Research

Design reports and capability notes from the Attestrum project. Every claim is byte-reproducible; the full source for each lives in the public repository.

Explainers

How Attestrum Works, End to End A user walkthrough — from a folder of files to a proof anyone can check. READ → Deterministic by Construction Seven disciplines that make every seal byte-identical across re-runs — and the precise claim they support. READ → Provenance Without Disclosure Cryptographic training-data provenance that keeps the corpus private — a design-and-explanation report. READ → The Disclosure Floor and the Defensible Position How verifiable training-data provenance backs the new transparency mandates — a positioning report. READ →

Capabilities

Corpus-Version Diffing What changed between two sealed corpus versions — added / removed / unchanged, composition shift, both Merkle roots — re-checkable byte-for-byte. READ → Benchmark Contamination Scanning Did the eval set leak into the training corpus? A read-only scan flagging exact, near-duplicate, and embedded overlap. READ → Training-Content Summaries What a sealed corpus is made of — modality, source-type, license, and language mix, weighted by count and bytes. READ → Intra-Corpus Near-Duplicate Rate How much of a corpus is near-duplicated against itself — rate, cluster-size histogram, and example clusters. READ → Corpus Removal Evidence Inclusion in the earlier version plus non-inclusion in the later one — proof a document was taken down, verifiable with stock cosign. READ →