- contact@verticalserve.com
- Every policy checked against what was agreed · part of the Insight broker suite
The architecture behind InsightProofing: how a document becomes canonical fields, how proofing and comparison work, what the agents add, and the state machines a check and a correction move through.
The UI is the account team’s workbench. The API holds the checks on PostgreSQL and runs the engines, the agents, the rules and the copilot on the shared Insight platform packages.
InsightProofing UI · React
InsightProofing API
Documents into the canonical coverage model, each value with its page.
Proof and compare: every pass, fail and change, cited.
Reason over numbered facts and draft.
A number not in the facts: the answer is dropped.
Prepared, answered and signed on the record.
Proofing controls (Rules Studio) fire on every Check and can block delivery. The copilot reads the check’s results and never acts.
Runs in your environment
Every document is read page by page. A field is a label and a value; a list is forms, exclusions, subjectivities, additional insureds or the schedule of underlying; a tower is its layers. Each value keeps its page and line, so every citation opens the page.
A carrier label, e.g. ‘Per Occ Limit’, tried in this order
A label a person confirmed is used first.
The taxonomy’s synonyms for the field.
Case, punctuation and filler words removed.
Used, but flagged for a person to confirm.
One field however the carrier words it.
No match at all: the label is shown as unmapped on the document, and once a person maps it every later document reads it.
Proofing compares canonical fields: the policy as amended by its endorsements against the binder, the binder against the quote. Derived checks add what no single field shows. Comparison reads the renewal against the expiring, field by field and list by list.
Named insured (legal suffix and punctuation folded, then flagged for review), dates, limits, sublimits, deductibles and retentions, premium, taxes, fees, commission.
Forms by form number with the edition compared, exclusions, additional insureds, subjectivities outstanding on the binder.
The policy term, renewal continuity with the expiring, premium + taxes + fees = total, a tower that attaches without gaps, a schedule of underlying that matches.
Each canonical field knows which way is broader — a higher limit, a lower deductible — so a change is read, not just shown. Changes over 10% warn and over 25% alert.
The agents run in order on the shared pipeline, each as itself at its autonomy level. Each reads the engines’ numbered facts, reasons with the model in a fixed shape that cites them, and prepares a decision.
sequenceDiagram participant AM as Account manager participant API as InsightProofing participant EN as Engines participant AG as Agent participant M as Model participant DR as Decision record AM->>API: Check API->>EN: read, proof, compare EN-->>API: results, each cited API->>AG: numbered facts AG->>M: reason over the facts M-->>AG: findings citing fact ids AG->>AG: grounding check: every number in the facts? AG-->>API: prepared decision (L1/L2) AM->>API: Accept, Reject or Ask API->>DR: signed answer
Sorts what arrived — expiring, quote, binder, issued policy, endorsements — and drafts the request for what is missing.
Asks; never sendsReads every document into the canonical coverage model for the line, each value with its page and the carrier’s own label.
Confirming a label is yoursPolicy against binder, binder against quote, field by field: pass, fail or needs review, cited on both documents.
Never decides a passThe renewal against the expiring as quoted, bound and issued: broader, narrower, costs more, costs less.
Endorsements dropped or added, edition changes, and restrictive forms nobody agreed to.
Layers that attach without gaps, and a schedule of underlying that matches what was bound.
Drafts the correction request to the carrier from the failed checks and follows each one up.
Capped at ask firstWrites the client summary — the coverage, what changed, what was corrected — once every failure is settled.
The AE signs and sendsRuns the proofing controls and prepares a practice leader’s sign-off when one fires.
Prepares; never decidesA check moves one stage at a time, and the stage machine says what stops it. A correction request moves from requested to corrected or waived.
A policy check
Moves on when the documents are in.
Moves on when every document is read.
Failures raised → with carrier; nothing to correct → ready.
Moves on when corrected or waived.
Only when no high-severity failure is unresolved.
The account executive signs.
A correction request
Raised from a failed check.
Sent to the carrier, with a due date and an owner.
The carrier has it; overdue past its due date.
A corrective endorsement re-proofed — or waived with a reason; a waived high-severity one refers to a practice leader.
The check’s decision record holds every control hit, every prepared decision with its facts and model, and every answer with the person who gave it. Two reports are built from it: the client delivery summary, and the proofing record for the E&O file.
Your coverage; what changed at renewal; the documents checked.
Documents and reading; proofing results with forms and tower; the renewal comparison.
The copilot answers over read-only lookups on the check in your scope, every figure checked against what it read, and never takes an action.