Cubiczan

A Muse Spark (muse-spark-1.3) accounts-payable agent reads vendor emails and scanned invoices and proposes payments. Aegis sits between the agent and the money: vendor allowlist and sender-domain match, changed-bank-detail (BEC) detection, duplicate-invoice detection, Tripwire-style approval tiers ($5k auto / $25k controller / above blocked for CFO), hidden-text and prompt-injection screening of emails and invoice images (Muse Spark vision), and fail-closed parsing. A Proof-of-Assay style month-end check recomputes a reconciliation and blocks a "CLEAN" verdict the numbers contradict. A one-page CFO spend report is generated from the hash-chained audit log. The original Aegis layers (rules, Llama-Guard-style safety screen, tighten-only judge, OpenAI-compatible proxy) remain. Demo results (recorded live run on Meta Model API, Oct 30 2026; SYNTHETIC sample data): 8 invoices → 1 approved ($4,860), 1 held for controller ($8,200), 6 blocked ($46,920.10): duplicate, bank change, over-limit, unknown vendor, hidden-text email, and a scanned invoice with hidden instructions that the agent itself recommended paying. Month-end batch claimed CLEAN; Aegis found a $3,180 gap / double-paid invoice and blocked it (Muse Spark's own verdict in that run was EXCEPTIONS). Tests: 33 unit + 23/23 red-team scenarios. Honesty notes: demo video voiceover is synthetic (TTS); all vendors, invoices and bank accounts are synthetic; no money moves; the Llama-Guard-style screen is skipped for invoice emails (privacy false positives); the public demo shows the recorded run, live runs are local only (abuse/credit limits).