What a promoted change does to the evidence base it re-enters.

The overlooked half

Promotion is a beginning, not a conclusion.

Most discussion of the feedback principle concerns the barrier — what a change must survive before it goes live. The barrier matters, but it obscures what happens next. The moment a change is promoted, every subsequent trade is generated under the new configuration, which means the evidence base immediately begins accumulating two kinds of rows: those produced under the old rules and those produced under the new. Every aggregate computed across that boundary is a blend of two systems, and nothing in the numbers marks where the boundary sits unless the change note does.

The batching argument

Changes queue so that each review window measures a known configuration.

This is the real function of the rule that a validated change enters production at the next scheduled review rather than mid-cycle. The queue does more than prevent hot-patching — it batches changes so each window measures one configuration from start to finish. A system that promoted changes continuously would produce a record in which no period was ever generated under a stable ruleset, and therefore no period could be attributed to one. The queue costs a few weeks of delayed benefit and buys the ability to say what any given month was actually testing.

The interpretability cost

Continuous improvement produces a permanently unmeasurable system.

There is a genuine paradox worth naming here. A system that adopts every validated improvement as soon as it is found will, at any moment, be running the best configuration its research has identified — and will never be able to demonstrate that any of those configurations worked. Each change is superseded before it accumulates a sample. The operator ends up with a system that is always improving in theory and never measurable in practice, which is indistinguishable from a system that is drifting. The batching rule accepts a slightly worse configuration in exchange for a knowable one.

What resets and what doesn't

A change's blast radius determines how much history it invalidates.

Not every promotion resets everything, and knowing the scope is what makes the post-promotion period readable. A coefficient ladder change affects trade management and therefore execution-quality metrics, while leaving branch classification and gate behavior untouched — so MAE/MFE history becomes non-comparable across the boundary and branch counts remain continuous. A branch weight change affects blended EV and quota floors while leaving individual branch EVs comparable. This is why the change note records scope: it is the document that tells a future reviewer which historical series survived the change and which restart.

• Execution-metric changes invalidate capture and giveback history; branch-level series continue.
• Weight changes reset blended EV comparability but not per-branch comparability.
• Classification or labelling changes are the most expensive — they reset nearly everything.

The loop's integrity condition

The feedback principle only works if the evidence it feeds on stays trustworthy.

The closing half depends entirely on the opening half of the whole system. A promoted change is validated against evidence, produces new evidence, and that new evidence informs the next change — which means the loop amplifies whatever quality the evidence base has. A journal entered honestly makes each revolution slightly better informed than the last. A journal with quiet omissions and generous adherence grades makes each revolution slightly more confident about something that is not true, with no mechanism anywhere in the loop capable of detecting it. The feedback principle is a multiplier, and multipliers work on whatever sign they are given.

Where most traders begin

The loop ends where the instinct starts, and that inversion is the design.

The page states it plainly: the loop ends where most traders begin, which is changing rules. The instinctive response to a losing stretch is to adjust the system, and the instinct arrives with genuine reasoning attached — something did go wrong, and something probably should change. MARS does not dispute either claim. It disputes the timing and the evidence, routing the impulse through diagnosis, sandbox validation, a review window, and a written note. The change may well be the same one the instinct proposed. What differs is that by the time it is made, it is possible to know whether it worked.