Human BEFORE the Loop · actuarial epistemics
Can anyone show that this AI decision was warranted when it was made?
Epistemic risk is the exposure an organisation takes on when it acts on claims whose standing it cannot show at the moment of acting. This site lets a domain propose a crossing, a proposed AI-mediated state change, and get an appraisal: typed checks, risk archetype readings, the evidence gaps that block warrant, and a receipt.
The graph reports the crossing; it does not allocate risk. The underwriter allocates.
No scores, prices or ratings. Not insurance, underwriting or actuarial advice. See non-claims.
Why now
In 2026 insurers began excluding AI-related losses from standard policies. ISO introduced generative-AI exclusions for commercial general liability in January, and Verisk is considering exclusions for agentic AI. Insurers cite three reasons: AI systems are opaque, loss scenarios are hard to model, and harm may be systemic.
Each reason is an epistemic problem before it is an actuarial one. Opacity is missing evidence of warrant. Unmodelled scenarios are missing risk archetypes. Systemic harm is unmeasured accumulation. That is the gap this site works on.
Sources: Fenwick, The Insurer, Claims Journal.
Three tests, three actuarial dimensions
Hazard
Local Admissibility: Can it act?
Is this crossing warranted: typed, bound to its evidence, authorised by someone other than itself, within policy, with its invariants measured?
Accumulation
Recursive Viability: Can it keep acting?
What fails together when many crossings share a model, vendor, policy or data source, or feed back into themselves?
Severity
Transformability: Can we still change course?
Can the crossing be undone, and how much of the reachable future does it close off?
What you can do here
Propose a crossing
Submit a structured proposal and get a determination (EXECUTE or HOLD with typed reasons), an archetype appraisal, a receipt and an underwritability conditions document.
Open the workbench →Browse crossing classes
3 reference classes, each with its policy, invariants, evidence requirements and what an underwriter would need.
See classes →Risk archetypes
10 machine-readable archetypes, each labelled with whether its physics is a formal measure, an analogy, or absent.
See archetypes →Portfolio and accumulation
Run many crossings together to see shared dependencies, HOLD rates and epistemic IBNR: crossings that took effect without a warrant that still stands.
Open portfolio →Where the evidence stands
36 claims about the machinery are tested by named tests. Every claim about risk, insurance and loss is proposed: the archetypes, all crossing classes and the reconstruction study have not been run against a real domain. The evidence ledger lists each claim and what backs it.
Every appraisal also reports two kinds of uncertainty separately: about values inside the risk model, and about whether the model has a path for the consequence that matters. No class has yet been tested for the second; see Late evidence.
No insurer or regulator recognises these receipts. Until one does, they carry no price. Recognition is an institutional project measured in years, not a feature of this site.