The compared read

The comparison

Requirements go in, coverage documents go in, a cited answer comes out. That has been the engine’s shape since version one: the standing checks are one requirement set, a counterparty’s insurance requirements are another. V3 adds the presentation: each requirement beside what the coverage provides, with the shortfall in dollars where it is provable. The demo’s requirement sets come from real published documents, page and printed words behind every row. A requirement the vocabulary cannot express is carried with the reason it is not checked. Each row can open the pages themselves: the requirement’s page beside the policy’s, rendered from the documents with the printed words highlighted, so the verdict and the evidence for it are read together. A counterparty’s requirements can also be uploaded: the document is compiled once, every line verified against its text in code, what the vocabulary cannot check carried with the reason, and lines that cannot be verified left out and said so.

Questions

The standing checks ask what somebody wrote down in advance. The question box takes anything: a person questions the documents in plain words and the answer comes from the extracted text, quoting the page, with what the pages do not settle said plainly. Provisions are found and quoted; what a provision does to a specific loss is a legal reading this build does not perform, and answers say so when asked for one. Every answer is checked against the page text in code before it is shown, and prints its cost and the content hash of its rules.

The two provable parts

• Retained exposure. Dollars the program keeps with the insured: a self-insured retention the insured pays before coverage attaches, or a deductible the carrier pays and collects back. Read where the file prints it, cited to the page.
• Gap exposure. Structural holes in the program. The space between one layer’s exhaustion and the next layer’s attachment, and an underlying policy that a schedule names, with a stated limit, that is not in the file.

Both are arithmetic over values the assemble step already produces, and both are citable. No model performs the arithmetic, which is what makes the figure reproducible rather than merely plausible.

Where the dollars stop

The exposure figure includes the first category and never the second.

• Provable from the documents: uncovered structure. Retained dollars below attachment, structural space between layers, a scheduled policy absent from the file, priced at the schedule’s own stated limit and marked as resting on the schedule rather than the policy. These are facts about the file.
• Not provable from the documents: predicted loss. What a claim would actually cost requires exposure data this system does not have and is not being given.

So the claim is that here is your uncovered exposure, proven from your own documents. It is never here is your liability in the sense an actuary would use the word. Claims entered by a user are that user’s facts, carried and displayed as theirs. The system does not forecast.

Where this goes next

Stated as direction, not as plan.

• Work-type attribution. Answering whether a specific scope is uncovered requires reading what endorsements and exclusions do to coverage, which this build does not do and does not pretend to.
• Loss runs ingested rather than typed. The program schema already carries the fields they would fill.
• Read-only connections to document stores. Policy packages are large and moving them is the real friction. If the documents already live somewhere the system can reach, fetch them at the source.