Your words. Your proof.

No. witnessd captures keystroke timing intervals (the milliseconds between key presses), not the actual characters you type. Your document content is hashed locally, never transmitted. There is zero telemetry, zero cloud dependency, and zero way for us to read your work. The code is open source under Apache-2.0, so you can verify this yourself.

AI detectors analyze finished text and guess whether it was AI-generated. witnessd takes a fundamentally different approach: it records the physical process of writing -- timing, rhythm, pauses, revisions, paste events with origin attestation. It captures what you actually did, not what the output looks like. If you pasted AI-generated text and then extensively revised it, the evidence shows that: AI-assisted, not AI-authored. witnessd proves what it can prove (real time passed, a human was typing, the document evolved incrementally) and honestly declares what it cannot.

No measurable impact. Keystroke capture runs in a background thread with under 0.1ms latency per event. SWF proof computation happens asynchronously during natural pauses in your writing. CPU usage during active capture is typically under 1%. witnessd is designed to be invisible while you work.

We collect nothing. witnessd runs entirely on your local machine. The data it generates (inter-key timing intervals, checkpoint hashes, SWF proofs) is stored in a local SQLite database at ~/.witnessd/. No data is ever sent to our servers, because we have no servers. External timestamp anchoring is optional and only transmits a hash, never your content.

witnessd evidence is structured for FRE 902(13) self-authentication, and external timestamps from RFC 3161 authorities or OpenTimestamps provide independent third-party verification. However, no software can guarantee legal outcomes. Admissibility depends on jurisdiction, judge discretion, and how evidence is presented. What we can say: the cryptographic chain, behavioral data, and third-party anchors produce a significantly stronger evidence package than AI detection scores or personal testimony alone.

Our dual-layer detection cross-validates hardware input events against application-layer events. If someone injects synthetic keystrokes via software, the hardware event counts will not match, and the evidence is flagged immediately. Faking at the hardware layer requires kernel-level access and physical device modification, which raises the forgery cost to the point of impracticality for most threat models.

Yes. All evidence capture, SWF proof generation, and hash chain construction works with no internet connection. The only feature that requires connectivity is external timestamp anchoring (OpenTimestamps or RFC 3161 TSA), which is optional and can be done later when you come back online.

The open source CLI supports macOS (Intel and Apple Silicon), Linux (x86_64 and ARM64), and Windows (x64). Native App Store applications with GUI interfaces are planned for Mac App Store (Summer 2026) and Microsoft Store (Q3 2026).