Sample output

Inspect the work, with the assumptions attached.

Two kinds of evidence appear on this page, clearly separated: genuine pilot-generated report pages from one synthetic workers'-compensation case (run 7347e4dd), and an illustrative walkthrough of the validation mechanics. Everything is synthetic; every report is draft.

Synthetic data Draft — pending review

The synthetic pilot case

One synthetic captive writing workers' compensation. Six annual TPA loss-run snapshots, evaluated 2021-12-31 through 2026-12-31, are reconciled claim-by-claim — 21 claims growing to 119 — with each snapshot tied to its printed totals before any analysis runs. The pilot application then produces a reserve analysis with a permanent chain-ladder reference, a modeled uncertainty range, a funding-by-confidence-level exhibit, and an audit trail that stamps every assumption, data snapshot, and model diagnostic onto the run identifier.

The workflow at a glance

Illustration of the Prior Experience workflow in three stages: raw loss-run data with inconsistent fields, then validation and analysis showing field mappings, an exception count, data-quality pass rate and a development check, then a review-ready reserve work product with an ultimate-loss table and uncertainty chart.
Representative workflow using synthetic data — a marketing illustration of the ingest → validate → deliver sequence, not an application screenshot.

Reserve analysis by accident year Synthetic pilot output

Page 2 of the draft synthetic reserve analysis: a table of accident-year ultimate loss estimates comparing paid losses to date, the credibility-model indicated ultimate, the chain-ladder ultimate, and the selected ultimate for accident years 2021 through 2026, totaling 2,083 thousand dollars selected against 1,838 thousand paid; below it, a chart of paid losses with indicated-ultimate uncertainty intervals and chain-ladder markers. The page carries a DRAFT — pending review watermark and the run identifier 7347e4dd.
Page 2 of the draft reserve analysis (run 7347e4dd, evaluation 2026-12-31). Generated by the pilot application from the six snapshots above.

Compares modeled ultimate losses with a permanent chain-ladder reference and shows the uncertainty range behind the selected estimate. The chain-ladder column reconciles to the age-to-age factors and cumulative development factors in the report's own appendix; selected ultimates default to the model indication, and any override would be documented in the audit trail with reason and approver. Where the indication sits below paid-to-date for mature years, the report presents it as a finding to review — visible on the chart — rather than smoothing it away.

Validation and lineage Synthetic pilot output

Before any estimate is produced, each loss-run snapshot is reconciled and tied to its printed totals, and every assumption is stamped onto the run. These excerpts come directly from the report's audit trail:

Data lineage — snapshots used

SnapshotEvaluationClaimsTie-out
792ed0a02021-12-3121tied
6b3d46a12022-12-3142tied
875374b52023-12-3162tied
a623c10d2024-12-3182tied
d29bb3d52025-12-31103tied
67737bd02026-12-31119tied

Stamped assumptions (excerpt)

KeyStatement
tail_factor1.000 — none assumed beyond selected maturity
recoveriesPaid/incurred totals are net of recoveries
latest_diagonalObserved data only; never extrapolated
credibilityEntity experience weight Z = 0.86

The full audit trail also records the benchmark prior, model diagnostics, and every failed and successful fit attempt.

Reserve funding by confidence level Synthetic pilot output

Page 1 of the draft reserve funding exhibit: a table of required funding at stated confidence levels — expected mean 245 thousand dollars, 50th percentile 224, 70th percentile 363, 75th percentile 409, 80th percentile 456, 90th percentile 596, 95th percentile 702, with a chain-ladder reference of 833 — above a histogram of the modeled distribution of total unpaid losses with each percentile line annotated. The page carries a DRAFT — pending review watermark.
Page 1 of the draft funding exhibit (run 7347e4dd). Confidence levels are modeled reserve-funding views under the selected ultimates — not premium recommendations, and no single row is presented as the answer.

How validation works — illustrative walkthrough

The pilot reports above show finished work products. The examples below are a separate, simplified illustration — built by hand with synthetic general-liability rows — showing the kinds of mapping and exception decisions the validation stage records. They are not application screenshots.

as received — tpa_lossrun_Q4.xlsx (excerpt)Illustrative
CLM #DOLPD_AMTCASE RSVINCURSTATUS
GL-441703/14/23$12,450.002,00014,450O
44172023-03-1412,4502,00014,450Open
GL-452214-Mar-23(1,200.00)N/A-1200RO
GL-461005/02/20248,900 8,900CLSD?
GL-4610A05/02/20241,15001,150C
normalized claim history (excerpt)Illustrative
claim_iddate_of_losspaid_losscase_reservereported_incurredstatussrc
GL-44172023-03-1412,4502,00014,450openQ4:r17
GL-45222023-03-14-1,2000-1,200reopenedQ4:r31
GL-46102024-05-0210,050010,050closedQ4:r44+r45

✓ duplicate row removed · identifier change merged after confirmation · original source values retained per row

Exception log — illustrative excerpt
IDTypeDetailResolution
EX-014Duplicate rowGL-4417 appears twice with different header casingRemoved duplicate; source rows retained
EX-015Missing valueCase reserve blank on closed claim GL-4610Set to 0 for closed status; noted
EX-016Identifier changeGL-4610A appears after Q3; same DOL and claimant keyMerged after confirmation; both ids retained
EX-017ReconciliationReported incurred column ≠ paid + case on 4 rowsOpen — pending TPA confirmation
Reconciliation to source — illustrative excerpt ($)
MeasureSource filesAdjustmentsNormalizedDifference
Claim rows142−3 duplicates1390
Paid loss4,096,850−26,250 duplicates4,070,6000
Case reserves579,000−2,000 duplicates577,0000
Reported incurred4,675,850−28,250 duplicates4,647,6000

Reported incurred = paid loss + case reserves throughout (4,070,600 + 577,000 = 4,647,600). Adjustments are itemized in the exception log.

Methodology notes and limitations

The pilot reports estimate ultimate losses with a credibility model that blends the entity's own development experience with a stated external benchmark, and print a deterministic chain-ladder estimate beside every indicated figure so that divergence between approaches is a visible finding rather than a hidden one. Assumptions — tail factor, benchmark prior and its weight, credibility, and data conventions — are stamped on the run, and model diagnostics are reported including failed fit attempts. A real engagement would involve the reviewing actuary's own method choices and selections.

These samples are not actuarial opinions. They are draft, opinion-support work products built entirely from synthetic data, shown to demonstrate how the work is documented.

Full report downloads

Complete PDF reports will be posted here after each one passes internal QA review. The pages shown above are from reports whose displayed figures have been verified; remaining reports in the same run are still under review and are deliberately not published yet.

Want this run on a loss run shaped like yours?

Synthetic or redacted data is welcome.