Ladder experiment: calculation checks
7 September 2026
Separate computational checks reconcile the raw CSV files for all three 1,000-path cohorts with the reported rung counts, payment components and path ledgers. The checks include every recorded simulation path. These are numerical verification procedures, not external peer review.
| Initial pool | Simulated paths | Pool paths stopped | Focal eligible rung observations | Completed | Unresolved | Rungs meeting the display rule |
|---|---|---|---|---|---|---|
| 50 | 1,000 | 886 | 20,936 | 20,857 | 79 | 1–24 |
| 200 | 1,000 | 153 | 20,974 | 20,972 | 2 | 1–37 |
| 500 | 1,000 | 5 | 20,585 | 20,585 | 0 | 1–37 |
The display rule requires all survivor-eligible outcomes for that rung to be completed and at least 30 observations. The last three rungs in the larger pools fail the sample-size requirement, despite complete coverage of the observed eligible releases.
Of the 886 stops in the 50-member pool, 596 follow the conservative two-member boundary and 290 occur at an infeasible exposure share. Corresponding counts for the 200-member pool are 25 and 128. All five stops in the 500-member pool occur at an infeasible exposure share. These are pool continuation outcomes; the focal member's unresolved payment count is a different quantity.
The model executed and audited 728,300 single-death allocations. Its largest numerical fairness residual was £2.69×10⁻⁹. The largest transfer-conservation residual was £3.73×10⁻⁸ and the largest whole-path ledger residual was £1.20×10⁻⁷. Each focal example's five payment components sum to the reported payment within £1.82×10⁻¹².
The independent funding calculation agrees with the implementation within £4.27×10⁻¹³. Additional deterministic cases check the dense expected-transfer equations, no-death investment returns, death during the distribution year and treatment of a stopped but survivor-eligible path. All passed.
Scope: these checks support the reported experiment under its declared synthetic mortality, investment and allocation assumptions. They do not establish complete lifetime funding, a guaranteed target, interval-certified fairness, an unordered batch policy or a contractual solution for retained assets when the research calculation stops. The three sizes use different fixed cohorts; their results do not isolate pool size as the only experimental variable.
The source and results package includes checks/audit_model.py, which reads the simulation source and raw outputs, writes model_audit_results.json and runs the deterministic checks described in checks/model_audit_note.md.