Our public research library combines market analysis, academic work, and implementation-oriented synthesis. The common thread is architectural: what changes when AI agents stop assisting work and start executing it.
The strongest non-brand search surface on this site is the research library itself. These pages are static, detailed, and intentionally opinionated. Use this archive when you want the full set instead of only the featured studies linked from the homepage.
A US export-control order removed Claude Fable 5 from every foreign national 4 days after launch. It cost us nothing, because nothing depended on it. Concentration risk read as an engineering problem: match each task to its tool, keep data in your own architecture, and run an owned-hardware floor no order can reach.
A US export-control directive disabled Fable 5, a leading frontier model, for every foreign national on 13 June 2026. A scenario had placed the same mechanism in 2029. Why single-model dependence is now realised concentration risk under DORA, and what owning the orchestration layer, building model-agnostically, and operating in-jurisdiction actually buys an institution.
A BaFin-regulated Wertpapierinstitut gave every employee access to a frontier model. The controls that mattered were not mystical: vendor file, AI literacy, output gate, logs, and a living usage review.
Salesforce Headless 360, read from the CIO seat. The first Tier-1 enterprise-SaaS vendor to voluntarily move its UI out of the centre of its own product. Three orders of effect, the Frankenstein problem, and where the vendor relationship has actually moved.
Dario Amodei's FT Lunch, read from the institutional-allocator seat. The Mythos disclosure that put US banks on emergency calls, the "diffuse at the speed of trust" adoption-throttle thesis, and the $380 billion underwriting problem sitting under five internally-tension-laden CEO positions.
Stanford HAI's 12 headline takeaways from the 2026 AI Index, reorganised into five operating signals for CIOs, heads of research, and compliance officers running AI under regulation.
A research note on why raw AI fails in regulated finance, and why the expensive distance between fluent output and governed institutional output is where SVRN ALPHA operates.
Systematic review of 88 papers and 29 multi-agent trading systems. The central claim is that architecture, governance, and role design matter more than choosing the newest frontier model.
Humwork.ai (YC P26) put human-in-the-loop on an API. The viral framing is "AI hires humans." The structural story is that expertise just got a ticker price, and the MCP layer is a training pipeline that funds its own replacement.
Five foundation models from four providers converge on the same economic problem: AI adoption is being priced on subsidies that will not last forever.
First-order effects are what you measure. Second- and third-order effects are where the strategy actually happens. An essay on why institutional AI programmes stall at the productivity dashboard, and what sovereign AI mastery actually requires.
Seven frontier models independently converge on the same structural outcome: compression of middle layers across labor markets, organizations, and value chains.
S-curve analysis of whether enterprise adoption is moving fast enough to justify current AI infrastructure spending and implied revenue expectations.
Software multiple compression is treated as a symptom of a deeper shift: AI agents automate execution work that traditional SaaS was built to organize.
A joint study with MPCM on how prepared German HDAX companies are for AI-native investor relations.