Operator brief · 171

Ask whether the run is trustworthy before asking what it says.

The key idea

The hidden interval

Every simulated statistic is an estimate with a spread.

A completion probability reported to three decimal places invites a precision the number does not possess. It is an estimate produced by sampling a finite number of paths from an infinite space of possible orderings, and like any sampled estimate it carries an interval around it. Run the same configuration with a different draw and the figure moves. The workbook does not hide this — it ships path stability measures, a confidence grade, and explicit error bands — but the headline tables do not carry the interval on their face, so the burden falls on the operator to remember that the digits after the decimal point are ornamental until the spread is known.

The dial

Path count buys precision, and precision is not accuracy.

More paths narrow the interval. The relationship is unglamorous and reliable: a ten-thousand-path exploratory run answers rough questions quickly, and the fifty-thousand-path standard exists because that count produces bands stable enough to be quoted. What additional paths cannot do is correct a misspecified configuration. A run built on a non-monotone ladder or a drifted branch mix converges beautifully on the wrong answer, and it converges harder the longer it runs. This is the distinction worth carrying: path count controls how tightly the simulation estimates its own model, and nothing whatsoever about whether the model resembles the world.

FigureError band width against path count — schematic convergence

Illustrative geometry: the interval narrows with more paths and settles rather than vanishing. The dashed reference marks a difference too small to distinguish from sampling noise at exploratory counts.

The reading position

Consulted last, and only when the result is marginal.

The manual's recommended reading order after a run puts the confidence table at the end, with a condition attached: if the result is marginal, check it before acting. That placement is a piece of practical wisdom rather than an oversight. Most runs are not close. A configuration with a lock rate near ninety percent does not need an error band to be actionable, and demanding a statistical ritual before every obvious answer would train operators to skip it entirely. The check earns its place by being reserved for the cases where it changes something — which are exactly the cases where the temptation to act on a small favourable difference is strongest.

The cheap pre-check

A fast in-sheet estimate that is explicitly not a simulation.

The workbook also holds a lightweight analytic approximation that runs no paths at all — a sanity estimate available before committing to a full engine run. Its value is triage: it catches the configuration that is obviously non-viable before anyone spends time on fifty thousand paths, and it gives a rough expectation against which a real run can be checked for gross error. Its labelling matters as much as its output. Because it performs no path simulation, it says nothing about sequence risk, drawdown distribution, or lock probability — the entire class of questions the Lab exists to answer. Treated as triage it is useful; treated as a shortcut it silently answers a different question than the one asked.

The failure it prevents

Thin differences are how research programmes go quietly wrong.

The damaging error is rarely a wildly wrong result, which tends to get caught. It is the small favourable difference — a slightly better completion rate, a marginally shallower adverse band — that is inside the noise and gets promoted anyway because it pointed the way the operator was already leaning. Do that repeatedly and a research programme accumulates a stack of changes each justified by a difference that was never real, with no single decision identifiable as the mistake. The confidence grade is the cheap instrument that interrupts this, and it interrupts it at the only moment interruption is possible: before the marginal result becomes a memo.

The key idea

The last question about a run is whether it was a run worth reading.

Research discipline in the Lab reduces to a short sequence: declare the configuration, clear the gates, control the draw, and then — when the answer is close — ask whether the instrument can actually resolve the difference being claimed. Skip the last step and the sandbox still produces numbers, still produces confidence, and stops producing knowledge. The boundary between research and deployment is only worth defending if what sits on the research side has been held to a standard. Confidence checking is that standard's final clause, and it is the one most easily skipped precisely because it usually confirms what the operator already believed.

Connected inside MARS

R&D Boundary →

Advanced EV Analytics Lab →

Monte Carlo Lab →

Every brief documents the same shipped system.

The complete MARS package — eleven workbooks, three TradingView indicators, the full manual library — $497.

Get MARS Pro — $497 Explore the Platform