// Reinsurers and fronting carriers
You are modeling someone else's spreadsheet.
Cedent exposure arrives in three competing schema families with multiple live versions each. Program data arrives in as many formats as you have counterparties. Everything downstream — the model, the aggregate, the statutory filing — inherits whatever came in. We build at the point of arrival.
Dearborn Labs is an AI-native software development firm built specifically for insurance. You hire us; you don't buy software from us.
// The situation
Schema versioning is a business blocker, not a nuisance.
Moody’s documents the case in full: a reinsurer received Florida exposure in CEDE 8.0, needed it in EDM to model it, and its conversion tool supported only CEDE 9.0 and 10.0. That left three options — a months-long tool update with revalidation, declining the business, or surrendering analytical control to somebody else’s conversion.
Moody’s · exposure data management · publication date not stated in our source.
None of the three is a technology decision. Each one is an underwriting decision made by a file format, which is why translation belongs to you rather than to a vendor’s release calendar. The first thing worth building here is the translator, with a validation report attached.
// The stack you actually run
Multiple live versions at once is the normal state.
Coexisting versions are not a failed migration. They are what a live book looks like when cedents, brokers and vendors upgrade on their own schedules. Schema version translation with validation is a recurring, budgeted line item at every reinsurer and cat-exposed carrier, it is unglamorous, and it is never finished — which is the argument for owning the translator rather than buying a conversion.
Fronting adds the second half of the problem. Fronting is roughly $22.6B, about 20% of US MGA premium, and a fronting carrier holds statutory reporting obligations and regulatory accountability for business it does not underwrite and whose data it does not originate. Conning · Managing General Agents · July 2026 · 2025 volumes. That book arrives as inbound bordereaux in N formats, on N cadences, at N quality levels, and the filing deadline does not move to accommodate any of it.
Meanwhile the manual path has a cost you already pay. Preparing cat-model input by hand means de-duplication, OCR, geocode validation and gap-filling with inconsistent assumptions, submission by submission. Your own hours per submission are the baseline, and week one measures them.
// The build
Four builds at the point of arrival.
Exposure schema translation with validation
CEDE, EDM and OED, across the versions actually in circulation, with a validation report that states what was lost or defaulted in translation rather than silently coercing it. A conversion that does not report its losses is how a portfolio gets modeled on assumptions nobody chose.
Surface: your exposure database and the model vendor's import path, versioned.
Cedent SOV to model-ready exposure
Spreadsheets, PDFs, scans and email bodies where “BLDG_VAL,” “BldgValue” and “RCV Building” all mean the same thing. Everything maps to a canonical model, geocoding happens with resolution recorded as a field, and gaps are flagged rather than filled silently. Secondary modifiers — roof shape, deck attachment, covering material — are usually not SOV columns at all; they live in narrative appraisals and engineering surveys.
Surface: the inbound submission channel plus your canonical exposure schema.
Ceded recovery reconstruction
Treaty terms held in contract PDFs joined against claims in the database, surfacing recoveries that were never identified. Terms in, claims in, unrecovered out, with the governing clause cited on each one. This build has a dollar answer, and the dollar answer is usually positive.
Surface: the treaty document set and your claims and cession ledgers.
Fronting bordereaux to statutory
N program counterparties, N formats, one set of statutory obligations. Inbound files normalize at receipt, reconcile to booked, and the variance surfaces before the filing rather than during the audit. Highest-value target in the ecosystem: high volume, unavoidable, currently manual, hard regulatory deadline, measurable error rate.
Surface: the inbound reporting channel through to your statutory close.
What we measure
Hours per submission from cedent file to model-ready exposure, baselined against your own range rather than an industry average. Percentage of locations geocoded at rooftop against ZIP centroid, which is the single input that most changes a coastal loss estimate. Unrecovered ceded dollars identified, with the clause cited. Days from period close to filing-ready aggregate, and how many counterparties were late.
// We ran one of these
Every build on this page reports what it inferred, because the file being assembled is the file somebody else will model and a filing will rest on.
The operating record →// What's hard about this
Two constraints, and neither is a modeling problem.
A normalized SOV looks authoritative whether or not the values were ever right.
This is the specific failure mode of AI in exposure work: cleaning the schedule raises confidence without raising accuracy, and a confident schedule is harder to argue with than a messy one. The answer is to ship the confidence signal alongside the data. Every field is labeled as extracted, inferred or defaulted, geocode resolution is recorded rather than assumed, and the validation report travels with the file so the actuary reviewing it can see how much of the aggregate rests on a default.
Insurance-to-value drift is a human and appraisal problem no model resolves.
Nothing in the data tells you an insured’s reported value is stale, and the margin clause or occurrence limit of liability endorsement will cap recovery at a multiple of that stale number. What is buildable is the flag and the arithmetic: surface schedules whose values have not been refreshed against replacement cost, quantify the cap exposure at partial loss, and route the location to appraisal as a named exception. The valuation stays a human judgment; the exposure to it stops being invisible.
// What ships with it
Lineage back to the inbound file, or it cannot be certified.
Roughly half the states have adopted the NAIC AI Model Bulletin, and New York’s Department of Financial Services states the position plainly: an insurer cannot rely solely on a third party’s claim of non-discrimination. Responsibility is non-delegable, which matters more here than anywhere — the business was originated by somebody else and the filing is yours.
NAIC and state bulletins · current at September 2026.
So every build leaves with a model inventory entry and risk tier, data lineage naming the provenance of every external source and every delegated reporter’s file, pre-deployment testing results, drift thresholds with remediation triggers, and a human-override specification naming who can act on a variance.
What the governance file contains →Start with the file that arrives, not the model that reads it.
Four weeks, fixed scope, one named senior engineer working through your actual inbound cedent and program files. You get the translation and validation spec, and a list of what is not safe to build yet.