Research | Oversight Protocol

The research program behind Oversight aims to produce rigorous, peer-reviewable documentation of every design decision in the protocol. Each document below addresses a specific dimension of the system: its cryptographic construction, the resilience of its watermarking layers, or the performance characteristics of its seal and open operations. Where possible, results are presented with enough detail to allow independent reproduction. The long-term goal is a USENIX Security submission covering the full protocol.

Threat Model New

Honest companion to the protocol spec. Per-layer watermark survival matrix across screenshot, reformat, manual retype, and vocabulary-attack axes. L3 collusion and canonicalization limits. Passive beacons as telemetry rather than guarantees. Jurisdiction-by-IP as soft policy. What RFC 3161 timestamps actually prove and what they do not.

Protocol Design and Construction

Full technical treatment of the cryptographic construction, threat model, and security arguments. Covers both the classical (OSGT-CLASSIC-v1) and post-quantum hybrid (OSGT-HYBRID-v1) algorithm suites, the container format, watermarking system, transparency logging, and policy enforcement.

Watermark Resilience Analysis New

Empirical evaluation of per-layer watermark survival across eight attack classes, ECC error tolerance under the R=7 repetition code, content fingerprinting as a server-side fallback, Shannon capacity bounds at various paraphrase rates, and an honest assessment of LLM-based stripping via the SIRA attack. Cites 19 papers from 2001 to 2026.

Performance Evaluation New

Measured seal/open throughput (253 MB/s at 1 MB), per-layer watermark embedding overhead, extraction and L3 verification scaling, content fingerprint cost (3.35 s/MB via winnowing), file size overhead from zero-width expansion (5.2x), sub-100us ECC timing, and qualitative comparison with C2PA and Digimarc. All data from N=10 runs on Intel Core i7.