Skip to main content
Pin the scope — In every question below, name the entity and the source-of-truth tables. A plausible answer from the wrong (summary) table is worse than no answer — if two sources could answer, run both and let your SME rule which is truth.
Run each of these against your validated warehouse. For every answer, note that Ana cites the governed surface, names the basis (incurred / earned / accident-year, by LOB), and shows the rendered SQL. The golden values below are pinned against the synthetic P&C warehouse (window 2024, horizon 2024-12-31) in validation/golden-queries.md — your numbers will differ, but the shape and the invariants should hold.

​
5.1 · Loss ratio (the headline metric)

Prompt
You’ll see: the governed surface return incurred losses (paid + case reserve) over earned premium, accident-year — the synthetic book pins 94,860,400/94,860,400 / 173,458,746 = 0.5469. Ana names the basis instead of silently picking one; the paid basis runs separately and comes in lower (it excludes case reserves).

​
5.2 · Combined ratio

Prompt
You’ll see: the combined ratio as loss + expense — synthetic 0.5469 + 0.2500 = 0.7969, comfortably under 100% (an underwriting profit). The invariant combined == loss + expense is asserted in the golden queries.

​
5.3 · Frequency and severity (the two levers)

Prompt
You’ll see: frequency 521.17 / 1,000 (13,000 claims / 24,944 PIF) and severity **7,296.95∗∗(7,296.95** (94,860,400 / 13,000) — paired, because the loss ratio alone hides which lever moved. Note frequency × severity ≈ pure premium (loss cost).
Know the exposure basis — The governed claim_frequency surface uses policies-in-force as the denominator. The actuarially correct base is earned exposure (car-years, house-years, payroll) — PIF is an approximation. If your warehouse carries an earned-exposure fact, re-point the denominator; Ana will say which it used rather than imply it’s earned exposure. See notes/frequency-severity.md.

​
5.4 · Reserve position

Prompt
You’ll see: the open book at case 32,340,903∗∗⋅paid∗∗32,340,903** · paid **12,936,361 · incurred $45,277,264 · reserve ratio 0.7143 — with Ana stating plainly that this is case-incurred, not ultimate.
Where’s IBNR? — The governed surface computes case-incurred (paid + case reserve) from the claim snapshot. IBNR is not in raw claims — it’s a reserving-study output — so case-incurred understates ultimate for recent accident periods. Ana will say so rather than fabricate IBNR; the decision record in notes/reserve-definition.md documents the scope. If you have an actuarial ultimate/IBNR table, add it as a backing.
Why everyone gets the same number — Loss ratio, combined ratio, and reserves can each be computed several ways. The ontology pins one governed definition — incurred / earned / accident-year, with the decision recorded in ontology/notes/ — so Actuarial, Underwriting, and Finance stop disagreeing about which number is “the” number.

​
5.5 · When the answer isn’t governed yet — watch the model grow

Now ask something from your shortlist that the starter doesn’t already cover — your earned-exposure fact, a reinsurance treaty structure, a loss-development triangle, the actuarial ultimate/IBNR table. This is the important beat: a starter pack is a head start, not the finished model.
Prompt
You’ll see: Ana explore only the frontier (not re-derive the whole warehouse), answer, and propose a write-back — a new metric committed to your repo with provenance. Review and merge it, and the next person who asks gets the governed answer for free. That’s the malleable loop: the ontology you ship is the one you grow, and it gets more complete every time you use it.
You ratify; high-stakes definitions get review — Ana proposes; humans ratify via normal git review. Anything in governance-pii.md scope (fair pricing, reserve MNPI, medical-claim sensitivity) or a core ratio surface should require review before merge (CODEOWNERS-style) — see STANDARDS.md. The point isn’t to let an agent rewrite your model unsupervised; it’s that discovered knowledge is captured instead of re-discovered next time.

​
✅ Checkpoint

  • Loss ratio, combined ratio, frequency/severity, and reserves all answered through governed surfaces, SQL and basis shown
  • You can point to the notes file explaining at least one metric’s definition decision
  • Ana proposed a write-back for a not-yet-governed question — and you saw it land as a PR