Skip to content
Decision-Evidence Operating System

Outcome · Loss ratio

The headline ratio held. The mix underneath it doubled its cost share, and nobody was asked to decide that.

A loss ratio is an average over decisions no one is reading individually. It can sit flat for years while the composition of cost migrates underneath it — and by the time the average moves, the decisions that moved it are three development years old.

The mix moved. The Irish data shows by how much.

94%of settled private motor claim numbers in H1 2025 were damage claims — Central Bank of Ireland, NCID
56%of settled claim cost those damage claims now carry, versus 28% across 2015–2019 — Central Bank of Ireland, NCID
36%movement in average fire-and-theft claim cost inside one half-year, in the same NCID series

How we count this

Central Bank of Ireland, National Claims Information Database, private motor. In 2015–2019 damage claims were 90% of settled claim numbers and 28% of settled cost; by H1 2025 they were 94% of numbers and 56% of cost. The count share barely moved. The cost share doubled. An average that stayed calm through that is not evidence that the book was calm.The 36% half-year movement in fire-and-theft average claim cost is from the same NCID series, and it does real work in our gate register: it is the recorded reason acceptance gate G-49 carries a tolerance of ±6% — a tolerance widened with a written reason, never silently.

What is specified against that reality

The demo-data corpus specification materialises 84 loss development triangles from 1.9 million specified claim transactions — designed, not generated, and gate-checked against the published NCID shape before it ships. Development is where a loss ratio tells the truth, which is why the triangles are specified first and the ratio is not specified at all.

On pricing conduct: the Central Bank of Ireland’s ban on price walking has been in force since 1 July 2022. The specified renewal_price_test dataset stores each charged premium against the year-one-equivalent counterfactual computed by the same model — so “would a new customer have paid this?” is a query, not an investigation.

The triangles and the renewal_price_test dataset are specification, carried by acceptance gates, never a measurement of anything. No figure in this section is a loss-ratio claim.

The pilot metric stays empty until claims develop

Projectedgross loss ratio movement attributable to appetite-aware referral, against a paired baseline lane

How we count this

We measure the same classes of risk in both lanes over the same period, and we wait for development rather than reporting on incurred figures early. A ratio that improves on immature claims is not a result.The falsifier: a loss ratio that improves only because the lane declined more business is shrinkage, not selection. Written volume is reported beside the ratio, and if the ratio only moves with volume falling, we record the result as negative. That rule is ours, set before any data exists so it cannot be adjusted after seeing some.

The mechanism: appetite arrives as a constraint, not as a report

Live portfolio accumulation state is injected as a hard constraint on the individual risk, inside the decision moment rather than in a report published after it.

Per-location hazard is scored while the quote is being made, and that run is logged into the decision it justified rather than kept in a separate modelling artefact.

Where the constraint and the risk disagree, the case is argued by seats working from different evidence and ruled on by a named chair — so a declined risk has a reason somebody signed, and an accepted one has a reason too.

SOV/services/nexus-disaster-engine/src/{lib.rs,sims/,feeds/}

PC/src/routes/rooms.routes.ts

The Chaired Deliberation Chamber

Versioned expert personas argue the case from different angles — including one seat whose only job is to attack the conclusion — and a named human chair rules, at a durable waitpoint.

PC/src/routes/rooms.routes.ts

What runs today

What runs today is peril physics and the analytics shell around it, behind the accumulation constraint above.

Where you would start

The part of this you can run today is the deliberation, not the catastrophe financials: a chaired room on the risks your own appetite rules already flag, with the reasoning recorded. That is a smaller claim than the metric above, and it is the one we can stand behind now.

See the room
Book the 20 minutes

A person replies with two or three times. Not a sequence.