Skip to main content
The single biggest mistake with a starter pack is to skip this step — to point it at your warehouse, decide it’s “close enough,” and ship. That bakes in decisions before you know which questions matter, and you get a model that’s 10% useful and 90% noise. This module is the antidote, and it takes 15 minutes.

​
0.1 · Answer one question

The North Star question — When this works, what does someone do differently on Monday morning?
Not “model all of insurance.” Something narrow and real: “the actuary stops waiting a week for an accident-year loss-ratio cut by line,” or “the underwriting lead sees frequency and severity moving apart in time to re-rate.” If you can’t name it yet, you’re not ready to build the ontology — you’re ready to have this conversation. Have it first.

​
0.2 · What a North Star looks like (examples to spark yours)

Most P&C / carrier teams land on a North Star adjacent to one of the three below. These are examples, not a menu to pick from — yours might be one of these, a blend, or something next to them. In Module 1, Ana drafts yours from your actual data; reading these now just tells you which surfaces and classification become relevant once it’s set. (Full detail in the repo’s NORTH_STAR.md.)
ArchetypeForLead with
A · Loss-ratio & reserving parity”Match the actuarial numbers we already trust”loss_ratio, combined_ratio, reserve_position + golden queries (loss ratio 0.5469, combined 0.7969) — reconcile first, govern second.
B · Underwriting & pricing”See frequency and severity before the loss ratio does”claim_frequency, claim_severity + the frequency × severity ≈ pure-premium note + LOB/peril rollups.
C · Claims operations”Manage the open inventory and the reserve position”reserve_position, claim_severity + the grain discipline + the layered reserve vocabulary.

​
0.3 · You won’t fill out a checklist — Ana drafts it

There’s no question bank to work through and no form to fill out. In Module 1 you’ll paste one prompt and Ana looks at your actual data, runs a short scoping conversation, recommends the archetype that fits, and writes your North Star down as north_star.md — a paragraph on what the ontology is for plus the 6–8 questions it must answer in 30 days. That single artifact is what every later module builds toward. The watch-outs that make you sound like the expert in the room — written vs. earned, paid vs. incurred, accident vs. calendar year — live in the repo’s decision notes, and Ana surfaces them as they come up.
Why the model gets cheaper as you use it — This isn’t just good practice — it’s the design. An agent that re-discovers your warehouse on every question pays an “amnesia tax” (most of its tokens go to rediscovery, and results aren’t reproducible). Reading a known model first and committing what it learns inverts that: discovery is paid once and amortized across every future question. See TextQL’s research, “Malleability Is All You Need” (VLDB 2026) — a measured ~30% token reduction and a model that gets more reliable the more it’s used.

​
✅ Checkpoint

  • You can state your North Star in one sentence
  • You have a candidate archetype in mind (loss-ratio parity / underwriting / claims ops) — or you’ll let Ana recommend one in Module 1
  • You’re ready to let Ana draft north_star.md from your data in Module 1
Not sure which archetype? That’s fine — in Module 1 Ana looks at your data and recommends one. You don’t have to decide here.