Self-test annotation (Claude Opus 4.8, max effort, Claude Code headless, 3 trials, all-or-nothing grading):
All three runs produced valid, schema-complete JSON; failures are narrow near-misses, not malformed output.
Run 1: 10/14 fields exact (missed the fee-audit pair and the payout/residual pair).
Run 2: 12/14 fields exact (missed only fee_overcharge_usd + fee_discrepancy_charge_refs — fit one global
       fee changepoint where the true schedule changes per currency).
Run 3: 12/14 fields exact (missed only net_expected_payout_usd + unreconciled_difference_usd — double-counted
       two duplicated settlement-export lines; the deviation equals their value to the cent).
Both wrong answers correspond to pre-registered, data-refuted alternative models documented in the package.
Bucket rationale: 5-10% reflects the Opus near-miss pattern together with a supplementary 23-run
Claude Sonnet 4.6 measurement (2 passes, ~9%); those logs are available on request.
