We tried to break our own strategy on two more markets. Here's the score.
When a strategy survives validation on one instrument over one year, the honest next move isn't to trade it — it's to attack it. We re-ran the entire gold battery on silver (~20.5M ticks) and EURUSD (~21.7M ticks), every parameter transplanted unchanged, the strategy and filter code literally imported from the gold scripts so no quiet re-tuning was possible, with predictions written down before each run. One layer confirmed beautifully. One half-confirmed. One failed twice.
The scoreboard
| Layer of the claim | Gold | Silver | EURUSD | Score |
|---|---|---|---|---|
| Market structure — direction is a random walk; volatility clusters and concentrates in the London/NY overlap; weekly regime doesn't persist | ✔ | ✔ | ✔ | 3 / 3 |
| Regime-conditionality — breakouts pay only in trending regimes (the tercile signature) | ✔ strong (81% of P&L, t=2.65) | ✘ inverted | ✔ directional, weak year | 2 / 3 |
| Gate forecastability — the trailing-days filter picks the right days in advance | ✔ (+0.185R causal, t=1.70) | ✘ inert | ✘ negative | 1 / 3 |
Silver: the mechanism refused to replicate
Silver's year was even more extreme than gold's (+74.8%, 57.7% annualized vol). The structural physics replicated almost line for line — same random-walk direction, same volatility clustering (|ret| autocorrelation 0.37–0.41), same ~45% overlap concentration, even the same faint multi-day reversion fingerprint. These now look like genuine properties of spot metals, not gold quirks.
The strategy layer did not. The transplanted breakout earned a statistically empty +0.063R (t = 0.80); the pre-registered 5-day filter did nothing (gold: +40% expectancy lift); and the defining tercile signature inverted — P&L concentrated in the choppiest weeks (+68% / −14% / +46% across low/mid/high), the opposite shape. One honest confound: silver's London ranges (~18 cents median) against 2–5 cent spreads mean costs eat 15–25% of risk per trade (gold: under 1%) — silver could barely express any edge with this geometry, which weakens the refutation as well as the confirmation. Per our no-tuning rule, the 7-day filter looking better on silver (+0.148R, t = 1.36) is reported, not adopted — picking it after the fact would be curve-fitting, and "silver needs different parameters" is itself refutation-leaning evidence.
EURUSD: the clean-cost test that split the question in two
EURUSD removed silver's cost confound (spreads ~1–3% of risk) and replaced it with a different challenge: a range-bound year (+1.2% total, 6.5% vol — barely any trends to detect). We pre-committed to the right readout: in a trendless year, annual P&L is uninformative; the tercile view is the test. Result: monotonic in every column — losing in chop (−0.155R), flat in the middle (−0.012R), profitable only in the trending third (+0.108R). The gold signature's shape, on a clean-cost instrument, in the least favorable year. Weak (t = 0.58, few trending weeks), but pointing the right way — the opposite of silver's inversion.
The trailing-days filter, though, failed again: the pre-registered 5-day variant scored −0.143R vs −0.027R unfiltered. In a year of weak, brief trends, the lagged trend state could not find the good weeks in advance.
What we now believe — and doubt
The structural layer is settled: three instruments, two asset classes, one physics. The conditional mechanism — "session breakouts pay only in trending regimes" — has modest independent support (2/3). The weak link, twice-failed outside gold's trending year, is the forecastability of the filter: the claim that you can tell in advance, from the last few days, that the regime is on. The most plausible failure mode of our candidate is no longer "regime doesn't matter" — it's "the regime can't be reliably predicted outside the kind of year gold just had."
That is precisely why the Trial's third test scores the filter's live day-selection separately, on every day including the ones with no trades. A profitable record with a coin-flip filter would mean the filter is dead weight — and the frozen rules won't let us launder that detail away.
Sources: silver_findings.md and forex_findings.md in our research record. Honest limitations carried from the docs: one year ≈ one regime per instrument; EURUSD costs are assumed (0.5–2 pips; we had no recorded EURUSD bid/ask during this period — our recorders do now); descriptive same-week terciles on silver/EURUSD versus gold's fully causal validation pass; small trending-bucket samples everywhere outside gold.