A vendor (software update distributor) and a customer (enterprise IT) have built up continuity across eight interactions. The continuity record at each side evolves with every legitimate exchange. Below: the record state on the customer side after each interaction.

Continuity record at step N is HMAC-SHA-256(record_{N-1}, fingerprint_N || obfusc_N || content_hash_N). The chain is one-way; rewinding requires the inputs at every step, which requires the relationship’s history.

Each interaction’s content yields a fingerprint that is stable for the relationship (same patterns produce the same fingerprint) and not directly invertible (fingerprint does not reveal content). The construction combines bigram statistics, token-length distribution moments, and a relationship-keyed hash. The fingerprint is one of three KDF inputs to the continuity step.

Each step contributes an obfuscation input that is not directly observable from interaction logs. Generated from a relationship-keyed PRG seeded at enrollment. Defeats the offline-replay variant where an attacker has logged communications — logs alone do not yield the obfuscation input.

Five transport paths carry one cover message plus four chunk-bearing packets. None are distinguishable from each other on casual observation. The cover is plausible business correspondence. The chunk packets look like opaque transit traffic.

Bernard’s recipient client receives the cover and chunks. σ_session is retrieved from local Secure Enclave (set at enrollment, never on the wire). Recovery: select k=3 chunks, order by σ, attempt authenticated decryption.

RECEIVED

RECOVERED REAL MESSAGE

Silent-failure attack mode: wrong σ returns the same error as “insufficient chunks.” The attacker gets no oracle signal.

Suppose Bernard is compelled by legal authority to disclose received content. He can comply with a partially-honest disclosure:

The compliance is harder to disprove than with simple threshold sharing. The real message remains hidden.

Endpoint compromise. If Bernard’s device is compromised AND σ_session is exfiltrated before session end, the recovered message is exposed.

Forward secrecy across credential compromise. Compromise of registry-stored long-term enrollment credentials may permit retroactive decryption depending on protocol parameters.

Anonymous operation. Both endpoints are identified to the registry. Registry metadata (which endpoints, when, what duration) is available to legal process. This is by design — the architectural answer to dual-use concerns.

Voice-channel embodiment. Written cover serves voice: at the sender, voice content is reduced to a written-form transcript that is embedded in a written cover; at the recipient, voice is locally reconstructed from the recovered text using pre-registered sender voice profile. The wire carries no voice or text of the real content — only the cover correspondence. Latency and quality constraints are operationally stringent and the demonstration build is queued.

Video-channel deniable communication with synchronised state attestation across multiple parties. Architectural design parallels voice-channel: written cover carries video reconstruction parameters; recipient locally reconstructs from pre-registered profile. Synchronised state attestation handles multi-party video where each participant must reconstruct a coherent shared video state.

Persistent storage on conventional storage media (disk, SSD, cloud-storage backend). Provides authenticated deniability with the disk-resident format indistinguishable from generic encrypted storage under casual inspection. Cover-content is selected from the user’s authentic data, per-document KDF derives from cover identifier and content hash, protected content is embedded steganographically into authentic cover documents.

Volatile-memory variant for in-process secret protection. Memory contents survive cooperative-verification challenge but become unrecoverable after process termination. Hardware-enclave integration: Intel SGX, AMD SEV, ARM TrustZone.

Virtual mailbox credential storage. Pseudo-email-accounts on conventional providers carry credential split. Inter-account correspondence has cover-coherence (the messages between accounts read plausibly), credentials are embedded steganographically in selected messages, account-level threshold distribution spreads the credential across N independent providers. Addresses the password-manager attack class: compromise of one vault does not yield the credential population.

Self-cooperative architecture across user devices. A single operator runs cooperative-verification across their own devices (phone + laptop + secure-element token). Provides quorum-based access from operator’s own devices without requiring a counterparty. Mutual device authentication on recovery; cooperative revocation flow on device loss.

Dedicated storage-device hardware integrating the ARIA-STORAGE primitive at the firmware level. Embedded controller, secure element, hardware RNG, cover-content flash region, embedded-protected-content flash region, sideband authentication channel. Default mode presents host file-system view as ordinary cover content; authenticated mode presents the extended view with protected content recovered.

Trusted-custodian configuration for institutional contexts where a regulated custodian holds part of the cooperative-verification quorum (legal escrow, financial trustee, regulated data fiduciary). Day-to-day access path: user portion only. Extraordinary access path: user + custodian cooperation under registry-recorded terms.

Attacker calls the treasury function impersonating the CFO with a cloned voice. Requests urgent wire transfer to attacker-controlled account.

The defence is architectural — not detection-based. The treasury function does not need to detect the voice clone; the absence of authentication is the signal.

Attacker has working credentials. PKI / MFA accepts them. Standard IAM provides no architectural defence against legitimate-credential lateral movement.

The per-customer continuity record is the SolarWinds amplification killer — the attacker has to compromise each victim individually rather than achieving one-to-thousands amplification.

All free-text → AI surfaces in SoafAii applications route through SCIM (Semantic Confidence Intent Mapping) for analysis, clarification, and brief generation before reaching any generation model. The demo’s single free-text surface is the Section 5 “Try your own scenario” input — that input goes through SCIM’s analyze and brief endpoints with graceful fallback if unreachable.

No browser-to-Anthropic call. Every AI invocation routes through the studio-brain proxy (Cloudflare Worker holding the API key as a Wrangler secret). The model string used by this demo is fixed: claude-sonnet-4-20250514.

The Health badge in the header is the SoafAii Health module — aviation-grade tiered reporting (GREEN/AMBER/RED) with fingerprinting, periodic checks, self-healing for DOM regressions, and security isolation. Click the badge or press Ctrl+Shift+Q for the embedded QA panel.

First-launch detection via aria_mehler_tour_seen_v1 in localStorage. Two-tier picker on first launch: 5-step quickstart for the productive-in-five-minutes path, or 14-step full walkthrough for the comprehensive orientation. Re-openable at any time from the (?) help button in the header. ESC or click outside to dismiss.