Claims development · bitemporal standing

Late evidence

In insurance, exposure, loss, claim and recovery each become known long after the one before. A determination made on what was available then can be defeated by what becomes available later. Both must stay on the record: what was warranted at the time, and what is warranted now.

Two times, kept apart

• Valid time is when the crossing takes effect.
• Record time is when a determination is made. A determination uses only evidence that was available at its record time; each evidence item may declare when it became available.
• Re-running the gate at an earlier record time reconstructs what was known then. A later determination references the earlier one and never edits it.

Worked example: INS-005 proposed

Denial warranted when made; a model withdrawal learned later defeats it. Synthetic.

• On 10 September the denial was warranted: the registry entry for the model was the evidence available, and it was current. The denial took effect minutes later.
• On 18 September the registry published a withdrawal of that model, effective 5 September, before the denial. The evidence arrived late, but it concerns the past, so it is checked against the valid time: the model version was already withdrawn when the denial took effect.
• The original determination still replays to the same receipt as of 10 September (PASS). Nothing was wrong with it given what was known; it is defeated, not falsified.
• Because the denial already took effect and is held now, it counts toward epistemic IBNR in the portfolio: incurred, and only now known to be unwarranted.

From defeat to an insured event

A defeated standing is an adverse epistemic finding. On its own it is not an insured event. The proposed trigger requires every part of F(C, R_C, P, T) ∧ defined loss: the determination C with receipt R_C was relied on for purpose P, it failed adjudication test F before time T, and that failure caused a defined loss. Reliance and causally attributable loss are what an insurer can observe and a receipt alone cannot supply.

Synthetic scenario COVER-INS-005: Denial relied on to release a claim reserve, later defeated, followed by a settlement loss. Scenario JSON · schema

TRIGGERED under these declared terms. Remove the loss record and the same defeat is NOT_TRIGGERED; remove the reliance record and it is NOT_TRIGGERED.

Structure challenged. CHALLENGED: an observed consequence or effect has no represented path, so this result may omit losses the model cannot express. CON-002 (independently observed) follows no declared consequence path.

• No observed effect ≠ no effect; no modelled path ≠ no possible path.
• Evaluates declared terms against declared records; it is not a coverage decision.
• Causal links are declared by the parties, not verified here.
• An adverse epistemic finding alone does not trigger: reliance and a defined, attributable loss are required.

Two kinds of uncertainty

U_V is uncertainty about observed values inside the declared consequence model: an amount not yet reported, or disputed. U_S is uncertainty about whether the model contains the paths that matter at all. Neither can be inferred from the other. Better measurement reduces U_V; it cannot, by itself, repair U_S. They are reported as a pair and never combined into one confidence number.

• The pending settlement is U_V: the path exists in the model (a wrongful denial is settled); only its amount is unknown. Waiting for the report resolves it.
• The ombudsman fee is U_S: no declared path in the claims class carries a regulator’s fee. Measuring more precisely cannot fix that; the class’s consequence model has to be revised and tested again.
• Before the fee was observed, structure was untested, not supported. No observation had challenged it, and no structural test had been run. An absence of challenges is not support.
• What this example demonstrates is narrow: the fee record states that it follows no declared path, and the site flags it. The site does not work out for itself that a path is missing. Noticing an omission without being told has been shown once, narrowly, in the refund fixture (R29 PASS, narrow: frozen event kinds and an independent provider-fee challenge): paths frozen in a commit before the observation existed, then a fee the provider recorded in its own log, reported as a challenge to the frozen structure. That detects unaccounted event kinds in an independent log. It would miss an effect absent from every log, or one folded into a kind the model already accounts for.

No observed effect ≠ no effect; no modelled path ≠ no possible path.

Standing effects

Causes: VALID_TIME_MOVED, EVIDENCE_ARRIVED_LATE, AVAILABILITY_DIFFERED, GRAMMAR_CHANGED (class or policy changed), PROPOSAL_REVISED. In the workbench, reassessing a proposal shows its standing change against the previous determination.