What It Took to Build This Desk

How this is estimated (from the box's own timestamps): the desk was touched on ~87 of its 100 calendar days — nearly every day — producing 1,900+ scripts, tools, and research write-ups. Git records 325 commits, but only across the June–July build phase; the entire research arc after that ran in working sessions and scratch scripts that were never committed, so commits undercount the work.

At ~6 engaged hours per active day — and the heavy days ran from early morning to late night — that's on the order of ~520 hours of hands-on build-test-review time over three-plus months.

Put another way: the finished body of work — 129 survivorship-free studies, 6 paper bots plus the HOUSE real-money bot, and a 34,220-name data warehouse — is what a traditional quant analyst would spend well over 1,000 hours (several months full-time) to produce by hand. The machine compressed months of research into weeks; these figures are honest estimates grounded in file and commit timestamps, not a time clock.

Every idea here was built, backtested on survivorship-free data, charged realistic costs, and split into in-sample / out-of-sample with random-entry controls. These are the ones that did not survive — grouped by family, with a few of the marquee autopsies.

The single fact that shapes the entire desk: almost everything you'd own is the same bet. Measured on 13 years of daily returns — and again inside the two stress regimes, because correlations change exactly when it matters. (Live, refreshed weekly, as the heat-map on the Charts page.)

Updated 2026-09-30: Atlas, Desk Blend, The Book and Balanced Futbot were retired 2026-09-27; today’s real-money answer is HOUSE v2.9. The ranking below is kept as it was written.

Not a backtest — the four-voice council's opinion (Bill the skeptic · Oracle the structuralist · Claude the builder · Athena the return-seeker) on where real capital should actually go, ranked by return-per-drawdown, robustness, and honest diversification value. The engines are at their ceiling and roughly equal on quality, so this is about fit and conviction, not a winner-take-all — and the council does not fully agree, which is the point.

We ran an adversarial "beat this bot" game on every live bot: build a challenger, prove it wins on the desk's own terms (return first, drawdown second), survivorship-free and costed — and all three reviewers must agree. The pattern was decisive: you can't out-signal the validated designs. No new strategy beat any engine. The only improvements that survived were small, free, or diversification tweaks — and one apparent "win" turned out to be a survivorship-bias trap, caught and reverted.

Getting the Ark ready for the real IBKR accounts forced three honest questions. Where should each piece live for tax? The dip sleeve realizes almost all its gains short-term every year — so it came out of the taxable account. Does it earn its place anywhere? Tested across the whole household, dropping it everywhere added ~1 point a year at the same Sharpe — so it was retired, and its 20% went to the engine. Can the engine run hotter? A dial sweep found only two robust levers (vol-target and Bitcoin weight); raising the vol-target 50% → 65% added ~3 points a year for ~2 points more drawdown, Sharpe unchanged, peak margin still under 2×.

The refined book — engine + gold + managed futures + Bitcoin — backtests at ~50%/yr, −26% max drawdown (2019–26, now charging margin interest). A new margin guard cuts the book to 1× if the account cushion ever drops below 35%. Real money maps as Taxable = Ark (margin), 401K = the same Ark via 3× ETFs at ⅓ weight (same exposure, no borrowing), Roth = Gringotts parking in DBMF instead of T-bills — running in dry-run first. Together: ~56%/yr, −29% worst drop in the 2019–26 backtest, vs 47%/−28% before — but a 30-year Monte Carlo replaying 1999–2026 (dot-com and 2008 included) puts the household's core at ~14%/yr, so the plan is to expect ~14–19%/yr. Its real edge showed up as survival: the smallest chance of being behind after 10 years of any aggressive option (1.6%) and a typical worst drop of −33% vs QQQ's −66%. Honest caveat: 2019–26 was the best tech/Bitcoin window on record; plan on far less.

If no new signal can beat the engines, the last gain left is how you combine them. Atlas is the answer: the Ark as a dominant core (70%), with Balanced Futbot (20%) and Gringotts (10%) as small, uncorrelated satellites — sized so crypto can lift the return without sinking the ship. It built nothing new; it was a paper overlay of three bots then running. Retired 2026-09-27 — HOUSE replaced it.

In a paired Monte Carlo (5,000 ten-year paths), Atlas beat the Ark on terminal wealth in 76% of worlds — ahead at every percentile, including the unlucky downside — at a catastrophic-drawdown tail that's statistically identical. More return, same risk (as long as crypto doesn't go to zero). The improvement the beat-the-bot games couldn't find in a signal, portfolio construction found in the sizing.

Atlas V2 (Sep 2026): a follow-up test asked whether Atlas is truly diversified — and found Futbot is the least-diversifying satellite (~0.56 correlated to the tech core; essentially more tech-trend in disguise). V2 swaps it for a managed-futures sleeve (DBMF — genuinely crash-positive, up +21.6% in 2022 while tech collapsed) and lifts BTC to 20%: 65 / 15 / 20. Same return, shallower drawdown (−27% vs −37%), and the best return-per-drawdown (Calmar) on the desk. It ran head-to-head with V1 on paper; both were retired 2026-09-27 when HOUSE took over.

Survivorship-free data. Every backtest ran on a warehouse that includes the losers and the delisted — the single most common way backtests lie.

Prove it, or bury it. Realistic costs on every trade, in-sample / out-of-sample splits, Monte-Carlo stress across thousands of shuffled histories, and a random-entry control on every "edge" to make sure it wasn't just the market.

Pre-registration. The news-sentiment trial was frozen and forward-tested with its kill-criteria written down in advance — so the "dead" verdict couldn't be rationalized away.

Three voices on every idea. The builder, a structural strategist (Oracle), and a professional skeptic (Bill) whose job is to kill fragile ideas. Nothing shipped without all three agreeing — and they killed a lot.

Outside red-team. The record was also stress-tested by an external AI reviewer. When it argued the intraday graveyard had only killed price signals, we ran the missing order-flow / cross-sectional / auction tests on full tape data — which sharpened the conclusion rather than overturning it, and surfaced the one open auction lead.