The investment decision
For the No gold account, choosing Lower turnover rather than Minimum risk allocation saved only about $110 annually per constant $1m at the assumed cost rate, while accepting about 27bp more realized tracking risk.
Prefer tighter tracking when mandate fidelity is the priority. Choose lower turnover only when the client's objective justifies the additional risk. A technical PASS is not an investment approval.
Challenge a result
Each case modifies an in-memory copy. Detection requires the relevant economic or structural check to fail—not merely a changed file hash. Restoring the unchanged original inputs must pass.
Inspect the implementation trade-off
Realized gross tracking error
Annualized standard deviation of daily active returns, in %. Lower tracking error is not higher expected return.
Annualized modeled costs
Basis points of portfolio NAV per year; 5bp per dollar bought or sold. Costs funded from the account.
Terminal NAV starts from $1m, with no external deposits or withdrawals; modeled costs are paid from the account. One-way turnover sums half of buys plus sells as a fraction of each day's pre-trade NAV, then annualizes on 252 sessions/year. At 5bp per dollar bought or sold, 100% annual one-way turnover therefore costs 10bp of NAV/year.
Wide table: swipe horizontally, or focus the table and use arrow keys.
Inspect a historical stress window
Selected familiar episodes within the already-examined sample, not an untouched holdout or a forecast. Uses the account selected above.
Wide table: swipe horizontally, or focus the table and use arrow keys.
Active gap is portfolio net cumulative return minus frictionless benchmark cumulative return, in percentage points (pp). Window drawdown includes the starting-close wealth as the initial peak; it is not full-history drawdown.
Independent control evidence
Separate contract, arithmetic and versioned SQL. No production optimizer or validation functions are imported. Inputs and the Ledoit–Wolf numerical dependency are shared; independent code is not independent market data.
Residual and tolerance share the displayed unit. Fractions: 0.01 = 1%; 0.0001 = 1bp. Boolean flags: 0 = pass, 1 = failure. Comparisons are checked values, not independent statistical observations.
Wide table: swipe horizontally, or focus the table and use arrow keys.
What these results do—and do not—establish
Evidence
- Eight ETFs; 32,192 CRSP observations, 2010–2025 source history.
- 2014–2025 evaluation, 30 account/method paths and 4,320 decisions.
- Self-financing trades, prior-close decisions, drifting holdings and independently linked return contributions.
- Traceable block → diagnosis → correction → retest workflow.
Boundaries
- Current-vintage, retrospectively selected universe; already-exposed historical sample.
- No live orders, actual clients, tax-lot engine, stock-level factor model or alpha claim.
- Injected failures are tests, not real production incidents.
- Licensed source rows and detailed security-day ledgers remain private.
Run provenance and tolerances
Tolerance definitions are in oversight_contract.json. Dollar/return/weight units differ; row-count checks permit zero missing rows. “Max error” for boolean controls is 0 or 1, not a financial amount.
Run receipt · Full validation summary · Incident evidence · Methodology PDF / source