Every paper here got tested.

The design, the pre-registration and the decision rules were drafted by an AI model and adopted by us; an AI agent ran an adversarial audit of the design before the reproduction test and found two defects that would have let the paper overstate a change. Both are fixed and reported. The reproduction test came first: the 2016–2019 data had to show the published sign before any later statistic was computed.

The honest part is Table 3. The 61% bound on the decline holds for the registered window; it is 90% when the window ends at 3:55, 91% over all sessions from 2020, and 2024–2026 alone cannot exclude a complete disappearance. We also release a table of first single-stock leveraged-ETF launch dates sourced from issuer documents, because the aggregator we first used placed nine launches up to ten months too early.

The gates were assembled over three days, most of them after a failure. The first fact-check failed its pre-registered bar by 0.013 against a label that turned out to be wrong; a real earnings calendar, parsed from the news feed itself, gave a second pre-registered hearing on rows nobody had opened (0.981 on 39 rows). The dumbest baseline, a headline count, scored 0.889 on the same rows, so the judge’s margin is borderline and the paper says so.

Most of what went wrong was plumbing, not the model: an unpaged feed that turned every six-hour window into the last hour, a window ending at the wrong clock that changed the answer on identical rows, a stamp test that searched source text for a function name and tested nothing. Table 1 lists all 22 uses of the judge and their verdicts.

This is the companion to the first paper. That one asked how often a backtest approves noise; this one asks whether the backtest is even running the same system as live. All six defects sat where two paths that should agree had drifted apart — a historical request and a same-day one, a field that exists only in the backtest, a data list built from what is traded rather than what is required. In five of the six, the path closer to live was right.

The uncomfortable finding is methodological: none of the six was caught by reviewing results. One erred in both directions, one moved almost every number and flipped no verdict, one changed a single decision in twenty, one produced a false pass when two errors cancelled. Written lessons did not stop recurrence — the clock trap was logged months before it came back — and one class recurred past its own test. The paper’s answer is structural: build the replay’s data request from the live path’s own list, make every missing input loud, re-apply the binding limit after every step that can increase size, and chase every unexplained parity divergence to its root.

The simulations are fully public and synthetic, so the numbers can be said out loud. Under a pure null with twenty variants per idea, tuning on the full sample falsely accepts 36.9% of candidates — and selecting on a reused holdout is the identical quantity, because a holdout consulted for selection is not a holdout. A disciplined single holdout achieves its nominal 2.3%; a trial-count-deflated hurdle achieves 0.4%. The popular fold-consistency heuristic passes 88% of noise — near-vacuous once selection is accounted for.

Against an era-locked alternative — a real effect confined to one regime — the standard stack fails almost completely (92% acceptance), while leave-best-era-out collapses acceptance to 1.2% at a null size of 0.04%. The price is measured rather than hidden: detection of a stable SR 0.5 on eight years of dailies is 28% under deflation, 7% under LBEO, and falls further when trial variants are realistically correlated. Severity is purchased with power at exactly the effect sizes that survive costs.

This is the paper we have spent the most research hours interrogating, and the field report is two-sided.

As a lens, it is permanently adopted. Decomposing every diagnostic into overnight and intraday components — the paper's basic move — changed how we read our own results: several behaviors that look like noise in close-to-close terms resolve into clean structure once the clock is split. The clientele framing also gives the most convincing answer we know to the question every systematic researcher gets asked: why would an edge survive being published? Because the other side of the trade is not uninformed — it is institutionally committed to trading at a particular time of day, and that commitment does not read papers.

As a timing overlay, it did not convert — for us. Overlays motivated by the paper's smoothed overnight-minus-intraday spread went through our standard protocol and no admissible configuration produced a single qualifying event out of sample (working note No. 9 records the funeral). We stress the scope of that sentence: our horizon, our instruments, our admission rules. The phenomenon in the paper is documented at monthly horizons on broad cross-sections; our failure to convert it says nothing about the paper's claims and a great deal about the distance between a documented premium and an adoptable signal.

Reading advice for practitioners: the introduction, Fig. 2, and Table 2 carry the argument; the international replication is the underrated part — nine markets with the same anatomy is the kind of robustness no backtest of ours will ever have.