WritersProof Security Model

Attack: An attacker copies existing text by retyping it character-by-character to generate synthetic keystroke evidence.

Mitigation: SWF (Slow Work Function) timing proofs enforce minimum computational delay between checkpoints. Behavioral fingerprinting detects unnaturally consistent typing cadence. The appraisal verifier measures SNR (signal-to-noise ratio) and Lyapunov exponents — retyped text produces detectable statistical regularity.

Cost to defeat: Requires physically retyping the entire document with natural-looking timing variation, pauses, and corrections. SWF proofs force minimum real-world time. Behavioral analysis flags cadence anomalies.

Attack: An attacker pastes AI-generated or copied text and makes minor edits to simulate a writing process.

Mitigation: Dual-layer keystroke capture (IOKit HID + CGEventTap on macOS) distinguishes typed input from clipboard paste events. Large pastes are flagged and recorded with exact byte counts. The checkpoint chain shows the ratio of typed vs pasted content.

Cost to defeat: Easy to attempt but immediately detected. Paste events are recorded at the OS input layer — they cannot be disguised as typed input.

Attack: An attacker manipulates system clocks or evidence timestamps to fabricate when writing occurred.

Mitigation: Evidence is anchored to external timestamp authorities: RFC 3161 TSA receipts and OpenTimestamps (Bitcoin blockchain). These are independently verifiable and cannot be forged without controlling the timestamp authority or rewriting the blockchain.

Cost to defeat: Would require forging RFC 3161 TSA signatures or rewriting Bitcoin blocks. No known practical attack exists.

Attack: An attacker modifies an exported evidence packet to alter the recorded process.

Mitigation: Every checkpoint is hash-chained (SHA-256). Modifying any checkpoint invalidates the chain. The export is signed with the author's Ed25519 key. The CBOR envelope includes algorithm version and commit ID for reproducible verification.

Cost to defeat: SHA-256 collision resistance makes selective modification computationally infeasible. Ed25519 signatures prevent wholesale replacement.

Attack: An attacker claims someone else's evidence as their own or generates evidence under a false identity.

Mitigation: Each user has a unique Ed25519 keypair generated at `writersproof init`. The identity is expressed as a DID (Decentralized Identifier). Private keys never leave the local machine. Evidence packets are signed with the author's private key — verification confirms the signature matches.

Cost to defeat: Requires possession of the victim's Ed25519 private key. Key extraction requires physical access to the machine or compromise of the user's keychain.

Attack: An attacker reuses a legitimate evidence packet for a different document.

Mitigation: Each evidence packet includes the SHA-256 hash of the final document content. Verification checks that the hash matches the document being claimed. The hash chain is document-specific — it cannot be transferred to a different file.

Cost to defeat: Easy to attempt but immediately caught. The document hash won't match the replayed evidence.

Attack: An attacker attempts to skip or accelerate the Slow Work Function proofs to compress fabricated evidence into a shorter timeframe.

Mitigation: SWF proofs are calibrated to the local hardware at init time. Each proof chain is verified by recomputing the sequential hash iterations. There is no shortcut — the work must be performed sequentially (not parallelizable). The verifier measures expected vs actual computation time.

Cost to defeat: Would require hardware significantly faster than what was calibrated, AND the ability to forge the calibration record. The sequential nature of the hash chain prevents GPU/parallel acceleration.

Attack: An attacker uses automation to simulate human-like typing patterns while inputting pre-written text.

Mitigation: The behavioral fingerprint captures inter-key timing, error correction patterns, pause distribution, and cognitive load signals. These form a multi-dimensional profile that is unique to each writer. Automated input shows detectable statistical regularity in Lyapunov exponent analysis.

Cost to defeat: Requires not just simulating typing speed but reproducing the specific author's pause patterns, error rates, correction behaviors, and cognitive load signatures. No known tool can replicate a full behavioral fingerprint.