A tech pack is complete only when the receiving team can use it

Structured generation is useful because it can turn an approved brief and technician's notes into a first draft with consistent field names. ChatGarment's research interface generates JSON containing textual garment attributes and continuous numerical parameters, then uses that file to drive a parametric sewing-pattern model.[S10] That establishes a route from multimodal input to structured pattern parameters. It does not establish a graded bill of materials, verified supplier identifiers, buyer tolerances, operation sequences or production approval. The output boundary of the research system must not be relabelled as a complete factory tech pack.

The safe drafting pattern is retrieval before generation. The system should first receive the controlled style record, approved block, applicable point-of-measure schema, material and trim identifiers, construction library and current revision. It may then propose text for fields whose source records exist. A construction sentence should cite the operation or seam identifier it paraphrases. A measurement entry should preserve the source unit and landmark rather than infer a number from a garment description. A material description should resolve to the selected material record, not to a generic phrase such as cotton fabric. NIST's generative-AI profile identifies confabulation and information-integrity risks, which makes source-linked verification especially important when generated content enters consequential work.[S20]

Unknown inputs need a different behaviour from missing prose. If fibre content is absent, the correct output is awaiting evidence: material record, with procurement or the material owner assigned to resolve it. If a tolerance has not been agreed, the system should not select a familiar value from another product. If a trim code has no approved supplier reference, a plausible alphanumeric code is more dangerous than a blank because it can be mistaken for a buying instruction. These are deliberate fail-closed rules. They preserve the distinction between an absent value, an inapplicable field and a rejected value.

Extraction requires the same discipline as composition. A model asked to read an incoming document should return the exact source location, the extracted text or value, a confidence or review flag where available, and any conflict with the controlled record. The reviewer should see the source passage beside the proposed field. Suggested actions in the NIST profile include empirical evaluation and verification of generated information, but they do not certify a particular extraction workflow.[S20] The local acceptance test must therefore include wrong units, conflicting revisions, ambiguous abbreviations and absent evidence, not only clean documents.

A practical field can carry four separate states: proposed, source verified, technically approved and superseded. The states should not collapse into a single coloured tick. Source verification establishes that the value matches its origin. Technical approval establishes that the accountable role accepts it for the intended use. Supersession records that it once applied but no longer does. This structure lets a model accelerate clerical drafting while leaving technical judgement and release authority visible. It also creates a usable audit trail when a later sample fails and the team needs to know which information was actually approved.

A filename ending in DXF does not specify enough information for a controlled handoff. CLO's documentation lists Standard DXF, AAMA and ASTM choices, together with controls for scale, grading, notches, annotations and the relationship between cutting and sewing lines.[S22] These options are useful because they make interpretation visible. They also show why the sender and receiver must agree settings before exchanging production work. Format support is a documented capability, not evidence that any two installations will preserve every required meaning.

The exchange agreement should name the sending and receiving systems, their releases, the selected DXF convention, the unit, the size group and the grading method. It should state whether grading travels inside the DXF or in a separate rule file, whether annotations are expected, how notches are placed, and which line represents the cut boundary. CLO notes that cutting and sewing lines may change depending on the exporting CAD settings, and its interface provides a swap option.[S22] That makes line-role verification a required readback check rather than a visual assumption.

The agreement should also identify a representative test packet. For the illustrative shirt, the packet would include a base size and the agreed size extremes, panels with straight and curved boundaries, a sleeve and cuff with intentional easing, a collar assembly, notches, seam allowances and internal lines. The purpose is to expose interpretation problems on geometry that the team actually intends to use. A simple rectangular panel can pass while curve points, grading or annotations fail elsewhere. The packet is a GFT test design, not a universal interoperability standard.

Exhibit 36 makes the exchange loop explicit. The loop ends only when the receiving technician has compared evidence against the approved source revision. A failed import returns to the responsible sender with settings, screenshots or measurement records attached. It does not become an unlogged manual repair at the receiving desk.

Procurement should require this evidence before dependence on a workflow, not after a production deadline has made correction expensive. A vendor demonstration can establish that an export screen exists. Only the receiving test can establish that the intended team can use the resulting artefact. Even a successful test remains bounded to the versions, options and garment features examined.[S19][S22]

Opening a file proves that the receiving parser accepted something. It does not prove that the resulting pattern is the approved pattern. Readback must compare meaning at the sewing and cutting boundary. The technician should identify every expected panel, check cut quantities and orientation, measure agreed landmarks, inspect curves and confirm that notches and internal lines remain associated with the correct pieces. Graded sizes need their own comparison because a correct base size can conceal a missing or altered rule.[S22]

The comparison should distinguish three outcomes. Pass means the required meaning survived without intervention. Corrected import means the receiver changed a setting or repaired an artefact, which must be logged and fed back to the sender. Fail means the packet cannot support the intended next action. Combining corrected and untouched files into one pass rate would hide recurring exchange work. Where a technician approves an exception, the reason, affected sizes and downstream retest should be attached to the exact revision. NIST's emphasis on context-specific evaluation supports defining these outcomes before judging the workflow.[S19]

Annotations deserve the same scrutiny as coordinates. CLO provides an option to import pattern annotations and separate controls for notch placement and duplication.[S22] A panel can therefore retain its outline while losing the label that connects it to the bill of materials or construction sequence. Similarly, a notch shifted from the sewing line to an unintended outline can change assembly interpretation without changing the panel bounding box. Readback needs semantic checks, not only geometric tolerance.

Exhibit 37 separates what the opened vendor documentation states from what remains for the local team to prove. The right-hand column is deliberately a test requirement, not a compatibility claim.

The table does not rank DXF conventions. AAMA may suit one receiving workflow, ASTM another and Standard DXF a different purpose. The correct choice is the one agreed and demonstrated for the actual information required. Nor does a passed shirt packet establish readiness for every garment. New symmetry, layering, padding, special notations or grading structures can introduce different failure modes and should extend the test set deliberately.