Clergy housing rules: final verification, Codex (2026-09-25)
Recorded by the orchestrator from Codex's final message (its sandbox cannot write here).
Verdict: FAIL
Codex recomputed all 45 cases. Every computed number matches. US-TC-18 and US-TC-22 fail on the warning property only.
Prior concerns
- Negative deductions from overlapping claims: RESOLVED (CA-05 clamps; CA-TC-14 returns 0).
- US reasonable-pay ceiling: PARTIALLY RESOLVED. It is a warning only; the numeric result is not bounded by reasonable pay.
- Part-year remuneration: RESOLVED (CA-06/07/08; CA-TC-04/12/13).
Findings
| id | severity | rule/case | problem | fix |
|---|---|---|---|---|
| F-01 | critical | US-02/03/06/09/13 | The formulas never enforce the reasonable-pay ceiling that US-09 states; an exclusion of 1,000,000 is possible when reasonable pay is 10,000. | Either add REASONABLE_PAY as a limit, or label the number as provisional and refer. |
| F-02 | important | US-TC-22 | The invariant omits reasonable pay, so no test exercises it. | Extend the invariant and add cases. |
| F-03 | minor | US-13 / US-TC-18 | US-TC-18 lacks reasonable_pay_warning_shown: true. | Add the field or narrow the rule. |
| F-04 | info | all | No other arithmetic or invariant defect found. Clamps, month validation and monotonicity all hold. | None. |
Orchestrator resolution (proposed, pending the Fable final)
Reasonable pay is a facts-and-circumstances test (Pub 517) that the tool cannot compute. So it becomes a required screening question: "Is your total pay (salary plus housing) reasonable for the work you do?"
- "No" or "not sure": REFER (US-09), with no number shown.
- "Yes": compute, and label the result "assuming your total pay is reasonable". Also add the warning field to US-TC-18.
This makes the stated invariant true for every number the tool displays: no number is shown unless the user has affirmed reasonable pay.