Rates by line
Burning cost is frequency × severity, where severity is indemnity-only, trended to the 2026–27 accident date, limited to a $1m basic limit on a lognormal, then loaded for defence cost, parameter uncertainty (by evidence class) and correlation. The gross rate divides that by the target loss ratio — what is left after an itemised expense stack, the net cost of reinsurance, and margin. Every row carries its evidence class.
What a buyer pays
Worked accounts at the current assumptions. The fintech pair is the whole sales conversation: same company, same revenue, differing only in whether a spend ceiling exists.
The market today
Bottom-up from the world's operational robot stock, at our rates. Caged arms sit on existing machinery programmes and are excluded from the addressable figure.
Reality check. Cyber insurance — the closest structural analogue — reached about $15.3bn of global premium in 2024, after roughly twenty-five years, and is still under 1% of global P&C. Any model that produces a multi-billion autonomy line inside five years is wrong.
Thirty years, four worlds
Annual simulation 2026–2056 under four worlds. BASE follows Goldman's humanoid curve; BULL follows Morgan Stanley's; BEAR assumes humanoids stall and AI liability is absorbed by incumbents; UMBRELLA is the bull curve with full execution. Roboline's economics are a real P&L — MGA commission net of distribution, then a loss-ratio path, quota-share cession, duration-weighted float, opex, and the external capital a carrier needs. This cannot tell you which world arrives.
Premium pool by world
annual, USDRoboline net revenue and EBITDA —
annual, USDWhat one commercial humanoid costs to insure
per unit-year, USDEvidence ledger
Every parameter, with what it rests on. Measured from data we hold; Borrowed from an adjacent line; Judgment is engineering opinion we have not yet earned the right to hold. Count the labels.
How to read this
The absolute level of any rate here could be wrong by a factor of two in either direction; there is no loss history for these classes and the page says so on every row. The ratios between segments are where the confidence is, and the ratios are what pricing decisions actually turn on. The simulation is the argument written down so it can be attacked one assumption at a time — which is what the controls are for.