Add com.writerslogic.text-semantic.1 (draft)

[dcondrey]

Draft, held pending publication of cross-platform int8 test vectors. Adds com.writerslogic.text-semantic.1 (fingerprint): a 256-bit semantic fingerprint via a pinned open embedding model under fixed int8 inference and seeded random-hyperplane LSH, so paraphrased or translated text still resolves to its manifest. A soft-binding value must be reproducible by any verifier, so this entry is not ready to merge until deterministic int8 inference is proven and the test vectors are published; it is opened as a draft to track that work. Follows PR #57 (ids 42-44); final identifier to be assigned on merge. Spec: https://docs.writerslogic.com/soft-binding/text-semantic

[domguinard]

@dcondrey are you ready for us to review the entry or is it superseded by: #60

[dcondrey]

Not superseded. #60 was a scoping fix to an existing entry (id 43's robustness claim), not a replacement for this one — #58 is a distinct new entry.

Not ready for review yet.

[dcondrey]

Closing in favor of a better approach. The semantic fingerprint captures meaning similarity, not origin specificity — it would fire on independently-written content covering the same topic, making it a weak provenance discriminator. The right class for paraphrase- and translation-robust text soft binding is post-hoc semantic embedding watermarking: a key-conditioned signal injected via semantic-preserving substitutions, detected by statistical hypothesis test, that carries origin evidence rather than just measuring topical similarity. Opening a replacement PR for com.writerslogic.text-watermark-semantic.1 based on dual-embedding watermarking (DEW, arxiv 2606.31602, June 2026), which survives paraphrase and translation without requiring generation-time access.