Exposures and candidate selection
How owner facts become bounded exposures and a current market shortlist.
Shared words prove little. The contract still has to pay when the business loses money.
Owner facts and hypotheses
Detect labels every exposure by origin:
- Owner stated: the owner named or confirmed this loss. It can proceed to selection and sizing.
- Suggested: the analyst inferred a possible risk from the business context. It can be shown as a hypothesis, but it cannot become a funded position until confirmed.
- Illustrative amount: a display aid requested by the owner. It is never presented as the owner's number or carried into a shared receipt.
Stable exposure references prevent an answer for one exposure from silently updating another exposure with the same risk type.
Live retrieval boundary
Live analysis does not use worked cases as recommendation evidence. Their embeddings support the example corpus and its integrity checks, while the public analyst builds candidates from the selected market generation.
The menu stays small. The application groups open contracts by bounded risk metadata, applies the exposure's driver, relevant local context, contract window and structural rules, ranks candidate series by relevance, and retains a small spread of current strikes; it keeps the complete raw market record application-side and gives the reasoning model only the facts needed to choose among them.
Broad admission, bounded reasoning
The raw generation broadly retains ordinary open non-Sports contracts, including unclassified contracts that may become useful after later semantic work. Selection starts from a narrower menu: closed, expired, malformed and wrong-scope contracts have already been removed.
Automatic metadata can support a funded candidate; curated mappings provide stronger evidence but are not the only route. Every selection still needs a plausible driver, defensible harm-paying side and valid displayed paid-side price.
Worked example: p019
The salon owner explicitly says financing costs are not the real problem and names a $12,500 recession-driven demand loss. Detect retains only the owner-confirmed demand exposure for funding. Suggested financing context cannot override that instruction. The candidate menu includes recession and related macro observations; the 2026 recession contract wins because it matches the stated year and driver more closely than 2027 recession, technical unemployment or one-retailer activity contracts.