Analysis14 August 20266 min read

Redacting a document without destroying its evidentiary value

The short version

Conventional redaction destroys the recipient's ability to distinguish a withheld passage from an edited one, which is exactly the distinction that determines whether the document still functions as evidence.
Committing to each element of a document separately, then publishing only the commitments for withheld parts, lets a recipient verify that everything shown is unaltered and that the hidden parts existed and have not been swapped.
This proves the remainder is untouched. It says nothing about whether the redaction was appropriate, and it cannot stop structure and metadata from leaking what was withheld.
The practical failure mode is not cryptographic. It is redacting the text while leaving it recoverable underneath, still the most common real-world breach of this kind.

You send a contract to a counterparty with the pricing schedule blacked out. They now hold a document you have demonstrably modified, with no way to tell which modifications were the agreed redactions and which were not. You have protected the confidential part by degrading the whole thing into an assertion.

For most correspondence nobody minds. In disclosure, due diligence, regulatory production or anything that might later be disputed, it is a real loss, and it is avoidable.

What redaction destroys

An unaltered document carries an implicit guarantee: what you are reading is what was written. Hash it, compare, done. Redaction breaks that guarantee globally rather than locally. Once one region is obscured, the recipient cannot verify any of it, because the hash no longer matches and there is no way to attribute the mismatch.

The recipient is left unable to separate three very different situations:

A passage was withheld, and everything else is as written.
A passage was withheld, and a sentence elsewhere was quietly reworded.
A passage that never existed was invented, and another was hidden to obscure it.

The point of redaction is to withhold. The problem is that withholding and altering become indistinguishable at exactly the moment someone has a motive to prefer one.

Commit to the parts, not the whole

The fix is to stop treating the document as one opaque blob. Commit to each element separately, each paragraph, cell, or field, and combine those commitments into a single value that identifies the document.

   root  =  H( H(A,B) , H(C,D) )
           /                  \
      H(A,B)                  H(C,D)
      /    \                  /    \
   H(A)   H(B)             H(C)   H(D)
     |      |                |      |
  para A  para B         [WITHHELD] para D
                          reveal H(C) only

Publish the leaf commitment for a withheld element instead of its content. The root still verifies, so the recipient can confirm the visible text is unaltered and that something specific was withheld from a known position.

The recipient recomputes the root from what they were given: the visible paragraphs plus the bare commitment standing in for the withheld one. If it matches the root that was signed, four things follow at once.

Every visible element is exactly as originally written.
The withheld element existed, in that position, at signing time.
It has not been substituted for a different one.
Nothing has been inserted or removed around it.

The recipient learns nothing about the hidden content, and gains the ability to rely on everything else. That is the trade conventional redaction fails to make.

Where this genuinely helps

Disclosure and production. Producing documents with privileged material removed, where the receiving side has every incentive to suggest the rest was massaged.

Due diligence. Sharing contracts with counterparty names or commercial terms withheld, while the buyer still needs to rely on the clauses they can see.

Regulatory submissions. Where a regulator receives a version with third-party personal data removed and must be able to treat the remainder as authoritative.

Journalism.Publishing a document while protecting a source’s identifying details, in a form where the authenticity of the rest is independently checkable rather than resting on the publication’s reputation.

The limits, plainly

It does not say the redaction was legitimate. Cryptography establishes that something was withheld from a known position; whether it should have been is a legal and editorial question. A verifiably redacted document can be a verifiably misleading one.

Structure leaks. The commitment tree reveals how many elements were withheld and where they sat. Often that is fine and occasionally it is the whole story, a recipient can see that you redacted precisely the three paragraphs following a particular heading. If the shape of the withholding is itself sensitive, this approach is the wrong one.

Granularity is a decision made in advance. Commit per paragraph and you can only withhold whole paragraphs. Finer granularity gives more flexibility and leaks more structure. There is no setting that is right for every document.

What we built

Redact seals a document so that individual elements can be withheld while the remainder stays verifiable, and the recipient checks it without needing us or an account. Reveal applies the same idea beyond documents, withholding rows or frames while still proving the original. You can watch the guarantee fail on purpose on the tamper demo, which is the fastest way to understand what it is actually claiming.