Statistical truth first: the gate battery every EV must pass.

The gatekeeper

The minimum sample gate — below it, everything is provisional.

The first check is brutally simple: the control panel carries a minimum sample gate, and any lens whose selected N falls below it gets its outputs treated as provisional regardless of how they look. This single knob defuses the lab's most seductive failure mode — filtered brilliance. Stack enough filters (this pair, this session, this ATR state, this stop-width bucket) and some tiny sample will always look spectacular, because with few enough trades, something always does. The troubleshooting table names it directly: a selected lens that looks too strong should send you to the sample size, the trimmed EV, and the active filters before it sends you anywhere near a conclusion. The gate doesn't say small samples are worthless. It says they're not yet evidence.

The confidence pair

Lower bound and P(EV>0) — the edge's worst honest reading.

The second rank of the battery asks not 'what is the EV?' but 'what is the EV when we account for how little we know?' The EV lower bound is the conservative floor after statistical confidence adjustment — the number the sample supports even if the luck in it is discounted. Probability EV-above-zero expresses the same humility as odds: given this sample's mean and dispersion, how sure are we the true edge isn't zero or worse? The manual's troubleshooting row captures the operational meaning perfectly: positive EV with a negative lower bound means an edge exists but is not statistically secure — keep trading and reviewing, but do not scale from this evidence alone. The realized EV is the claim. The lower bound is what the claim can currently prove.

The outlier strip

Remove the best trade and see what's left standing.

The third rank interrogates the shape of the evidence. Trimmed EV and winsorized EV cap the extreme tails; median EV reports the middle trade's truth; and the bluntest instruments — EV without the best trade, EV without the worst — answer the question every honest operator should fear: is this edge a distribution, or one memorable Tuesday? A system whose expectancy collapses when its single best outcome is removed isn't necessarily broken — fat-tail branches are supposed to earn their keep in the tail — but it is fragile in a specific, known way, and the tail-reliance and skew-dependency reads quantify exactly how much of the claimed edge is riding on rare events repeating. The battery doesn't forbid tail-dependent edges. It forbids not knowing you have one.

• Tail reliance is a property to know, not automatically a flaw — Trend No-Partial exists to be tail-reliant, on purpose and on quota.
• An edge that survives trimming, stripping, and the median read is the kind the deployment system can lean on.
• The comparison against the all-blended baseline runs at every rung: a lens is judged relative, never in isolation.

Why this family goes first

Every other family's diagnosis assumes this one passed.

The workflow's ordering — statistical truth first — is logical necessity, not preference. Mechanical leakage analysis assumes there's a real edge to leak; behavioral analysis assumes a real edge for behavior to degrade; context slicing assumes a real aggregate to slice. Run those diagnostics on a statistically insecure sample and they'll happily produce findings — capture ratios, post-loss deltas, session pockets — all of them noise wearing the costume of insight, each one an invitation to repair something that was never measured. The battery's pass isn't a formality before the interesting analysis. It's the license that makes the interesting analysis mean anything.

The key idea

The lab's first loyalty is to what the evidence can prove.

Everything in this battery serves one sentence from the manual's closing standard: protect MARS from false confidence. The realized EV flatters; the sample gate, the lower bound, the probability read, and the outlier strip each remove one specific way of being fooled. What survives all four is a number the rest of the system — the scorecard's verdicts, the benchmark's comparisons, the edge-change threshold's second pillar — can safely consume. Statistical truth first isn't caution for its own sake. It's the foundation everything downstream is standing on.