A policy improves the factory only when measured service and cash improve together

GFT recommends beginning evaluation with a precisely specified question: would the candidate rule have made a better feasible commitment using the information then available? Specify the material family, location, customer promise and authority being tested. A trial of planner recommendations should not quietly become a trial of supplier selection or customer pricing. Record the incumbent rule, including its exceptions, before interpreting the candidate's result. Otherwise the comparison risks testing a documented new process against an imaginary version of the old process.

Research on lost-sales inventory distinguishes observable transactions from underlying demand and supply. Research on simulation-based inventory optimisation compares policies within particular modelling assumptions. Neither source establishes that a historical factory ledger contains every input needed for a counterfactual decision. GFT therefore recommends a replay admission check before scoring policies: the evaluator should demonstrate the availability, meaning and timing of the required records.[S22][S23]

For each review event, preserve the forecast vintage, usable stock by status, reservations, outstanding purchase commitments, released receipts, current cost basis, cash availability and applicable constraints. Keep both physical event time and recording time. A receipt physically delivered but not recorded until later was not necessarily visible to the planner. The replay should reproduce the agreed information boundary rather than granting the candidate a cleaner view of the past. Corrections discovered afterwards belong in a separate reconciliation layer.

Supplier observations need the same discipline. GFT proposes estimating a decision-date lead-time distribution only from information available by that date. Open orders remain open observations with elapsed time and promised dates; they are not completed deliveries with invented arrival dates. Retain partial receipts and rejected quantities. The censored-data problem identified in inventory research is a reason to expose these gaps, rather than silently fill them with convenient averages.[S22]

Separate tuning from evaluation. GFT recommends retaining an earlier calibration window, a subsequent validation window and a final untouched evaluation window wherever the available history permits that division. Choose boundaries around the actual ordering and selling cycle. Freeze parameter-selection criteria before reading final outcomes. Report when a short or interrupted history prevents an honest holdout. A replay with repeatedly adjusted parameters should be labelled exploratory, even when its final score looks favourable.

A recorded receipt path also has limits. A different order quantity, order date or supplier may have changed dispatch priority, partial fulfilment or payment requirements. Holding historical receipts fixed is therefore a conditional experiment, not a reconstruction of an alternative reality. GFT recommends reporting a fixed-receipt replay separately from scenarios that model supplier response. Where that response cannot be estimated, retain an explicit assumption and a sensitivity range rather than calling the alternative result observed.

The evidence ladder in Exhibit 54 is a proposed sequence of questions, not a numerical ranking of credibility. Equation checks establish implementation consistency; later stages expose additional failure modes. None automatically substitutes for the next.

GFT would retain a replay manifest containing input extracts, clock convention, parameter version, exclusion reasons and the exact decisions scored. A reviewer should be able to reconstruct an unfavourable recommendation as readily as a favourable one. If stock states or purchase approvals cannot be reconstructed, the appropriate finding is limited evaluability. That finding supports improved capture before a stronger claim; it does not establish that the candidate policy has failed operationally.

Inventory research explicitly identifies censoring in both demand and supply observations. When a customer cannot obtain an unavailable item, a sales record may contain no transaction for the missed request. Equally, a short delivery need not reveal the supplier's full available capacity. These are information limits, not permission to invent the missing quantities. The cited research manuscript develops a specialised learning model; its theoretical guarantees are not evidence of local factory service.[S22]

GFT proposes a request ledger that records the requested item and specification, quantity, request time, confirmed due date, offer, acceptance, substitution, cancellation and final fulfilment. Link repeated enquiries to the same commercial request where possible. A refused quotation, a cancelled confirmed order and a late shipment should remain distinguishable. They answer different questions about demand, the promise and execution. Preserve the original confirmed date when a later date is negotiated.

Do not make unavailable items disappear from the denominator. For a proposed immediate unit-fill measure, count eligible requested units and identify those supplied immediately under the agreed definition. For order delivery, retain the complete confirmed order and due-date boundary. GFT recommends disclosing cancellations and customer-approved substitutions alongside the principal result. A policy should not receive an apparent service improvement merely because difficult requests were removed or a different material was shipped without equivalent approval.

Use bounds when request capture is incomplete. In an illustrative stockout interval, assume recorded sales of 80 garments, documented unfilled requests for 10 distinct additional garments, and an explicitly assumed further unrecorded-demand range of zero to 20 garments. Eligible requests would lie between 90 and 110 garments. With all recorded sales counted as immediately supplied, the illustrative fill range is 80 divided by 110 to 80 divided by 90, approximately 72.7 to 88.9 percent. The range expresses an assumed missing-demand bound, not a confidence interval.

That illustration deliberately avoids treating stockout duration as a demand estimate. GFT recommends using availability timestamps to identify exposure and seeking corroboration from enquiries, backorders or comparable in-stock periods. Comparable periods still require a judgement about season, price, customer mix and selling opportunity. If the ordering decision changes customer willingness to request, historical demand is itself an imperfect comparator. Report that dependence rather than presenting a single adjusted sales series as measured truth.

A useful sensitivity test asks whether the policy ranking survives plausible missing-demand bounds. GFT would calculate the service and cost differences under the same declared assumptions for both policies. If the ranking reverses, the decision is sensitive to information that has not been observed. Prioritise better request capture or a bounded field comparison. Do not choose the assumption that makes the preferred candidate win and then describe the result as a backtested improvement.

Supply uncertainty requires an equally visible register. Distinguish ordered, confirmed, dispatched, received and quality-released quantities. A request for more material may expose a supplier capacity limit that the historical smaller orders never tested. GFT would therefore reject a replay that assumes the alternative order could always have been fulfilled merely because its simulated warehouse had room. The proposed record should identify whether evidence, a contractual commitment or an illustrative response model supports each alternative receipt.

Stochastic lead-time research examines how replenishment variation interacts with ordering behaviour. Its assumptions and application context do not calibrate Cameroon transport or supplier performance. GFT proposes using that work as a warning against replacing an uncertain replenishment clock with a deterministic average, while testing local dependencies separately.[S24]

Construct a scenario register around mechanisms. A demand surge can coincide with supplier congestion; a partial fabric receipt can coexist with missing labels; a quality hold can delay material that the ledger already shows as received. A cash freeze may prevent replenishment just as overdue customer receipts increase. GFT recommends recording the combination, affected decision, observable warning, fallback and recovery condition. Avoid attributing a probability to a scenario unless an appropriate local basis has been established.

The heatmap in Exhibit 56 identifies which synthetic shocks are switched on, rather than assigning measured risk scores. In this illustrative register, an active demand shock multiplies requests by 1.3, the delay shock adds five days, the quality shock releases only 75 percent of a planned receipt in time, and the funding shock funds only 60 percent of planned purchases. Binary cells show scenario composition, not equal severity or joint event frequency.

A useful stress run should explain the first binding constraint. GFT would identify whether cash blocks the order, a missing component blocks production release, quality approval blocks usable stock or a transport delay blocks receipt. Record the time until the breach becomes visible and the action actually available then. A warning issued after the last feasible purchase date is not an effective protection, even if the dashboard eventually diagnoses the shortage correctly.

Test supplier correlation through physical dependencies rather than supplier names alone. OECD describes subcontracting and multi-tiered sourcing in the garment sector. GFT therefore recommends checking whether apparently distinct suppliers rely on the same fabric processor, carrier or constrained route. A scenario may impose a common outage without asserting its probability. Two alternative purchase addresses do not establish independent recovery capacity.[S25]

Connectivity loss needs an operational test as well as a data test. GFT proposes retaining a frozen authorised manual plan during the outage, recording local movements and checking reconciliation before reopening recommendations. The evaluation should count delayed decisions and manual effort. An automatic retry that duplicates a recommendation would be a control failure even if material service happened to remain acceptable. Detailed integration design belongs to the decision-system chapter; this chapter asks whether the fallback preserved the measured boundary.

When no feasible plan meets the customer promise, report that infeasibility. GFT recommends an exception decision on allocation, authorised substitution, negotiated delivery or reduced commitment, with its service consequences retained. Stress testing should not assume unlimited emergency freight, supplier credit or overtime as free recovery. Record the resources and approvals required to recover. The purpose is to expose where the policy stops being acceptable and whether a controllable response exists.