Methodology

A geodesic exposure unit for path-dependent AI liability.

The holonom is not an AI risk score. It is a candidate actuarial exposure unit for institutional AI decision liability where historical loss data is insufficient and model-level metrics are incomplete.

Primitive

H (d, γ) = \frac{ℓ ( γ ) - ℓ ( γ ^{*} )}{κ}

d is an AI-enabled decision. γ is the actual institutional path the decision travels; γ* is the minimum-distortion governance path. ℓ measures realized path distortion. κ is a normalization constant equal to one underwriting-relevant unit.

One holonom equals one normalized unit of avoidable institutional distortion in an AI-enabled decision pathway.

Rating factor

P = P_{0} + λ i = 1 \sum n H_{i}

A candidate additive rating factor. P₀ is the baseline premium and λ is the price per holonom. This is a bridge for actuarial testing, not a full pricing model.

Reserving lens

R = R_{0} + α H_{claims}

The worse the evidentiary distortion on claims-side decisions, the higher the expected claim adjustment expense, defense cost, dispute duration, and uncertainty margin.

Default sub-score weightings

Variable	Weight
volume	0.10
severity	0.20
evidence	0.15
defensibility	0.15
audit	0.10
override	0.10
coverage	0.10
friction	0.10

Per-row H is computed as volume_factor · Σ(wᵢ · sᵢ) / κ with evidence, defensibility, audit, and override inverted so that "poor" always increases H.

Sub-score glossary

Decision volume: Log-scaled count of AI-assisted decisions. Weighted low: severity and evidence dominate.
Severity class: Low / Moderate / High / Critical. Credit denial, fraud flag, claims denial, pricing change.
Evidence preservation: Logs, provenance, reason codes, policy basis. Poor evidence increases H.
Legal defensibility: Adverse action notices, consent, explainability, appeal rights.
Audit reconstruction: Can the decision be reproduced after the fact from stored artifacts?
Human override clarity: Documentation of who changed what, when, and why.
Coverage ambiguity: Is the event clearly covered, excluded, or actively disputed?
Claims friction: Expected cost of proving or disputing liability at claim time.

What this is not

Not an AI governance score.
Not a compliance rating.
Not a replacement for full portfolio pricing models.
Not a substitute for legal review of specific claims.

v0.1 status. Holonom v0.1 is a proposed exposure unit and an indicative exposure model. It is not an accepted actuarial standard, not a filed rating plan, and not an actuarial standard of practice. The framework is being hardened through adversarial advisor review.