Optimisation earns authority through reproducible decisions and safe fallback

GFT would begin with an item-location policy that a planner can calculate independently. It should identify the stock object, the customer promise, the review event, the replenishment rule and the owner of exceptions. A policy also needs its effective date, expiry condition and permitted material substitutions. An engine that cannot state these fields has no stable instruction to execute, even if its forecast is statistically elaborate.

The foundational teaching material distinguishes order-point policies from periodic review and defines inventory position separately from physical stock. Those distinctions support an executable baseline: a trigger, an order calculation and a review calendar can be specified without a learning algorithm. Their suitability still depends on the demand and replenishment process.[S05] GFT proposes keeping this baseline available as a comparator and as a documented manual procedure, subject to current input checks.

The forecast interface should accept an identified forecast vintage, item-location scope, horizon, demand unit, time basis and uncertainty representation. It should also describe whether the demand represents customer requests, observed sales or component consumption. GFT would reject an interface that sends a bare number with no definition. Forecast estimation belongs to the forecasting paper; this chapter owns the contract by which an estimate becomes a candidate replenishment action.

Calendars require equal care. GFT would distinguish the factory consumption calendar, supplier dispatch calendar and delivery acceptance calendar. A supplier closure should change the feasible receipt date before a recommendation is approved. The planner should be able to inspect the calendar version used in the calculation and see any assumption about working days. A recommendation recalculated after a calendar change should supersede its predecessor rather than appear as an additional requirement.

The exception queue is part of the baseline. GFT proposes grouping exceptions by the decision they block: uncertain availability, incompatible specification, unaffordable quantity, late receipt or missing approval. Each entry should identify an accountable role, the affected promise and the evidence needed to clear it. The North Carolina cycle-counting procedure separates count entry from approval and explicitly addresses discrepancies; it is a procedural reference, not proof of apparel-factory performance.[S27]

The proposed chain in Exhibit 49 keeps data capture, calculation and commitment distinct. Its layers are design responsibilities, not claims about deployed software. A small factory might initially perform several responsibilities in a controlled workbook. The separation remains useful because the planner must know which action changes a record and which action changes a commercial obligation.

GFT would preserve a specimen decision packet that another authorised planner can reproduce from the saved extract. Reproduction means the same accepted inputs produce the same policy calculation, rounding and constraint results. It does not mean that the next shipment will arrive as predicted. The packet should distinguish reproducibility of a calculation from reliability of the commercial assumptions on which it depends.

A baseline also needs a declared response when no replenishment is proposed. GFT would distinguish a valid decision to wait from an absent calculation, an expired policy and a blocked input. Each state should have a reason visible to the planner. Silence should never be interpreted as confirmation that the customer promise is protected. The review record should show the next event that can change the decision, such as a usable receipt, a revised requirement or an approved parameter update.

A solver needs an economic question rather than a general instruction to improve inventory. GFT proposes quantities by material and release date as decision variables, with explicit choices for approved substitutions where permitted. The objective should identify relevant purchase, ordering, holding, shortage and expediting costs over a stated horizon. The treatment of remaining stock must be visible. A terminal assumption can otherwise make buying beyond the selling window appear attractive without revealing the residual exposure.

The MIT material on lot-sizing extensions provides a basis for examining purchase discounts alongside inventory costs. It does not establish local supplier prices or factory constraints.[S03] GFT would therefore bind the objective to dated commercial inputs and retain the input currency and cost boundary. A discount schedule from a supplier quotation should be distinguished from an assumed teaching schedule. Cost coefficients should never imply precision that the underlying shortage or recovery evidence cannot support.

GFT proposes separate constraints for spending dates, order multiples, minimum lots, receiving space, available conversion capacity and material compatibility. Service commitments need their own mathematical definition. A promised garment quantity cannot be supported by an aggregate fabric total if required components are unavailable. MIT's material-planning lecture explicitly raises capacity and uncertainty as issues beyond an initial requirements calculation; a requirements schedule alone does not prove feasibility.[S07]

For implementation, GFT would write the inventory balance as opening usable supply plus accepted receipts, less consumption and other authorised outflows, equals closing usable stock. The arithmetic must use the same material state throughout. A receipt under quality hold should enter the physical ledger without automatically entering usable supply. The finance constraint should evaluate payment obligations on their actual assumed dates. A plan passing an aggregate budget check may still fail at a deposit date.

The engine should return a status alongside its proposed quantities. GFT's proposed statuses in Exhibit 50 separate independently checked feasibility from a claim of optimality. A time limit may leave an admissible candidate, an inadmissible candidate or no candidate. The planner needs to know which occurred. The status should preserve termination reason, objective value, any available bound, the reported gap definition and the result of independent constraint checks.

A reported optimum remains conditional on the submitted formulation. GFT would independently recalculate pack divisibility, cash timing, material balances and capacity consumption from the proposed quantities. Integer rounding after optimisation should trigger those checks again. If costs can take values close to zero, the report should explain how a relative gap is defined rather than presenting an unstable percentage as an intuitive measure of economic importance.

Infeasibility is a business finding when the formulation is correct. GFT would present the conflicting obligations and controlled alternatives: change the delivery promise, split the order, qualify another source or obtain additional approved funding. A diagnostic model may allow labelled service shortfalls to identify a trade-off. Its output must remain a diagnostic until the authorised owner accepts the changed promise. A penalty coefficient should not quietly purchase permission to breach a hard commitment.

GFT proposes admission checks before any calculation can produce an approvable purchase recommendation. The checks should distinguish missing, invalid, stale and disputed inputs. A missing lead time is unknown. A negative lead time is invalid. A quotation beyond its accepted review window is stale. A receipt that conflicts with the physical count is disputed. These states need different remedies, even when they all block order release.

The item definition should carry the purchase unit, consumption unit and governed conversion. GFT would forbid conversion between rolls and metres without an accepted roll-length basis, and between mass and length without the relevant specification. Apparent numerical compatibility is insufficient. Material revision, shade restrictions and approved substitution should be checked before netting supply against demand. The engine must not manufacture interchangeability by collapsing different stock identifiers.

For time inputs, GFT would require event definitions and calendars. Promised dispatch, physical arrival and usable acceptance are different events. An imported timestamp without a time-zone interpretation should not silently determine which review window contains the receipt. Future-dated movements and out-of-sequence releases should enter an exception queue. The data owner can then explain a recording delay or correct a genuine error while preserving the original event.

Count discrepancies require controlled resolution. The opened North Carolina procedure explicitly covers entering results and approving counts.[S27] GFT would use that separation to design a local discrepancy workflow: record the observation, investigate relevant movements, approve an adjustment and recalculate affected recommendations. The recommendation engine should not overwrite the ledger to make its material balance fit. Model feasibility cannot adjudicate a disputed physical quantity.

Shipment reconciliation is similarly distinct from successful data transfer. Project Zipper compared barcode and RFID information across shipping and receiving stages; its study-specific findings should not become a guaranteed accuracy target for a different factory.[S31] GFT proposes reconciling expected, dispatched, physically received and accepted quantities as separate states. Duplicate messages should not create additional stock, and a successfully uploaded shipping notice should not become a usable receipt.

The failure response should remain proportionate to the affected boundary. GFT would block recommendations using an unresolved item-location record and flag any dependent garment commitments. Unrelated records may continue only if their independence has been established. A shared unit-conversion error may invalidate a wider scope than a local count dispute. The interface should explain the blocked scope and the rule that caused it, making recovery an accountable decision.

GFT would also apply numerical admission checks: finite values, valid probability ranges, consistent horizons and adequate precision for the declared units. Zero variability should be accepted only when it is a deliberate input supported by the chosen method, rather than a replacement for a missing estimate. Extremely large proposed quantities should be compared with approved bounds before display as a recommendation. Validation should reveal uncertainty rather than hide it behind defaults.

Finally, freshness should follow the decision's consequence. GFT would set review rules by input class and record the last accepted observation. A stock extract can become stale after an important movement even if little time has elapsed. A stable material specification may remain usable longer than a delivery promise under disruption. The system should evaluate these explicit rules, retain their versions and state which changed input would make recalculation necessary.