs an “air‑gap check” proving it pulled nothing from outside the server.

Matters · demo.aperturelegal.comlive
Enron Showcase — 4,202 docs
EDRM corpus · incl. 297 OCR’d scans · litigation
litigation1,179 issues
Hertz Global Holdings
10‑K FY2014 · restated FY2011–13
restatement19
Molson Coors Beverage Co.
10‑K FY2018 · restated FY2016–17
restatement8
Marvell Technology Group
10‑K FY2007 · restated FY2003–06
restatement20
Mattel, Inc.
10‑K FY2019 · restated FY2016–18
restatement18
read‑only · nothing is uploaded · the shipped engine, on public data
Claims you can verify

Every number names its set. The misses sit in the same file as the wins.

95/100caught
Planted contradictions caught, end to end
SET‑10 · 440 documents, sealed and both‑sides codenamed, scored through the real front door. Five misses are named in the file.
benchmarks/set10_score.json
90·94replicated
Blind adversarial sets — replicated
Two independent 100‑needle sets, each authored by a different model family and spliced into real correspondence. Quoted as “90+ replicated,” never the single high number.
BENCHMARKS-2026-07.md
98.3%grounded
Emitted claims with a verbatim citation that survives re‑verification
Shipped extraction build. Failures are dropped, not displayed. A rate, not a guarantee — spot‑check it on your own matter.
training/ROLLBACK-REGISTRY.md
zerobytes out
Bytes of your documents that leave the box
Verifiable by your own network monitoring. Disconnect the interface and run the pilot; the signed licence is checked offline, on the box.
your own packet capture
We audit ourselves, and publish what breaks

On our last internal audit, every headline number was recomputed from the artefacts on disk. Most reproduced exactly. Three were withdrawn or restated — one of them a headline comparison. That audit ships with the pilot, unedited. We would rather hand you our worst review than have your diligence team write it for us.

The offer

Test it on a closed matter where you already know the answer.

Seal your list of what mattered before we start, and open it at the read‑out. The validation runs on your hardware, on your corpus — you grade the recall, not us. If it also surfaces contradictions your team missed, that is the conversation worth having.