The demo data

Every claim in the pilot is synthetic and built for demonstration. Several hundred are loaded, across four lines of business, in GBP for the UK product and USD for the US one.

Generated against the taxonomy

The generator (CLI-only; see Generating claims) writes each claim against the Davies DGPH-34 claims-processing taxonomy: client and handling rules, policy, claim data, money, third parties, litigation, investigation, fraud indicators, compliance, feedback, diary, tasks and documents. It is held to minimum depth: at least 8 diary entries, 6 documents, 5 tasks and 5 involvements, and payments whenever money has been paid.

Scenarios and gotchas

Each claim is written around one of 34 scenarios (api/src/generator/scenarios.ts): 10 auto, 11 property, 10 workers compensation and 3 general liability. A scenario names:

• a brief: what the claim is and what it should let the chat demonstrate;
• the capabilities it showcases;
• the gotchas to plant in the record itself, in the diary, documents, payments or policy wording, so the assistant has something real to find;
• a one-line demo hint.

Some of the planted gotchas:

• a policy that lapsed a week before the loss;
• a contractor paid twice for the same job;
• a bereaved policyholder nobody flagged as vulnerable;
• three expressions of dissatisfaction that were never logged as complaints;
• a supplier's email that tries to give the AI instructions;
• a subcontractor indemnity sitting unused.

The full list, with every gotcha, is the generated scenarios reference.

Demo aids on screen

• Demo hint. A floating tip bottom-left on each claim says what it is good for, with one-click capability buttons and a spoiler list of the planted gotchas. It comes from metadata.demo.
• Random claim. Landing on / opens a random claim in the region's currency. Random claim in the top bar and Next claim in the banner do the same. The id is always in the URL, so any scenario can be bookmarked or shared.
• Claims Data (/data). What is loaded, the models in use, the scenario catalogue with counts per scenario, and the generator command.

Fixtures

packages/shared/src/fixtures/ holds three hand-written claims, one each for auto, property and workers compensation:

Id	Line	Scenario
CP-AUT-260381	Auto, GBP	Policy lapsed at loss
CP-PRO-260214	Property, GBP	Flood exclusion and a duplicate invoice
CP-WCP-260097	Workers compensation, USD	Late reporting, a return-to-work and a provider pattern

They are the generator's worked examples of depth and tone, and the seed data for the tests. The app never loads them on its own; the end-to-end suite seeds them into its throwaway database.

Quality

Flash Lite produces good records when held to the depth minimums; a richer GENERATOR_MODEL produces richer records at more cost. Read a sample from each scenario before a big demo and delete anything weak. There is no delete endpoint; use SQL against the local store.