The eighteen-month roadmap replaces priors with owned evidence

The illustrative months nine to eighteen window turns an accepted candidate into an operating capability. Scale here means reliable use in defined planning decisions, not a universal roll-out across every product and channel. A model that won on consumer sell-through does not automatically graduate to institutional orders. A calibrated style total does not establish calibrated size-level demand. Each extension receives its own evidence pack and may remain outside the release boundary.

Exhibit 65 uses pilot launch as the common origin. The illustrative preparation allowance precedes that origin. Twelve weeks are approximated as three months for drawing; the subsequent phases cover months four to eight and nine to eighteen after launch. Readiness testing belongs inside controlled comparison. Gate failures can extend any phase without creating permission to release.

Later operation adds structure only where the decision needs it. Hierarchical reconciliation becomes relevant when independent size, style, channel and regional forecasts contradict one another. MinT reconciliation seeks coherent forecasts across a collection of series under an unbiasedness assumption, so its use requires the assumption and resulting coherence to be inspected rather than treated as a formatting step.[S1405] The hierarchy follows the planning structure. It does not conceal weak variant evidence inside an apparently sound total.

Drift response becomes a standing control. Concept-drift research concerns changes in the relationship between inputs and targets and reviews adaptation and evaluation strategies.[S38] For this service, drift includes an assortment shift, a new channel mix, a changed price policy and a break in availability capture. Monitoring therefore watches input meaning and population as well as forecast errors. A response can be renewed evaluation, a wider range, selective retraining or reversion to the approved baseline.

Retraining is selective because a new data cut is not automatically a better training set. The release package identifies the training cut-off, feature snapshot, model version, calibration window and intended decision horizon. Operational guidance calls for summary statistics, realised performance monitoring, notifications and rollback when values deviate from expectations.[S40] The venture tests the replacement on the same launch-replay logic before release. If the new candidate fails, the prior approved version or baseline remains available.

Retirement is equally important. Weak models, unused features and duplicated reports leave the service when their evidence no longer justifies their burden. The register records why each component was retained or removed and whether any downstream decision still depends on it. This keeps the final system proportionate to the planning problem. The illustrative Phase 3 succeeds when the organisation can reproduce, challenge and reverse every retained release without relying on an undocumented individual or supplier console.

A gate is a decision meeting supported by a reproducible evidence pack. It is not a demonstration of the preferred model. The pack contains the forecast contract, prior register and sensitivity analysis; the pilot charter and observed conditions; the frozen corpus and launch-replay design; candidate and baseline results; calibration and residual uncertainty; operating readiness; and the proposed reversal conditions. Reviewers can then distinguish what was observed, what was derived and what remains an assumption.

Prior sensitivity belongs at every gate, not only before launch. Post-hoc calibration procedures can support calibrated uncertainty when enough suitable data are available, which makes the sufficient-data qualification part of the review.[S15] If the interval changes materially under a defensible alternative prior, analogue pool or calibration window, the sponsor should see that variation. A single selected specification would hide an important source of doubt.

The pilot section of the pack preserves the difference between designed and observed conditions. It records assignment, actual exposure, availability, returns, deviations and any human interventions. Controlled design supports causal interpretation only within its valid assignment and analysis conditions.[S32] Where those conditions failed, the gate can still approve learning or another pilot, but it should not approve a causal claim or a production scale-up that the design cannot support.

The backtest section shows who and what was held out. Independent launch counts appear beside row counts. As-of feature snapshots, preprocessing fits and calibration windows follow the forecast origin. Results are segmented by horizon, product family, channel and availability state. The sponsor should be able to see whether a candidate won on the deployment population or only on a convenient aggregate. Any excluded launch carries a reason fixed before the final comparison.

Operational readiness receives the same scrutiny as statistical performance. The pack identifies the service owner, data owner, planning owner and release approver. It includes freshness checks, runtime, failure behaviour, monitoring, fallback and a tested restoration path. It also states which supplier components are used, what data leave the venture's control and how the record can be exported. These are release conditions rather than procurement details to be resolved after planners become dependent on the output.

Each gate records an explicit decision. Proceed authorises the next bounded phase. Hold retains the current prior, baseline or service while a named evidence gap is repaired. Narrow limits the release to the products, channels or horizons that passed. Reverse restores the last approved state and records the trigger. The decision log names the owner, evidence version, conditions and review point. Silence or elapsed time never counts as approval.

Sponsors also receive the unresolved uncertainties in decision language. The pack states which commitments would change under the low, central and high cases, which assumptions drive the difference and what evidence could resolve it. It does not convert a predictive interval into a promise of sales, profit or investment return. This chapter offers a planning and measurement method, not investment, legal or tax advice. A specific venture commitment requires its own commercial, financial, legal and operational review.

The durable research asset is a relevant, well-governed launch corpus. Fashion forecasting research describes uncertainty, missing history and seasonality as recurring conditions, but that general framing does not supply effect sizes for a particular Cameroon apparel venture.[S1401] Visual-similarity research likewise offers a candidate transfer mechanism based on corrected demand from similar historical products, not proof that a new local manufacturer will inherit the reported performance.[S05] Local evidence must test both the mechanism and its limits.

Future work should assemble anonymised launch distributions with clear product, price, channel and timing descriptors. Stockout annotations and listing intervals should accompany sales so that later researchers can distinguish observed purchases from latent-demand corrections. Attribute taxonomies should be stable enough for comparison while retaining version history when the offer changes. Channel-specific validation should separate consumer sell-through, distributor orders and institutional commitments rather than pooling them into one demand label.

The corpus should preserve disappointing and constrained launches. Exclusion rules may remove unusable records from an analysis, but the register retains their existence and exclusion reason. Otherwise published evidence will overstate launch predictability.

Sharing requires rights and confidentiality to be designed before aggregation. An anonymised research product should remove direct identifiers, assess re-identification risk from rare orders or small channels and document the transformation applied to every released field. Commercially sensitive price, customer and supplier details need not be published to make the method inspectable. Aggregated distributions, data dictionaries and validation designs can contribute evidence while keeping the underlying demand record under its owner's control.

Policy context must remain outside the demand label. The inherited sector register records about 95 percent of Cameroon's cotton fibre exported raw and about 5 percent processed locally, with the original publication date not recorded; it separately records a national target of 50 percent local processing by 2030.[R06][R07] These figures explain an industrial ambition, not the sales distribution of a garment, a channel or a new firm. They are inherited sector observations and targets, not current GFT operating results.

The research question is whether local-processing ambition can be matched by repeatable demand learning. Evidence must come from actual offers, availability and buying behaviour. Publish where models fail, where stock censoring blocks inference and where simple baselines remain sufficient.

For GFT's Cameroon textile programme, the recommended close is therefore practical. Build the contract and prior, run the reversible pilot, protect the launch record, compare candidates on whole launches and operate only what proves useful. Publish or share evidence only where rights and confidentiality permit. Keep the national ambition visible as context while letting owned observations decide the next forecast revision.