Media-Mix Model and Incrementality Honesty Gate Automation: A Coefficient Can't Speak Without Its Fit

A media-mix regression always produces coefficients — that is what regression does, regardless of whether the data behind it actually supports a conclusion. The dangerous version of this is a model trained on one channel, or on a handful of noisy weeks, that still confidently reports "Instagram drives 100% of results," because with one input variable a regression has no choice but to attribute the entire outcome to it. Nothing about the math itself signals that this is an artifact of the input, not a discovery about the business — that has to be checked separately, by a rule that knows what a trustworthy model looks like.

A coefficient by itself is an answer with no context attached — it doesn't say how much data trained it, how well the model actually fits reality, or whether that fit even had enough channels present to attribute anything meaningfully. A single-channel model is the sharpest version of this: mathematically, a model with one input variable has to attribute 100% of whatever variation it explains to that one variable, so "100% contribution" from such a model is not a signal about the business — it is a property of having asked a one-variable question. Treating that number as a budget recommendation converts an artifact of the setup into decision-grade evidence it never earned.

Every media-mix model trained on an organization's data is evaluated against four fixed thresholds before it is allowed to carry a "usable" label: r² of at least 0.5 (the model actually explains a meaningful share of the outcome's variation), a MAPE of 30% or less (predictions are close enough to be worth acting on), at least 60 observations (enough data points to have trained on), and at least 30 days with an actual measured outcome (enough of the window had something to explain in the first place). A model failing any one of these does not get partially trusted — `usable` is a single boolean, and the function that computes it collects every failed threshold into a plain-language reasons list rather than stopping at the first failure, specifically so a report can say ALL the reasons a model isn't ready, not just one.

On top of the four thresholds sits a fifth, structural check aimed specifically at the single-channel trap: a model trained on exactly one channel showing that channel at or near 100% contribution is flagged with its own distinct message — "single-channel model: a 100% contribution here is arithmetic, not discovery" — a wording chosen specifically to stop that number from being read as a finding. The file's own governing comment cites the real, current example this rule is built to catch: as of this reading, the one MMM model actually trained in this database explains 8% of variation (r²=0.08), errs by an average of 102.9% (MAPE), and is recorded with `supports_budget_simulation: false` — exactly the kind of model this gate is designed to intercept before its number reaches a budget conversation.

When a model fails the gate, the interface shows a "does not decide budget" badge instead of "utilizável," with the specific list of reasons rendered in plain language next to it — phrased deliberately to be readable out loud in a meeting, not as an error code. The coefficients themselves are not hidden; they remain visible for inspection, because a rejected model can still be useful context for someone who understands its limitations — it is simply barred from being cited as if it had passed.

Incrementality experiments get the same discipline applied to statistical significance: at a fixed α of 0.05, a result is significant only if its p-value is below that threshold. When a p-value hasn't been computed yet — because an experiment hasn't run long enough to produce one — the significance field renders as `null`, not `false`, and the code's own comment states the reasoning directly: a missing p-value means "not yet measured," which is a different claim from "no effect was found," and collapsing the two into a single "not significant" result would misrepresent an experiment that simply isn't finished as one that already failed.

A team wants to know how much of their revenue comes from paid social, and trains a media-mix model using only that one channel's spend as an input. The regression, correctly following the math, attributes essentially all of the modeled variation to that single channel — a 100% figure that would look decisive on a slide. The single-channel check catches this specifically, labeling the result an arithmetic artifact of a one-variable model rather than evidence that paid social drives all revenue, and the badge reads "does not decide budget" rather than presenting the number as a real finding.

A marketing lead can no longer accidentally cite a barely-fitted or single-channel model's coefficient as if it were a validated insight — the gate names the specific reason it isn't ready, out loud, before the conversation gets that far. And an incrementality test that simply hasn't finished collecting data reads as exactly that — "not yet measured" — rather than being quietly reported as proof that a channel had no effect, which protects a genuinely promising channel from being cut based on an unfinished experiment.

A team wanting a model to always produce a confident, actionable contribution number will find the opposite here by design — a model that doesn't clear the four thresholds is explicitly barred from making a budget recommendation, whatever coefficient it produced. A team wanting the significance threshold configurable per report will find one fixed α value applied everywhere instead, specifically so the same underlying result can't be called significant in one place and not in another.

Publishing every media-mix coefficient a regression produces, with no check on how well the model actually fits or how much data trained it, treats "the math ran" as equivalent to "the finding is real" — a gap a single-channel model exposes immediately, since it mathematically has no choice but to attribute everything to its one input. Centriu Dash instead requires a model to clear four independent, named thresholds before its number is allowed to inform a budget decision, names the single-channel arithmetic trap explicitly rather than let it pass as insight, and keeps a missing p-value from ever being read as proof of nothing happening.