Add com.writerslogic.text-fingerprint.1 (fingerprint, text)

[dcondrey]

Add com.writerslogic.text-fingerprint.1 (fingerprint, text)

This PR adds a new entry (identifier 40) for a fingerprint-type soft
binding for text content.

Algorithm

com.writerslogic.text-fingerprint.1 computes a durable 256-bit fingerprint
from the words of a document (rather than embedding hidden markers into it):

1. Normalize to a single character stream: Unicode NFC; remove zero-width /
   formatting characters (U+200B, U+200C, U+200D, U+FEFF, U+2060, variation
   selectors U+FE00–U+FE0F); lowercase; collapse whitespace runs to a single
   space; strip punctuation; trim.
2. Character 4-grams: overlapping character 4-grams over the normalized
   string (sliding window of 4 chars, step 1; a single n-gram of the whole
   string when it is shorter than 4 chars). Character grams keep a single-word
   edit local, so the fingerprint stays stable on short text where word shingles
   moved too far.
3. SimHash-256: SHA-256 each 4-gram to a 256-bit vector; for each bit
   position sum +1/−1 across grams; the final bit is 1 when the column sum
   is > 0. Output is 32 bytes, hex-encoded.
4. Match: Hamming distance ≤ 32 bits (12.5%) indicates the same content
   under light edits.
5. Windowed fingerprints: the normalized stream is also split into
   overlapping windows of 512 characters with 50% overlap (step 256) and each
   window fingerprinted separately. The soft binding records the whole-document
   fingerprint as block 0 (empty scope) and one block per window
   (scope: {start, length}), so an extracted excerpt or truncated copy can be
   matched against a window block even when its whole-document fingerprint has
   drifted.

Why a computed fingerprint (vs. an embedded ZWC watermark)

The fingerprint is derived from the visible words and recorded in the
agent-signed C2PA manifest, so it is non-destructive (nothing is added to
the document), normalization-proof (reformatting, case, whitespace, and
punctuation changes happen before hashing), ZWC-immune (normalization
removes zero-width characters and variation selectors, so injecting them does
not change the fingerprint), and forge / transfer-resistant (the value is
bound by the manifest signature and cannot be lifted onto another document
without re-signing).

Honest limit: robust to edits and formatting (including a single-word edit
on a one-sentence snippet, thanks to character 4-grams), not to paraphrase.

This algorithm is part of the WritersLogic CPoE proof-of-effort authorship
attestation system, alongside the existing com.writerslogic.zwc-watermark.1
entry (identifier 29).

Checklist

• Entry conforms to the schema and includes all mandatory fields.
• File remains valid JSON (python -m json.tool).
• Next free integer identifier (40) assigned; no collisions.
• Submitted by an individual affiliated with WritersLogic.
• informationalUrl provided.
• Commit signed off (DCO).

[mrappard]

I would approve this if (https://writersproof.com/cpoe/text-fingerprint) had more information.

[domguinard]

Thanks a lot for the entry, we reviewed this during the WM TF meeting and would ask if you could add some information on the reference page as it is currently blank. No need for a great deal of details, a marketing page would do for instance.

[dcondrey]

Thanks for the review. The reference page is now published at https://docs.writerslogic.com/soft-binding/text-fingerprint with the algorithm details — normalization, the 256-bit SimHash over character 4-grams, the windowed scoped blocks, the match threshold, and limitations — and the entry's informationalUrl points there. Happy to expand it further if useful.

[domguinard]

Thanks a lot for the documentation page, please approve the ID change as ID 40 is already in the queue.
With that approval we should be good to go.

[domguinard]

I was able to accept the change myself.

[domguinard]

Merged, thanks a lot for your entry @dcondrey.