The four branches: one edge, four ways of monetizing it.

The family portrait

Four architectures, from banked certainty to fat tail.

Normal is the core branch — the architecture the simulation itself is centrally built around: half the position banked at 1R, the remainder run to a static 2R target — a structure built for ordinary continuation, monetizing early and capping ambition. Trend Partial is the hybrid fat-tail accelerator: it keeps Normal's 1R partial cushion but replaces the static target with a trail that unlocks at 1.6R — one foot in banked certainty, one foot in the tail. Trend No-Partial is the true fat-tail accelerator — not a hybrid: no partial at 1R at all, the full position held to the 1.6R trail unlock. It exists for geometric account growth, accepted alongside the highest variance in the family, which is why its use is earned through quota and evidence rather than merely chosen. Overflow is the stability branch: a small partial at 0.75R and full closure at 1R — the shortest, most certain monetization in the family. Its consistent 1R take-profits exist to supply stability when variance runs high or EV runs low, steadying the account while the accelerators do their volatile work.

Why identity is fixed at entry

Branch drift is how evidence gets laundered.

The branch is selected before entry and does not change mid-trade — the selection matrix names the violation exactly: turning Normal into a trend trade after entry is branch identity drift, not adaptation. The rule protects the evidence system more than the trade. Each branch's EV is computed from its own population — Normal's 1R and 2R hit rates, the trend branches' 1.6R continuation behavior, Overflow's 0.75R and 1R conversion — and a trade that switches architecture mid-flight lands its outcome in a population whose assumptions it didn't live under. A few drifted trades per month is enough to make every branch's hit rates quietly wrong, which makes branch EV wrong, which makes the blended EV and the deployment decisions built on it wrong. The discipline is one sentence at entry; the alternative is a corrupted denominator everywhere.

Different questions per branch

Each architecture is graded on its own conversion chain.

Because the exit structures differ, the evidence questions differ, and the Weekly Summary tracks each chain separately. Normal answers: how often does 1R hit, and — conditional on 1R — how often does the runner complete 2R? Trend Partial answers the 1R question plus the trend-specific one: how often does the 1.6R unlock arrive and what does the trail harvest after it? Trend No-Partial lives or dies on the continuation rate to 1.6R and beyond — with no partial cushion, its whole expectancy is the right tail showing up often enough. Overflow answers the humblest chain: 0.75R and 1R conversion, nearly binary. Reading a branch against another branch's chain is a category error the tab structure makes hard to commit — each branch's columns carry its own probabilities, and no summary statistic averages across them until the blending layer does so deliberately, with weights.

• Normal's health is a two-link chain: P(1R), then P(2R | 1R). A strong first link with a weak second is a different diagnosis than the reverse.
• Trend No-Partial's variance is structural, not a flaw — its sample needs more time before any verdict, and its use is quota-governed for exactly that reason.
• Overflow's near-binary structure makes it the cleanest branch to read and the worst to extrapolate from.

The family's logic

Coverage of outcomes, not prediction of them.

The reason the family has this shape is that trade outcomes have this shape. Some continuations run one clean leg and die — Overflow monetizes those completely instead of round-tripping them. Most run a moderate distance — Normal's banked-half-plus-2R is tuned for exactly that middle. Some become genuine trends — the Trend Partial structure claims them while keeping a cushion. And a few become the outsized moves that pay for everything — Trend No-Partial exists so the system has a vehicle that doesn't sell those early. The operator's discretionary skill is matching the trade in front of them to the right vehicle; the system's job is keeping each vehicle's record clean enough that the matching improves with evidence. Four branches is the smallest family that covers the outcome distribution without asking any single structure to be right everywhere.

The key idea

Branches are the unit of truth below the system.

Almost every serious diagnostic question in MARS decomposes to the branch level before it resolves: whether the edge is intact, where a weak week came from, whether a profile's weights still fit the market being traded. The four architectures are what make those questions answerable — separate populations, separate conversion chains, separate EV, blended only at the end and only by explicit weights. Declare the branch honestly at entry, manage it as declared, and the family does the rest. Blur it, and the system's sharpest instrument goes soft first.