DEV Community

尚毅段
尚毅段

Posted on Originally published at coveredcheck.com

An HO-4 policy is a rule engine: four traps when modeling insurance clauses as data

A friend's apartment flooded during a storm. Wind had opened part of the roof, rain came through, and the ceiling came down over a bedroom. Her renters policy covered wind damage. It also excluded flood. The claim paid nothing.

The sentence that decided it was not in the exclusions list. It was a preamble sitting above the list, and it read, in the Safeco HO-4 I later pulled apart: "We do not cover loss caused directly or indirectly by any of the following excluded perils. Such loss is excluded regardless of any other cause or event contributing concurrently or in any sequence to the loss."

That is a priority rule, not a paragraph. It says: if an excluded cause touched the loss at any point, in any order, the whole loss is excluded, even when a covered cause also contributed. The wind did not need to be the main cause. It only needed to be part of the chain.

I build CoveredCheck, a site that reads filed renters policy forms and quotes them. The work is less about insurance than about turning regulated prose into something a program can evaluate, and most of the difficulty is in places where a naive data model gives a confident wrong answer.

Here are the four that cost me the most.

Limits attach to the cause, not the item

Ask whether renters insurance covers jewelry and you will get a yes. Ask what happens when a ring is stolen and the answer changes.

The form caps theft of jewelry at $1,500 on any one article and $3,000 in the aggregate. Read the phrasing carefully: the cap is on loss by theft. It is not a cap on jewelry. A ring destroyed in a fire is measured against the general personal property limit. A ring taken from a drawer is measured against $1,500.

So coverage is not a property of the object. It is a property of the pair, cause and category. Any model that stores "jewelry limit = 1500" is wrong in a way that only shows up on the claim that matters, and it will be wrong in the homeowner's favour right up until the adjuster's letter arrives.

The per-article cap and the total cap do not multiply

The same clause gives two numbers, and they interact. Four rings, each under $1,500, do not pay $6,000. They pay $3,000, because the aggregate binds.

Writing that as code is a two-line function that people get wrong anyway, usually by applying the total only when a single article exceeds it:

def theft_payout(articles, per_article=1500, aggregate=3000):
    capped = [min(a, per_article) for a in articles]
    return min(sum(capped), aggregate)
Enter fullscreen mode Exit fullscreen mode

The interesting part is not the arithmetic. It is that the two limits live in the same sentence and serve different purposes. The per-article cap stops one expensive item from eating the policy. The aggregate cap stops a collection from doing the same. A schema with one "limit" column per category cannot hold both, and the loss is silent: the site still renders, the number still looks plausible.

Exclusion order matters more than exclusion text

The flood denial above comes from combining three separate rules in the right order.

First, the perils the policy insures against. Second, the exclusions, and the anti-concurrent causation preamble that makes them win ties. Third, the carve-backs, which pull specific situations back in. A burst pipe is water damage, water damage is excluded, but the ensuing-loss language and the specific pipe provision put it back.

Each rule is a sentence someone wrote for a reason. Evaluated out of order they produce nonsense: apply the carve-back before the exclusion and everything is covered; apply the exclusion last and nothing is.

This is the part I underestimated. I assumed the hard work was extraction, getting clause text out of PDFs and into a database. Extraction was tedious but mechanical. The hard work was deciding what questions the data has to answer, because a clause that looks like prose is usually a conditional, and the condition is often a word rather than a table.

"The form we read" is load-bearing

Every judgment on the site points at a form number, and the site says which one. That is not legal caution for its own sake.

Renters policies differ by carrier and by state filing. The $1,500 and $3,000 figures come from the Safeco HO-4 form I read. Another carrier may file different numbers, and the same carrier may file differently in another state. Mold shows up as a $10,000 sub-limit on many forms; temporary housing runs up to 12 months on some and has a dollar cap on others.

Once you accept that, the design changes. A coverage claim without a form, a section and a quoted sentence is not a fact, it is a guess with a confident tone. So the unit of content is not the article, it is the clause. The article is a view over clauses.

That has a cost in how the site reads. Pages open with the quote, verbatim, then explain it. Some readers want the short answer first, and they are right, which is why the short answer sits above the quote. But the quote has to be on the page, and every number I publish has to be traceable to one.

What I would tell another builder

If you are modelling regulated text, three things held up.

Ship the source next to the conclusion. Not a link in the footer, not "according to industry standards" in a paragraph. The sentence itself, with its section name, on the same screen as the claim it supports.

Store the question, not just the answer. "Does renters insurance cover jewelry" has at least three useful answers depending on the peril. If the schema has one answer field, the schema is the bug.

Let the quote be ugly. Policy language is repetitive, capitalized in odd places, and full of defined terms that mean something narrower than the word suggests. Cleaning it up into readable prose is where the meaning leaks out.

The friend's claim was denied correctly, under the terms she signed. She had read the exclusions page and remembered flood. Nobody reads the preamble. That is the gap this kind of site exists to close, and it is not closed by writing friendlier summaries of a document whose power is in its precision.

If you want to see how the clauses are laid out, the jewelry and anti-concurrent-causation pages are the two I would start with: https://coveredcheck.com/renters/jewelry/ and https://coveredcheck.com/clauses/anti-concurrent-causation/

I am not an insurer, broker or lawyer, and none of this is legal advice. Check your own policy wording, and start from the clause your denial letter cites.

Top comments (0)