// What we build
The hard part was never the layout.
Every carrier writes a different format, and parsing them is the easy half. What actually breaks comparability is valuation-date alignment, ALAE inclusion stated differently or not at all, open-versus-closed reserve treatment, and claim counts against occurrence counts.
// The build
Four things, stated rather than assumed.
Valuation-date alignment
Two loss runs valued three months apart are not comparable, and nothing on either document says so. We align them and show the shift rather than absorbing it.
ALAE treatment, stated
Included, excluded, or unstated — and unstated is the common case. We resolve it explicitly and label which of the three we found.
Open-reserve development
A five-year loss run priced at face value systematically understates ultimate cost. We surface the development assumption instead of burying it in a total.
Claim versus occurrence counting
Claimant-level and claim-level counts differ, and the difference changes frequency. We normalize, and we label which convention each source used.
A loss summary that hides its normalization assumptions is worse than no summary.
Every output states which fields were extracted, which were inferred, and which were defaulted. That is not an optional feature of the build — a cleaned file looks authoritative whether or not the underlying values were ever right, and a system that raises confidence without raising accuracy has made the book harder to reason about, not easier.
“The assumptions are shown because of what happens downstream when they are not. A valuation-date or ALAE choice becomes a number in a rate or a reserve, and somebody has to defend that number.”
// Where this comes from// What's hard about this
Normalization raises confidence faster than accuracy.
That is the specific failure mode here, and it is why the confidence signal is built alongside the data rather than added afterward. We will tell you what the file cannot support, and we would rather deliver a summary with four caveats attached than one that reads clean and is quietly wrong in the third year.
Bring us one workflow.
We will tell you which assumptions your current summaries hide, and what changes when they are shown. How the read works →