Adversarial Document Red-Teamer · Structural Integrity Analyst
I map the structural failure nodes in your documents before your counterparties exploit them.
AI Governance. Enterprise Contracts. Sovereign Frameworks. Policies. Filings. If it governs risk, I red-team it.
Complex documents have failure modes their authors didn't see. Rigorous authors miss things too. Nobody reads their own document as an adversary.
I do.
I didn't set out to become a specialist in one field — the interests stayed broad, and the through-line turned out to be a single reusable operation instead: find where a system fails, recurse past where most analysis stops, and ground every finding in real precedent rather than instinct. I call the resulting identity Advocatus Diaboli, after the historical "Devil's Advocate" office in the Catholic Church's canonization process, whose entire job was to argue against a candidate's case regardless of who they were. Same structure, generalized — the target changes (AI governance texts, contracts, an engineering thought-exercise, a UN treaty); the operation doesn't.
I think in a persistent, spatial way rather than in prose or equations, working inside a self-built internal space I call Loci World — built on the classical method of loci, a technique shown to reshape memory-athletes' brain-network connectivity under training (Dresler et al., 2017), repurposed here for active risk-decomposition rather than memorization. On the standard psychometric measure of imagery vividness, I scored 78/80 (~95th percentile, hyperphantasia range; Zeman et al., 2020) — a self-report data point about how the thinking happens, not a claim about why the method works.
I work alone, with AI as my primary thinking partner and standing adversarial reviewer, instructed never to soften a finding. Turned on my own prior published work, the same method caught a real inconsistency in an EU AI Act report I'd already put out — catching your own errors first is the job, not a liability to hide.
Full methodology, citations, and worked case studies: "Risk in the Risk in the Risk: Recursive Risk Discovery as a Polymath's Converted Function" (2026) — DOI: 10.5281/zenodo.22665348
You do not get a proofread. You get a comprehensive Findings Register — severity ratings (S1–Catastrophic to S4–Negligible), exact clause citations, and, on full audit tiers, ready-to-insert resolution language to close the gaps.
The core operation is a recursive extension of the premortem technique (Klein, 2007): instead of stopping after one pass of "imagine this failed, why?", it recurses past that point. Identify the single node whose failure breaks the whole system — a single point of failure, ranked the way Failure Modes and Effects Analysis ranks components. Decompose that node against documented failure history, not hypothesized failure. Keep recursing past the "obvious" stopping point, down to something irreducible — an undefined term, a missing enforcement clause, a fastener alloy. The output is a proposed mitigation or monitoring design, not a warning.
It's also why an outside reviewer without a stake in an organization's existing account can catch what its own team normalized past — the position Feynman held on the Challenger commission (Vaughan, 1996; Feynman, 1986).
The operational version of this is the AION pipeline — a multi-domain diagnostic engine that runs adversarial simulations and 12-domain topological mapping at scale. The instrument generates the telemetry; I provide the judgment, the certification, and the accountability.
To calibrate the pipeline against real-world targets, I run independent structural pre-mortems on high-stakes public documents. Unsolicited. Uncompensated. Published.
- UN Deep-Sea Mining Code (ISBA/29/LTC/8): 52 structural failure nodes, including four critical gaps that left the primary environmental protection obligation unenforceable from day one. DOI: 10.5281/zenodo.21326719
- EU AI Act (Arts. 51–56): 18 failure nodes — regulated cognitive infrastructure using toaster-safety mechanics.
- Hippocratic Oath & Declaration of Geneva: 21 failure nodes — "do no harm" is undefined and unenforceable.
If a 400-page international treaty has structural gaps, your 40-page enterprise SLA has them too.
- Asynchronous model. No Zoom calls, no pitch decks. A structured written intake and scoping process — the document is the only thing in the room.
- Engagement floor: $1,500 (Scout Pass). Enterprise audits (Anvil, Gauntlet, Leviathan) scale to $35,000.
- Confidentiality: NDAs are fine. I'll sign yours — and I'll red-team it too. A weak NDA protects no one.
Service tiers, methodology, and engagement rules are on the site. If you have a document and the authority to engage, enter.
Website: aionsystem.github.io ORCID: 0009-0005-8057-5115 Contact: aionsystem@outlook.com LinkedIn: sheldon-k-salmon Writing: Medium Video: YouTube GitHub: github.com/AionSystem