When the Model Is Silent: Who Carries the Tolerance?

For most of the history of precision manufacture, the drawing carried two jobs at once. It described the geometry, and it carried the agreement about how far that geometry was permitted to depart from nominal. The second job was always the commercially important one, and it was always visible: a title block, a general tolerance note, a set of feature control frames, all signed and dated.

Model-based definition separates those two jobs. The model becomes authoritative for shape; the tolerance moves into annotation, or into a standard, or — in the cases that cause difficulty — into nothing at all.

What a model states, and what it cannot

A CAD model is a statement of nominal geometry. Its surfaces are perfect, its intersections are mathematically exact and its edges have no radius of cut. It says nothing, of itself, about the thickness of a machining pass, the deflection of a thin wall, the effect of heat treatment or the capability of the process chosen to realise it.

Tolerance is an agreement about permissible departure. It is a commercial and contractual artefact as much as an engineering one, and it does not exist in the model unless someone has put it there deliberately. A feature that is modelled perfectly and specified nowhere is not a well-defined feature; it is an unresolved question waiting for a dispute to attach itself to.

PMI: annotation, or wishful thinking?

Product and manufacturing information is the mechanism by which the gap is supposed to close. The relevant frameworks — ISO 16792 and BS 8888 in the UK context, ASME Y14.41 where US customers are involved — set out how dimensions, tolerances and notes should be presented in a digital dataset.

Two distinctions matter more than the standards themselves. The first is semantic versus graphical annotation. Semantic PMI is machine-readable: a CMM or CAM system can query it and act on it. Graphical annotation is a picture of a tolerance, attached to the model but not intelligible to software. Many release packages in circulation are graphical only, which means the information exists for a human reader and for nobody else.

The second distinction is placement versus association. An annotation that floats near a face without being attached to a specific feature or pattern can be interpreted in more than one way, and on a busy shop floor it will be. The standards are clear that annotation should be associated; practice is frequently less so.

Where the argument surfaces

Disputes of this kind rarely begin in a meeting. They begin at a machine.

A CAM programmer working from the model may never open the annotation at all, particularly where the model has been supplied without a viewer that displays it. The machine operator sees a screen, not a drawing, and has no reason to suspect that a pocket depth carried a tolerance in a document nobody printed. The inspector programs a CMM from the model, aligns to nominal and reports a set of deviations with no pass or fail attached, because the tolerance was never loaded into the inspection plan.

The part is then rejected at goods-in against a requirement that the supplier did not know it had accepted. The buyer's position is that the requirement was in the release package. The supplier's position is that the model, which was described as the authoritative definition, was silent. Both statements can be true at once, which is what makes the argument expensive.

There is a further wrinkle. Where a drawing is retired, the buyer may inadvertently warrant that the replacement dataset is complete. If the specification hierarchy was never written down, the party with the least information — usually the supplier — ends up carrying a risk it never priced.

Clauses worth settling before the drawing goes

Four points are worth agreeing in writing before a 2D drawing is withdrawn from a part family.

A stated hierarchy. The contract should say, explicitly, what governs in the event of conflict: the purchase order, the drawing where one survives, the model, the referenced standards, or a separate specification document. Silence here is not neutral; it simply transfers interpretation to whoever is holding the part when the problem appears.

A default rule for unannotated features. Where the model is silent, something must apply. A general tolerance to BS 8888 or an agreed ISO 2768 class, referenced in the contract rather than assumed, removes most of the ambiguity at negligible cost.

A query duty with a window. The supplier should be obliged to raise queries on any feature it cannot specify from the dataset, within a stated period after receipt, and the buyer should be obliged to answer within another. A supplier that stays silent and machines to its own assumption should not be able to rely on that assumption later.

A record of acknowledgement. Revision control on a model is less obvious than on a drawing. A checksum, a release note and a signed acknowledgement of the dataset — including a declaration that the PMI has been read and is complete — gives both parties something to point at.

A practical test before retirement

Before a drawing is withdrawn, it is worth running one exercise: take the final drawing and attempt to reconstruct every requirement on it from the model and its annotation alone. Anything that cannot be reconstructed is either missing from the dataset or present only as graphical decoration. That list is the exact scope of the argument you would otherwise have later.

Model-based definition is not the problem. It is a sound approach that removes a great deal of duplication and error. The problem is the assumption that retiring the drawing retires the question of tolerance. It does not. It merely changes who is expected to answer it, and a supplier cannot answer a question that the release package never asked.