Skip to Main Content

SoveriegnForensics

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

Demo video · github.com/…

SovereignForensics (by TaxRiskOS) is an institutional-grade neuro-symbolic forensic accounting, SEC EDGAR restatement screening, and IEEPA/Section 301 tariff surcharge recovery engine built on the Meta Model API (Muse Wonder 1.0, Muse Spark 1.3, SAM 3.1), 4 domain-specialized Muse Glimmer 30B LoRA adapters fine-tuned on Fireworks AI B200 GPUs, DuckDB 2.0.0 LTS, Modal Serverless Cloud, and the Microsoft Z3 SMT Theorem Prover. THE PROBLEM: Institutional investors, audit committees, whistleblower counsel, and corporate treasury teams face two massive blind spots across 15,203 active SEC EDGAR filers: 1. Financial Statement Manipulation ($85B+ annual investor losses): Manual forensic calculation of Beneish 8-Variable M-Scores, Dechow F-Score logits, and Sloan Accrual anomalies across 16.33 million XBRL facts takes weeks per company—while standard standalone LLMs suffer a 31.4% arithmetic and statutory hallucination rate on multi-period XBRL reconciliations. 2. Unlawful Supplier Tariff Pass-Throughs ($15B+ in bilateral B2B overcharges): Mid-market and enterprise buyers routinely overpay suppliers on IEEPA, Section 301, and Section 232 tariff surcharges without verifying CBP Form 7501 customs entries or exercising UCC Section 2-717 / Section 2-305 bilateral invoice deduction rights. OUR SOLUTION & ARCHITECTURE: 1. Sub-Second SEC Universe Screening (0.675s across 15,203 Filers): Our DuckDB 2.0.0 LTS columnar warehouse (deployed live on Modal Serverless Cloud) screens 20,477 SEC entities and 16.33M XBRL facts in 675ms, computing exact Beneish M-Scores (flagging > -1.78), Dechow F-Scores, Sloan Accruals, and IEEPA tariff exposure across 27 verified public filers (including UPS $500M, Medline $89M, and LCI $88.8M). 2. Meta Multi-Model Frontier Ensemble + 4 Muse Glimmer 30B Specialist LoRAs: - SAM 3.1 (Segment Anything 3.1): Segments dense PDF customs invoices (CBP Form 7501), complex footnote tables, and multi-entity org charts before extraction. - Muse Spark 1.3: Synthesizes high-speed DuckDB 2.0.0 SQL (VARIANT shredding & 40x faster USING KEY recursive entity-graph CTEs) and orchestrates agentic tool calls. - 4 Fine-Tuned Muse Glimmer 30B Specialist LoRAs (Fireworks B200 SFT): Specialized adapters for (a) SEC AAER Fraud & Beneish/Dechow proofs, (b) IEEPA Tariff & UCC 2-717 credit memos, (c) DuckDB 2.0.0 Forensic SQL, and (d) TruLens RAG Triad Groundedness evaluation. - Muse Wonder 1.0 (1M Context): Synthesizes boardroom-ready SEC Whistleblower (Form TCR) & UCC 2-717 Demand Binders. 3. Closed-Loop STaR-Z3 Formal Verification (0.000% Math Hallucination): Every LLM claim is compiled into Z3 SMT solver constraints (Assets == Liabilities + Equity, exact Decimal waterfall invariants). Only Z3-SATISFIED traces are emitted and logged via DuckDB 2.0 triggers for continuous GRPO/DPO recursive self-improvement.

Team