Known gaps
What this build does not do, and what a number on a screen does not mean. Read this before trusting a figure.
The world
- Every row is synthetic. Quillmark Mutual, Estmark, its towns, people, garages and claims are generated from a seed. Volumes, delays, costs, shares and fault rates are assumptions chosen to look like a mid-sized motor and property book; none is measured from a real insurer.
- The answer key is written by the same hand as the detectors. Ten of the twelve kinds are planted with a pattern the detector looks for; scoring against that key measures whether the code finds what was planted, not whether real fraud looks like this. The subtle variants (a duplicate under a neighbouring peril, a hit-and-run filed as a collision, a double invoice with tax added, a recovery whose fault is only in the police report) are there so the scores are not perfect; real misses will be of kinds nobody planted.
- Photos are metadata only. No image is read. A real photo check would also compare the image itself.
- One country, one currency, one clock. No cross-border claims, no currency conversion, no daylight-saving change inside the 90 days.
The detectors
photo_before_lossis the weakest check by design: phones with a wrong clock and claimants unsure of the date look like old photos. A referral to Special Investigations needs a person, and photos under a week before the loss only ask the claimant to confirm the date.estimate_above_benchmarktrusts the benchmark table (version 3). A stale benchmark makes every estimate in its band look high or low.ring_indicatorsees shared phone numbers and bank accounts only. Shared addresses, devices, garages without a shared identity, and rings below three households are not seen.stalled_claimcounts any claim event as activity, including the operator's own task in the adjuster's diary. A task nobody acts on makes the file look alive for another ten days.- Tuning was done on seeds 7 and 11; seed 23 is the hold-out the published numbers come from. Three seeds of one generator are not three insurers.
The operator
- No real system is called. Every action changes the DuckDB world; the endpoint on a ledger row names the call a real adapter would make. No adapter exists.
- Rules decide every replayed decision. The model path (an OpenAI-compatible endpoint, asked only where the runbook leaves a real choice) is wired and tested, and was not used for the published numbers.
- The people answering the queue are scripted. Their rules are in
adjust/agent/humans.yaml; their approvals, rejections and reverts are labelled "scripted claims-desk responses (synthetic)" everywhere. - The rulebook assigns by free places only. Skills, languages, licences and complexity beyond the value bands are not considered.
- Nothing is sent. Document requests and notes wait in an outbox; a person sends them.
In shadow mode
With ADJUST_SHADOW=1 every tier 0 and tier 1 action is a "would do" row and nothing changes. There is no answer key on real data: precision and recall come from people reviewing a sample of those rows, which this build does not organise for you.