INDUSTRY FIELD GUIDE
Insurance ratemaking input provenance for AI review
A rate analysis can change because the underlying exposure data changed, because the denominator changed, or because someone replaced a table.
NAICS 524298 · All Other Insurance Related Activities
A rate analysis can change because the underlying exposure data changed, because the denominator changed, or because someone replaced a table. Census includes insurance ratemaking services in this industry. Actuarial data-quality materials provide context for assessing inputs. An organization can test whether a model detects a version or unit mismatch before calculating anything. Census NAICS manual · Actuarial data quality standard
Build two invented exposure extracts. Give the second a revised reporting basis and one added cohort. Add a commentary note that references the earlier file. Ask the model which fields can be compared, what changed, and what requires the owner’s confirmation. Include a change log the reviewer can use as ground truth. Keep premium decisions outside the exercise.
Do not share filed or internal models, customer details, actual exposure or loss figures, or proprietary factors without authority. Start with fictional fields, invented quantities, and an independent reviewer.
Describe an input-provenance test
Original Fiund exercise
An exposure-unit change and an added cohort prevent a raw total comparison
Invented exposure extracts use artificial cohort names and arbitrary counts. No premium model, loss history, customer record or filed rate is represented.
| Record | Source | What it records |
|---|---|---|
| RI1 | Extract E1 | Period P, cohort A only: 120 policy-months. Commentary M cites E1 and describes its month basis. |
| RI2 | Extract E2 change log | Same period P: A=10 policy-years, added cohort B=5 policy-years. Explicit rule for this fixture: 12 policy-months per policy-year. |
| RI3 | Review note | B absent from E1. No evidence that E1's omission means zero B exposure, and no loss figures or rate calculation supplied. |
Question for the model
Which A comparison is supported after unit reconciliation, and can E2's 15 total be called a drop from E1's 120?
Show the reference answer and checks
RI2's stated conversion maps A's ten policy-years to 120 policy-months, matching RI1's A exposure. The raw figures 120 and 15 mix units and populations: E2 also adds cohort B. RI3 supplies no prior B amount, so the complete-population change remains unresolved. Commentary M describes E1 and should not be silently relabeled as analysis of E2. No premium or risk conclusion follows from this input check.
- Comparing exposure totals across different units without conversion.
- Treating an absent earlier cohort as a measured zero or reusing old commentary for a revised extract.
Fields to preserve
- Extract version and cohort scope
- Defines the population.
- Exposure value, unit and conversion rule
- Supports a comparable measure.
- Commentary parent and missing component
- Preserves the source of interpretation.
Test another case
Add E1's B exposure as 60 policy-months. Require a same-scope total comparison of 180 policy-months in each version, without inventing a rate effect.
Use original fictitious extracts. Exclude actual exposure and loss data, proprietary factors and rating decisions.
Discuss a contribution
Describe the sources you control, the task they could support and what has already been checked. Use the existing-data assessment for an archive or the collection brief for a new recording project. The contributor agreement defines Fiund’s licensing authority and compensation. A guide or fictional exercise does not establish an accepted collection or a sale.