How this site works

AI and automation policy

This site is built and maintained with substantial automation. Here is precisely where, and precisely where not.

Short version

Automation is how this site is built: it researches sources, extracts facts and records the excerpt behind each one. Deterministic code, not a language model, produces every number in our calculators. The publishing design holds anything uncertain for a human rather than publishing it. The pipeline that would run those checks on a schedule is described below as a design, not as something operating today.

Status of the systems described here

4 of 4 switched on

Running now

Automated editorial runs, QA checks and publication blockers

Re-verification of stored facts against their sources, run on demand rather than on a schedule

Change history recorded for every field change

A human review queue where held items land, which a person opens, reviews against the evidence and settles

Where automation is used

These are in use today.

Finding and retrieving primary sources.

Extracting individual facts from those sources and recording the excerpt.

Drafting explanatory prose from an evidence package.

These are what automation is designed to do, and will do when the pipeline runs.

Detecting when a stored fact has gone stale and re-checking it.
Comparing a re-checked source against what we already stored.
Checking pages for duplication, missing evidence, technical errors and disclosure problems.

Where automation is not used

Calculations

Every figure a tool produces comes from explicit, tested, deterministic code. A language model may help explain a result. It never produces one.

Community content

We do not generate reviews, questions, answers, usernames or experiences. Everything in a community section came from a real person.

Authors and reviewers

We do not generate people.

Facts from memory

A model is never asked what it knows about an insurer. It is only asked to work from documents we retrieved.

Generated images

Every image on this site is a generated illustration: one at the top of each of the seven insurance hubs, and one at the top of each of the eight guides. All fifteen were made with Recraft V4.1 (pro tier), via Freepik on 2026-10-03, and there are no others.

The rule we apply is about what a picture claims rather than about how it was made. A flat geometric drawing of a car is obviously a drawing: nobody reads it as evidence and there is nothing in it to be misled by. A photograph claims far more. A generated photograph of a satisfied customer, an adviser at a desk, an office, or a damaged house after a storm asserts that a person, a place or an event exists, and every one of those would be a fabricated trust signal. That is the same failure as inventing a rating, and none of them appear here.

So the illustrations were generated under explicit constraints, and the full instruction given for each one is recorded in the source alongside the image: No people, no faces, no hands, no text or lettering, no logos or brand marks, nothing photorealistic. Each image was checked against that list before it was published. Their backgrounds were cut away afterwards so they sit on the page; that is the only edit made to them.

The guide images go a step further from depiction: each is a diagram of the idea the guide explains rather than a picture of anything. A bar split in two for a deductible, a wedge cut out of a circle for an exclusion, a timeline with a mark before it starts for a pre-existing condition. The guide on what to do when a claim is denied could have been illustrated with a photograph of someone distressed at a kitchen table. That picture would have invented a person and a story in order to borrow feeling it had not earned, which is the thing this policy exists to rule out.

No photograph on this site depicts a real customer, employee, adviser, property or event, because there are no photographs on this site.

Limits we place on publishing

The configured ceiling is one new indexable page per day, and it is a ceiling rather than a target. The design expects most automated runs to conclude that the right action is to update an existing page or to do nothing at all. Creating a page merely because a search phrase exists is not a valid reason, and the system is built to refuse it. No automated run has published a page here yet, so this limit has not yet had to bind on anything.

Human control

The design puts a human above automation: any automated decision can be overridden, any proposed change edited, and any published change rolled back. The schema carries a change history recording each change with its source, its reason, whether a human or automation made it, and its check results. That log is not being written yet, because the automated runs that would write to it are not running yet.

What automation has not done

It has not replaced human editorial review, because no human editorial review has happened yet. We are not presenting automation as though it were oversight.