Exhume
private beta · invite only · Sydney
single-tenant · your region private beta

Redacted is not removed.

ex·hume /ɛksˈhjuːm/ v. — to dig up what was buried; to bring to light after concealment.

The same page often goes out twice, redacted twice — differently. One copy hides the name; the other never got the memo. Exhume pairs every version of every document in a production and reads the difference — then a human decides what’s real.

Opens your mail app — it reaches a person, not a list.

350,000+ documents paired 100,000+ spans recovered 32 min to pair 83,000 documents

The same page, produced twice.

Different passes miss different things. Exhume lines the versions up, reads what one bar covers and the other forgot, and hands the difference to a reviewer — every span in one of four states.

AS PRODUCED DOC-004481 · 02/26

17. On or about February 18, defendant Marcus W. Doe met with co-conspirator Anton P. Vesely at the offices of Halbright Capital.

18. Funds were subsequently routed through Meridian Trust (Curaçao) and held on deposit at Verdant Holdings SA until 2 April.

SIBLING · EARLIER PASS DOC-004466 · 02/26

17. On or about February 18, defendant Marcus W. Doe met with co-conspirator Anton P. Vesely at the offices of Halbright Capital.

18. Funds were subsequently routed through Meridian Trust (Curaçao) and held on deposit at Verdant Holdings SA until 2 April.

One lasso, a hundred decisions.

The reveal map plots every recovered span in the matter by meaning — names cluster with names, OCR garble forms its own island, and a lasso resolves the lot in one gesture. Gold is a candidate, green confirmed, red dismissed: the same vocabulary everywhere you look.

PRIVATE BETA
EXHIBIT · BUILD 0.9 · THE REVEAL MAP · one dot = one recovered span seeded demo matter · every name fictional

Built for the piles nobody can read.

Anywhere the same records go out more than once — redacted by different hands, in a hurry, at scale. Tens of thousands of pages, a meaningful fraction of them behind bars.

FOR NEWSROOMS + FOIA DESKS

The agency answered the same request twice. The releases don’t agree.

Different requesters, different reviewers, different years — the same memo comes back wearing different bars. Feed Exhume the whole release history and it reads what one pass missed, across all 40,000 pages — not just the folder you had time for.

FOR LITIGATION TEAMS

Their production is 80,000 documents deep — half of them versions of each other.

Point it at what opposing counsel produced, and every inconsistent redaction surfaces with its evidence pair attached. Then point it at your own production before it ships. Same engine, both hats.

FOR ARCHIVES + OVERSIGHT

Declassified in 1998. Declassified again in 2011. Not identically.

Court unsealings, declassification reviews, inspector-general releases — public record that disagrees with itself. Exhume finds the disagreements at corpus scale; a human confirms what’s real before anyone quotes it.

The machinery.

No magic in the pipeline. Every stage is inspectable, and the parts that learn, learn from you — not from the internet.

01 / PAIR

Every version meets itself

Whole-corpus embeddings and clustering, rebuilt from scratch on every ingest — because the document you add today can be the sibling that cracks one you already had. Nothing pairs by filename.

02 / RECOVER

Read the difference

Three detection layers on every page, four OCR engines raced per document family, recovery methods cross-checked against the version’s own headers. High recall on purpose; noise is handled where it’s reversible.

03 / REVIEW

A human keeps the receipts

Rank, lasso, confirm. Gold is a guess, green is a fact, and every decision lands in a hash-chained audit log your opposing counsel can read.

04 / LEARN

It learns your matter

Every confirm and dismiss trains a small model, per matter, that re-ranks what’s left — noise sinks, real reveals rise. On demand, a vision model reads an evidence pair both ways and cites what it finds. Neither ever decides for you.

ENGINE 83,000 documents ingested in ~32 min on 16 cores · resumable
HOSTING single-tenant, in your region · hosted by us on AWS in Sydney, or deployed into your own tenancy — it needs a container runtime, Postgres and object storage, nothing else
OCR four engines, scored per document family · the best reader wins the page · re-reading is never assumed to help
PAIRING similarity proposes, positive evidence confirms · shared boilerplate alone never makes a pair
MODEL trained on your reviewer’s labels, per matter · it ranks, it never resolves
AUDIT hash-chained log of every decision · exports carry SHA-256 manifests

Where it stands.

Exhume is in private beta, invite only. Not a mockup: the engine has paired and recovered across real document productions, and the review surface in the exhibit above is the one the work gets done in.

WHAT EXISTS TODAY

The whole pipeline — ingest, pairing, recovery, a review queue with a hash-chained audit trail, and exports that carry their own SHA-256 manifests. Running against public productions totalling over 350,000 documents.

WHAT WE WANT

A small number of matters to run end to end alongside the people who own them. You supply a production; we pair it, recover what the versions disagree about, and hand back the confirmed reveals with the evidence pair behind each one.

WHAT IT COSTS

Nothing, for now. There is no price while the beta is finding its shape, and we would rather learn what this is worth from matters we have actually run than guess at it in advance.