AI Pathways’ QQQ-and-gold core: the full test

What we granted, and what we actually tested

Granted, not measured: AI Pathways, the channel run by Brendan, publishes a testable core in How To Actually Use GPT-6 Astra For Trading (Everyone’s Wrong). The video recommends passing criteria before outcomes, costs, next-open fills, splitting data, neighbourhood checks and paper trading. Those are useful safeguards, and the positive primary finding receives full credit.

Tested: An unlevered QQQ / GLD switching account: completed QQQ close above its simple moving average selects QQQ; below selects GLD, our gold proxy. Equality retains the previous target, with GLD as the initial fallback. Those proxy, equality, cost and date choices fill details the source does not specify. Each window starts fresh capital; decisions fill at the next exchange-session open. Primary friction is 5 basis points per traded dollar per side, including entry and terminal liquidation.

§1 · The locked pre-registration

The original protocol was committed as ddc4876 at 2026-09-24 17:20:27 UTC. It received append-only amendments before outcomes: local-window and distribution-bound rules, the exact-specimen limitations, and a correction to input serialization. The final input/config seal is 936fbf4 at 18:37:04 UTC; the execution-code seal is d65f8fc at 18:48:13 UTC. The first IS access starts at 18:48:24 UTC. The table below contains only values present in the d65f8fc configuration blob, checked against the retained file. It does not recast post-run findings as design. The full IS ledger was sealed as 9db2ebf before OOS-search (41422bc), then OOS-confirm (148f7eb), then the overlapping local and full windows (bcb12db and 77f4540). These local records document our release order; they do not undo the creator’s historical selection or establish prospective independence. The cited commit hashes come from our private research repository; they are our own attestation of this timeline, not a public record a reader can inspect.

The falsifiable question

Under our fixed proxy, accounting and dates, does this core retain QQQ-like return while reducing the worst account drawdown?

• A yes looks like: Net CAGR no worse than QQQ minus 2 percentage points AND maximum drawdown at least 5 percentage points smaller, in both long windows.
• A no looks like: Either inequality fails in either required comparison window. A failure limits this implementation, not the complete portfolio.
• Ambiguous residue: An exact headline check remains NOT EVALUABLE without source dates and conventions. Dividend-bound verdict flips would also be NOT EVALUABLE.

Our stated prior: QQQ-like return with a smaller fall, not the portfolio headline

This prior is in the original protocol before results. The result could have contradicted it; it is not a prediction of future performance.

The primary configuration (locked)

Deliberately NOT tested — named in advance

1. Arms 2 and 3 were closed as exact tests before the run because source specifications were incomplete. No substitute funnel or sleeve demonstration ran.
2. No leverage, taxes, deposits or interest; no stock-universe or prospective live test.
3. Exact source dates and return convention remained unavailable; local and bound-disclosed labels are compulsory.

The fixed questions and their distinct statuses

Primary R passes. Exact H is not evaluable, not a failed exact claim. The build-years and later-slice findings are disclosed below; short-window annualized rates are not promises of persistent annual returns.

§2 · The headline arithmetic

§3 · Was it the path? The stress tests

Build-years miss and the two short later stretches — disclosed amendment, 2026-09-25. The table puts the unfavourable build years next to the short later slices and the long overlapping windows. The first later rate covers about two years; the second about a year and a half. These dates already participated in the source’s selection. Annualizing them does not turn them into independent prospective confirmation.

The first later slice’s drawdowns are equal at displayed precision, and R fails both halves there. The approximate headline diagnostic (at least 30% a year with drawdown no deeper than QQQ’s) passes on both later slices, on the first only through that equality; it counts as a diagnostic on short, source-selected dates, never as H evaluated, and exact H stays NOT EVALUABLE. The build years miss the return allowance; the second short slice clears both R inequalities. Across these three non-overlapping periods the risk bar passes once. The long comparison windows overlap them and cannot be counted as additional independent periods.

Fixed settings clear R in 5 of 7 local settings and 4 of 7 full-history settings. PRIMARY stays primary. Price-only omits distributions from both accounts and is never called total return.

Five-day removal is a nontradable stress test. Set the rule’s five best positive-session returns and the identical QQQ dates to zero, preserving elapsed time. These diagnostics do not replace the primary result. The local drawdown advantage misses its bar; full history keeps the drawdown margin but loses the return allowance in all eight variants.

A smaller drawdown is not uniform across calendar years. Across the full history the rule’s within-year drawdown was deeper than QQQ’s in 7 of 23 calendar years (2005, 2006, 2010, 2011, 2012, 2015 and 2020), shallower in 8 and equal at displayed precision in 8, identically in all eight dividend-bound variants; 2004 and 2026 are partial years, and the diagnostics download carries every year. Of the three years shown below, the fast 2020 crash was worse for the rule; the slower 2008 and 2022 declines favoured it. These are within-year drawdowns, not QQQ’s entire 2007–2009 decline. Calendar rows have no registered standalone pass bar.

§4 · The controls

The controls use the same accounting as the core. MA-labelled passive-control cells repeat the same account deliberately and remain in the requested denominator. The cells are correlated, not independent strategies. A passive QQQ row cannot beat itself by the risk bar’s required drawdown margin; its FAIL label is not a claim that holding QQQ is unsuitable.

§5 · The strongest objection and the implementation limits

The strongest objection is that these are selected historical dates. A once-only release stops additional tuning by this analyst; it does not make the creator’s already-selected history independent. Both long windows overlap, all fixed settings are correlated, and the short later periods are especially fragile.

Daily vendor opens are fill references, not proven auction fills at real account size. GLD is our stand-in for gold. Costs are modelled, not a broker-fill audit. We use a simple price average, fractional shares, zero cash interest, initial purchase and scheduled final liquidation. Dividends accrue to prior-close holders on ex-date, become cash at payment-date close, and reinvest at the next open with cost. Unrealized receivables are valued at face value. Taxes and leverage are excluded.

The model counts reinvested distributions separately from split-adjusted prices. Two QQQ quarters in 2005 and one SPY payment date in 2006 remain uncertain: all eight registered combinations are bounds, not recovered facts. No result was chosen from a preferred corner. The full-history primary ranges shown above retain all corners.

§6 · Their own terms

The source’s reported survivor counts and iterative sleeve process are attributed disclosures. We did not reproduce them and do not infer an overfitting penalty without the actual search record. Missing specifications are limits on what this test can say.

All 12 funnel items and all 10 sleeve/process items remain exact-test limitations; none is a measured failed strategy. The whole-portfolio percentage is not a promised core return.

§7 · The complete ledger — all 462 rows

§8 · What would change our mind

1. Exact dates and return conventions. Supply the source window and what the portfolio percentage measures; register that literal comparison before its outcomes.
2. Complete source rules and records. Supply funnel floors, candidate catalogue, cohort mapping, sleeve and portfolio rules and ledgers; missing specifications are not replaced with a convenient demonstration.
3. Corrected evidence or implementation defect. A documented bar, corporate-action or code error reopens affected cells with the original evidence preserved.
4. A prospectively frozen record. New outcomes must be frozen before inspection. Choosing a nicer setting, cheaper cost or endpoint from this panel does not reopen the verdict.

§9 · What survives

Under the registered primary conventions, in both long comparison windows, this published switch historically reduced the worst fall while keeping return within our tolerance of QQQ. That is the measured favourable result. It did not hold in every period: the risk bar held in only 1 of 2 later slices, and in 1 of 3 separate periods once the build years are counted. Settings, costs, period choice and concentrated gains constrain it. Which component of a portfolio can you actually test, and what evidence exists for the rest? That question survives even when the component does well.

§10 · Tools for your own testing

Everything we publish is free — the reports, the result ledgers, all of it. The only thing that costs us real money is the market data.

Prices and the original local corporate-action records came from Massive (formerly Polygon.io): split-adjusted daily QQQ/QQQQ, GLD and SPY OHLC. QQQQ is joined to QQQ by identity. Missing historical SPY events were supplemented from State Street’s historical distribution workbook. Invesco issuer material and the issuer’s SEC-filed prospectus informed the missing-quarter assessment; GLD’s issuer disclosure supports no ordinary income distributions. Those issuer statements do not recover missing events. Registered bound scenarios resolve the uncertainty only for the stated sensitivity verdict.

All buy-and-hold and mixed controls use those same series. Cash and the Sharpe risk-free convention are zero: no FRED, Treasury or other cash-yield series contributes a number. exchange_calendars 4.13.2 (XNYS membership) supplies the session calendar, not prices. EODHD, the affiliate below, did not supply these measurements, and has no say in the verdict.

The diagnostics file has 9,324 rows: 8,736 descriptive calendar-year and pre-2023/2023+ returns, 462 five-day-removal concentration diagnostics and 126 rows where the parent window has no sessions in that interval. Its reason column tells them apart.

Download all 9,324 diagnostic rows (CSV) Download the locked design values (JSON)

Recalculation check: the original account-metric, criteria and diagnostic functions reproduce 462 of 462 archived cells and 9,324 of 9,324 diagnostic rows from hash-matched account ledgers. A planted CAGR discrepancy failed first. This checks numerical consistency, not vendor truth or future performance. No strategy was rerun for this report.

Get every report + the 7 Questions one-pager (free)