Diminishing Returns Detection¶

A measurement method — instantiates Plateau Detection and Switching

Detects a plateau by comparing each added unit of input against the extra output it buys, flagging the point where marginal response has flattened even while total output still looks healthy.

Diminishing Returns Detection is the sensor that fires the whole archetype. It works on the derivative, not the level: it compares each additional unit of input — a dollar, an hour, a rep, a headcount — against the extra output that unit produced, and reports when that marginal gain has fallen flat. The distinguishing move is that it deliberately ignores total output. A campaign, a training block, or a production line can still be posting high and even rising totals while the return on the next unit has already collapsed to nearly nothing; a level-watching eye misses that, and a margin-watching one catches it. Everything downstream — the stop rule, the switch — needs this signal to have something to act on.

Example¶

An agronomist is dialing in nitrogen fertilizer for a wheat field. The first 40 kg/ha lifts yield sharply; the next 40 adds noticeably less; by the fourth increment an extra 40 kg buys only a sliver of grain and starts pushing nitrate into the groundwater. Diminishing Returns Detection is the step that names this precisely: it scopes the pathway (nitrogen rate on this soil and crop), fixes the response (marketable yield), and tracks marginal yield per added kilogram. Plotted, the curve is the textbook law of diminishing returns — rising but bending over. The detector flags the increment where marginal yield per kilogram drops below the value of the nitrogen itself, roughly ≈160 kg/ha here. Above that point, "add more fertilizer" is no longer the answer, and the field is a candidate for a different lever — variety, drainage, timing — rather than a heavier dose.

How it works¶

The method turns a stream of input–output pairs into a marginal curve and reads where it bends:

Scope one pathway. Fix exactly which input is being varied and which response should move, so the margin is attributable and not smeared across confounds.
Difference, don't total. Compute output change per input change between successive increments, not cumulative output — the flattening lives in the first difference.
Compare margin to a floor. Flag the increment where marginal response falls below a preset floor (often the cost of the unit itself, so the test is "does the next unit still pay for itself?").

Tuning parameters¶

Increment granularity — how finely input is stepped. Fine steps catch the bend early but amplify noise; coarse steps are stable but can overshoot the plateau before it's seen.
Smoothing window — how many increments are averaged before reading the margin. More smoothing rejects noise but lags the true turn; less smoothing is jumpy.
Floor definition — whether the plateau line is an absolute marginal gain, a percentage of the first unit's gain, or the unit's own cost. Cost-based floors make the detector answer "still worth it?" directly.
Raw vs. cost-adjusted margin — reading marginal output flags saturation; reading marginal output per dollar flags unprofitability, which turns sooner.

When it helps, and when it misleads¶

Its strength is separating "we are doing well" from "the next unit is doing well" — the confusion that keeps money and effort flowing into a spent path long after the totals stopped it being justified.^[1]

Its failure mode is calling a plateau on a margin that is merely noisy or lagged — a run of small increments that regresses back up, or a response that simply hasn't arrived yet. Because it reads only the instantaneous margin, it has no view of whether the pathway was given a fair trial; that check belongs to a sibling. The classic misuse is running it backwards: choosing the increment window that makes the margin look flat in order to justify quitting a path someone already wanted to drop. The discipline is to keep detection separate from the stop-and-switch decision, and to hand the "is this trial long enough / is this real?" question to the mechanisms built for it.

How it implements the components¶

Diminishing Returns Detection fills the pure measurement subset — the signal, not the response to it:

input_pathway_scope — fixes the specific input/pathway whose margin is being read, so a plateau is diagnosed against a defined path rather than as vague disappointment.
response_metric — pins the outcome that should move if the path is still productive.
marginal_response_metric — the core output: extra response per extra unit of input, differenced increment by increment.
plateau_threshold — the floor the margin must fall below to be called flat.

It does not judge whether the trial was long or valid enough (observation_window, false_plateau_check — see Controlled Experiment After Plateau and Plateau Review Cadence), nor does it stop escalation or pick a new path — those belong to Escalation Stop Workflow and Strategy Switch Decision Tree.

Instantiates: Plateau Detection and Switching — it supplies the marginal-flattening signal the whole archetype is triggered by.
Sibling mechanisms: Marginal Gain Dashboard · Controlled Experiment After Plateau · Escalation Stop Workflow · Plateau Review Cadence · Ad Fatigue Switching · Training Plateau Adjustment · Product Growth Plateau Response · Process Redesign After Plateau · Saturation-Aware Resource Allocation · Strategy Switch Decision Tree · Expert-Governed Modality Change

References¶

[1] The law of diminishing returns — in production economics, holding other inputs fixed, each added unit of a variable input eventually yields a smaller increment of output. It is the formal object this detector measures; the flattening of the marginal curve is exactly the plateau signal. ↩