Assurance demonstrator

One architecture. Two legitimate answers.

This demonstrator computes what a contract requires and tests what evidence licenses. It does not infer a causal effect from settlement data alone.

01 · Requirement observed

Move the assumption. Watch the boundary.

An unnamed Medicare ACO with 8,838 beneficiaries settled at an underlying expenditure below its benchmark and elected a 2.00% threshold. The contract fixes the requirement and what turns on it. The true causal effect remains unknown.

Observed gap
$6,359.76
Corridor
$693.54
Causal-necessity requirement
$5,666.22
Paid per beneficiary
$4,452.99

$6,359.76 = $5,666.22 + $693.54

Attributable per beneficiary
$2,101
Attributable in total
$18.6M
Payment required this performance
No

Below the boundary, the untreated counterfactual would also have cleared. Under this declared state, only the share added by the intervention is attributable.

Contract values and portfolio diagnostics are reproduced from source-pinned data. Slider positions are hypothetical causal states, not estimates. Amounts displayed in prose are rounded; the underlying source values drive the calculations.

02 · Estimation refused

What it looks like when the answer is no.

A consumer lender recorded 40,000 applications, their outcomes, and the covariates used by its four-tier pricing rule. Yet within every tier, assignment was so consistent that no usable comparison survived.

stratum       n       treated   support   verdict
------------------------------------------------------
Tier 1     17,448     17,099       7%     CONSUMED
Tier 2      8,265        194      11%     CONSUMED
Tier 3      8,977        559      15%     CONSUMED
Tier 4      5,310          0       0%     CONSUMED

strata estimated          0 of 4
population covered        0.0%

REFUSAL: no interval exists to compare with a decision boundary.

The refusal is the finding.

This is not estimator failure or missing outcome data. More observations generated by the same assignment rule would not restore the missing comparison. Other identification designs may change the answer—for example, preserved randomization, a defensible instrument or discontinuity, or explicitly justified extrapolation. Each requires its own assumptions.

03 · Enforced discipline

What the architecture will not do.

  • Report a causal effect without a counterfactual and written identification model.
  • Estimate where characterization found no usable comparison.
  • Let a conditional result masquerade as a plain true-or-false answer.
  • Approximate a payment rule outside the contract class being modeled.
  • Report a requirement for a contract that was never paid.
  • Accept a threshold without a stated unit.

From example to engagement

Your contract. Your evidence. A saved assurance record.

The public demonstrator reveals behavior. A private run executes the architecture against the real decision.

Discuss a private run ↗