# Specification contract for a Marketing Mix Model

## 1. Decision

- Budget in scope: €3.2M per quarter
- Levers: video, search, social, CRM, retail media
- Decision: reallocate at most 20% per quarter
- Horizon: net contribution at 13 weeks
- Review: quarterly

## 2. Estimand

Average, then marginal, incremental effect of each lever on weekly net contribution, conditional on prices, promotions, stock-outs, distribution, seasonality and documented events.

## 3. Domain of use

- Observed support per channel, published with each curve.
- No recommendation more than 15% outside the observed minimum or maximum.
- No reading by creative or platform when the variation does not allow identification.

## 4. Data and grain

- Grain: week × market.
- Minimum: 156 weeks, a versioned data dictionary and a log of breaks.
- Outcome: net sales minus variable costs, returns and channel costs.
- Controls: net price, promotion, availability, distribution, calendar and competition.

## 5. Validation

- Out-of-sample temporal validation.
- Checks on signs, contributions, residuals, stability and marginal slope.
- Comparison of several windows, adstocks, curves and priors.
- Test of a retrospective allocation against the status quo.

## 6. Calibration

Compatible experiments are integrated as external information, with their population, treatment, horizon and uncertainty. Any divergence is documented before the update.

## 7. Authorised decisions

- Compare scenarios within the support.
- Identify a learning tranche.
- Estimate a range of contribution and regret.

## 8. Prohibited decisions

- Presenting a contribution as causal certainty without calibration.
- Extrapolating a curve far beyond the support.
- Optimising automatically without constraints or human validation.

## 9. Acceptance gate

Acceptance only if identifiability, decision validation and calibration are sufficient for the exposure. Otherwise: a bounded pilot, targeted data collection or keeping the current plan.