How AI Reads GD&T

On a precision-machined aerospace housing I was reviewing last month, the customer's drawing had exactly one GD&T frame on Sheet 4. A position tolerance. Diameter zero point zero five, with a material condition modifier, referenced to three datums in a specific priority order. The quality engineer at the supplier missed the modifier. Their CMM program measured the position against the wrong reference. The first article failed inspection. Three weeks of production paused.

The cost of that single misread GD&T frame was somewhere around fourteen thousand dollars in rework and lost margin. The supplier did nothing wrong on the part itself. The dimensions were perfect. The surface finish was perfect. The material was correct. They just missed one symbol in one corner of one feature control frame.

GD&T is the densest, most-misunderstood part of an engineering drawing. It is also the part where AI extraction matters the most, and where the gap between "works" and "fails" is widest. If you understood how AI reads the rest of a drawing in Chapter 03, this chapter is the deep dive into the hardest layer: the geometric and dimensional tolerancing system that aerospace, precision machining, and high-end fabrication all depend on.

Same approach as Chapter 03. No marketing. Just the genuine technical story of how a modern AI system parses a GD&T frame, where it works well, where it fails, and what the honest accuracy numbers actually mean.

First: Why GD&T Was Engineered to Be Hard to Read

Most of an engineering drawing is text and numbers. Linear dimensions, material specs, surface finishes, general notes. AI extraction tools have been able to handle text and numbers reasonably well for years.

GD&T is different. It was designed by humans, for humans, as a compact visual language for expressing relationships and conditions that plain dimensions cannot capture. Each GD&T frame packs five or six independent pieces of information into a single rectangular callout the size of a postage stamp.

That density is the entire point. A senior quality engineer can read a feature control frame in two seconds and immediately understand: what type of tolerance applies, how tight it is, which datums govern it, what material condition modifies it, and which features it controls. That is incredible expressiveness in a tiny visual unit.

For an AI system, that same density is exactly what makes GD&T the hardest layer of a drawing. The system has to do five independent things, and get all five of them right, or the row in Form 3 is wrong. There is no partial credit on a GD&T frame. Either the inspector measures against the right reference with the right method to the right tolerance, or the first article fails.

Before we explain how AI handles each of those five things, it is worth understanding why GD&T is structurally different from the rest of the drawing.

It is a visual language, not text

A linear dimension is text: a number with a tolerance, sitting between two arrows. OCR can read that. A GD&T frame is a structured visual object: a rectangle divided into compartments, each compartment containing a different type of information. Reading it requires understanding the structure, not just the characters.

Each symbol carries multi-axis semantic load

The position symbol (⊕) does not just mean "position." It implies a tolerance zone of a specific geometry (cylindrical, between parallel planes, etc.), a measurement method (CMM with datum simulation), and an interpretation that depends on the datum reference frame that follows. The same visual symbol can mean dramatically different things depending on what comes after it in the frame.

Datum references make it relational

A linear dimension is self-contained. A GD&T frame is relational: the tolerance is meaningful only relative to one or more datums defined elsewhere on the drawing. Reading the frame requires the AI to resolve those datum references against the rest of the drawing, in priority order, with any modifiers applied.

Modifiers make it conditional

The (M) modifier after a tolerance value or datum changes the entire meaning. Maximum Material Condition (MMC) introduces bonus tolerance. Least Material Condition (LMC) introduces it from the other direction. The same numeric tolerance with a different modifier becomes a different physical acceptance criterion.

Composite frames make it nested

Some GD&T frames are composite: a single feature control frame contains two or three rows, each row imposing a different tolerance with different datums on the same feature. The position of a hole pattern, for example, might be controlled both to a part-level datum frame and to a pattern-level datum frame, with different tolerance values for each. Reading this requires understanding nested logic.

The Five Layers of a Feature Control Frame

To explain how AI parses GD&T, you need to understand the structure of a feature control frame. Every frame, no matter how simple or complex, decomposes into the same five layers.

Layer 1: Recognizing the Symbol

The first job AI has, when looking at a GD&T frame, is recognizing the symbol in the leftmost compartment. This sounds easy. There are only fourteen official symbols. Most of them look nothing like each other.

In practice, the failure modes are specific.

Visually similar symbols. Perpendicularity (⊥) and parallelism (∥) are similar enough that a low-resolution OCR pass can confuse them. So can angularity and parallelism. Position (⊕) and concentricity (a circle with crosshairs) are different but rendered at small scales become near-identical pixel patterns.

Drawing scale and resolution. A GD&T frame at the title-block scale of a B-size drawing might be 4mm tall. Rasterized at 300 DPI, that is roughly 48 pixels. Modern vision models handle this. Lower-resolution scans, particularly photocopies of older drawings, can degrade the symbol below the threshold of reliable recognition.

Custom or proprietary symbols. Some defense and aerospace primes use non-standard symbols that mean something company-specific. These look like GD&T but follow internal conventions documented only in the prime's drawing standard. AI trained on public ASME Y14.5 standards will miss or misinterpret these.

Modern vision-language models hit roughly 96 percent accuracy on symbol recognition for clean ASME Y14.5 frames at standard scales. That sounds great until you consider that a drawing with 30 GD&T frames will have one frame with a wrong symbol on average. And the symbol determines everything downstream.

Layer 2: Parsing the Tolerance Value

Once the symbol is identified, the next compartment contains the tolerance magnitude. This is text: a number, sometimes preceded by a diameter symbol, sometimes followed by a modifier.

This is the layer where AI usually performs best. Numbers are what OCR was designed to read. With modern vision models, accuracy on tolerance values approaches 98 percent for clean drawings.

The failure modes that remain are specific.

Decimal interpretation. European drawings use comma as decimal separator. American drawings use period. A "0,05" on a European drawing means 0.05 mm. An AI trained primarily on American drawings can read this as the number "5" with a comma artifact. Wrong by a factor of 100.

Diameter symbol detection. Whether the ø appears before the tolerance value changes the tolerance zone from rectangular to cylindrical. This is a single character but a massive semantic difference. AI has to detect the presence or absence of that symbol reliably, even when it overlaps with the compartment border.

Bilateral notation. Some tolerances are written bilaterally: "0.05/0.02" meaning a 0.05 zone with 0.02 inner. AI trained on simple frames may parse this as two separate values.

Layer 3: Resolving Modifiers

Here is where accuracy starts to degrade. Modifier detection is the single largest source of GD&T parsing errors today.

The modifier symbols themselves (M, L, F, P, circled) are visually compact and often render at the smallest font on the drawing. A modifier in a frame at the title-block scale might be 1.5mm tall. Rasterized at typical resolution, that is around 18 pixels of information distinguishing M from L from O. Modern OCR handles this reasonably well at clean resolution, but degrades quickly on scanned or photocopied drawings.

Beyond detection, the harder problem is interpretation. A modifier in a specific compartment position changes the meaning of a specific element of the frame. (M) after the tolerance value means MMC applies to the feature being controlled. (M) after a datum letter means MMC applies to the datum simulation, not the feature. Same symbol, different position, different physical meaning.

Modern vision systems hit roughly 88 percent accuracy on modifier detection and placement for clean drawings. On scanned or older drawings, that drops to roughly 80 percent. This is the single biggest accuracy gap in GD&T parsing today.

For Mavlon, this is one of the areas we route low-confidence extractions to the human review step explicitly. A frame with a detected modifier is always flagged for visual verification by the quality engineer, because the cost of getting this wrong is so high relative to the cost of taking ten seconds to confirm it.

Layer 4: Datum Reference Chain

The datum reference compartments are where GD&T becomes truly relational. The frame says: this tolerance applies to this feature, measured relative to these datums, in this order, with these modifiers.

AI has to do four things here.

First, identify which compartments contain datum letters versus which contain other content. The frame format is standard, but compositing varies. Some drawings use vertical separators between compartments; some use horizontal separators with multiple rows (composite frames).

Second, parse the priority order. Primary, secondary, tertiary. The first datum letter encountered is the primary datum. The order is physical: the primary datum constrains three degrees of freedom, the secondary constrains two, the tertiary constrains one. A reordered datum chain produces a physically different measurement setup.

Third, detect modifiers attached to each datum letter. B(M) means the secondary datum is simulated at MMC. C(L) means the tertiary datum is simulated at LMC. Same letter without a modifier defaults to RFS simulation.

Fourth, resolve each datum letter against the rest of the drawing. The frame references "A," but A is defined elsewhere on the drawing as a specific face or feature or axis. AI has to walk the drawing graph to find where datum A is defined, what kind of feature it is, and what its size is. Then it has to do the same for B and C.

The fourth step is where AI capability becomes really interesting. Modern systems do this through a graph representation of the drawing. Each datum is a node. Each datum feature symbol on the drawing is an edge from the datum to the surface or feature that defines it. The feature control frame is a node with edges to each datum it references, with priority and modifier metadata on each edge.

Datum reference resolution accuracy in modern systems sits at roughly 82 percent for typical aerospace drawings. The errors are usually one of three things: missed modifier on a datum, wrong priority order (typically a swapped secondary and tertiary), or failure to resolve a datum letter to its definition (the datum feature symbol is on a different sheet or in an unusual location).

Layer 5: Composite Feature Control Frames

A composite frame is what happens when a single feature control frame has multiple rows, each row imposing a different tolerance with different datums. The classic example is a hole pattern with two position requirements: a tight pattern internal position (e.g., ⊕ 0.05 to A B C) and a looser overall location (e.g., ⊕ 0.25 to A).

Composite frames are the hardest GD&T case for AI today, and they are unfortunately common on aerospace drawings. Pratt & Whitney engine drawings frequently have composite frames on the critical hole patterns. Boeing structural drawings use them for coordinated holes.

The challenge is twofold.

First, recognizing that the frame is composite at all. A composite frame has a single tolerance symbol in the leftmost compartment that spans multiple rows. AI has to detect that the rows below share the upper symbol, rather than treating each row as a separate frame with a missing symbol.

Second, interpreting the relationship between rows correctly. In a position-position composite, the upper row controls the location of the pattern as a whole, and the lower row controls the location of the features within the pattern relative to the pattern. This is a different mathematical interpretation than two independent frames would imply.

Accuracy on composite frames in modern systems sits at roughly 75 percent for typical aerospace drawings. This is the weakest area in GD&T parsing today. For shops doing high-end aerospace work with Boeing or Pratt & Whitney, composite frame parsing reliability is often what determines whether AI extraction is usable for their specific drawings or not.

Building the Relational Graph

Everything we have described so far happens during extraction. The AI reads each GD&T frame, parses its five layers, and stores the result. But the parsed result is only useful if it connects to the rest of the drawing.

Modern systems build what we call a relational graph during extraction. Each drawing element becomes a node in a graph. Each relationship between elements becomes a typed edge.

For GD&T specifically, the graph looks like this. A feature (say, hole_001) is a node. The dimension that gives it a diameter is another node, connected by a "dimensioned by" edge. The feature control frame that controls its position is another node, connected by a "controlled by" edge. The datums that the FCF references are nodes too, each connected to the FCF by a "references" edge that carries priority and modifier metadata.

This graph is what allows the AI system to answer questions like: "what tolerances apply to hole_001?" "what is the inspection method for the position tolerance on hole_001?" "what is the bonus tolerance available given the current size of hole_001?"

Without the graph, the AI knows that "⊕ ø 0.05 M | A | B(M) | C" exists somewhere on the drawing. With the graph, the AI knows that this specific frame controls the position of hole_001 with a tolerance of cylindrical 0.05 mm at MMC, referenced to face A as primary, hole B at MMC as secondary, and surface C as tertiary, with bonus tolerance available proportional to how much hole_001 departs from its MMC size.

Form 3 generation is then a graph walk. Each row of Form 3 is produced by traversing the graph: find every feature, find every tolerance applied to it, find the inspection method implied by each tolerance, emit a row.

For drawings with dense GD&T, this graph is what makes Form 3 generation reliable. Without it, the AI is just producing a list of extracted frames with no connection to the features they control.

Honest Accuracy Numbers on GD&T

Here are the actual numbers from real production work. These reflect what modern AI systems achieve on aerospace drawings of moderate to high GD&T density. Numbers degrade meaningfully on older or scanned drawings.

What these numbers actually mean

The end-to-end number, 78 percent, is the percentage of GD&T frames where every layer is parsed correctly. If a drawing has 30 GD&T frames, the AI will produce 23 or 24 of them with full accuracy, and the remaining 6 or 7 will have at least one layer wrong.

This sounds discouraging. It is not, when you understand the workflow it enables. The AI's job is not to be perfectly correct on its own. It is to do 95 percent of the cognitive work and surface the remaining 5 percent for human review. A quality engineer reviewing 30 AI-extracted GD&T frames, with 7 of them flagged as low-confidence, takes roughly 5 minutes. Reading and extracting the same 30 frames from scratch takes roughly 90 minutes.

The accuracy number that matters is not "how often the AI is right by itself." It is "how often the AI plus the review step produces a correct result." That number, with a competent QE review, is 99 percent or higher. Which is the same accuracy as a competent QE doing the work manually, in a fraction of the time.

Where AI Still Gets GD&T Wrong

Below the headline accuracy numbers, the specific failure modes are consistent and worth knowing.

Handwritten modifier corrections

Old drawings frequently have handwritten corrections layered over printed GD&T frames. A printed (M) might be crossed out and replaced with a handwritten (L). AI vision models trained on clean drawings struggle with this. The model either reads only the original printed modifier or both, producing an ambiguous extraction. This is a "flag for review" case in every modern system.

Non-standard datum schemes

Some legacy drawings use datum schemes that predate ASME Y14.5-2009. The datum letters might be defined with non-standard symbols, or the priority might be implied by sheet location rather than by frame order. AI trained on the modern standard can misinterpret these.

Customer-specific frame conventions

Boeing, Airbus, Pratt & Whitney, and Lockheed each have internal conventions that layer on top of ASME Y14.5. A frame on a Boeing PVS drawing might use a Boeing-specific tolerance modifier that means something different than the equivalent ASME modifier. Generic AI will get the ASME interpretation right and miss the Boeing-specific layer.

Profile tolerances with complex zones

Profile of a surface with a non-uniform tolerance zone (different tolerance values on different parts of the same profile) is a specific edge case where AI accuracy drops sharply. The frame itself is readable, but interpreting which portion of the profile gets which tolerance requires reading additional annotation that is often implicit.

Frames adjacent to dense annotation

When a GD&T frame sits next to a cluster of other annotations, leader lines, or dimensions, the AI's region segmentation can split the frame incorrectly. A frame may be detected as two pieces, or the boundary between the frame and its leader may be confused. This produces missing or duplicated extractions.

Very small frames at low scale

Frames at less than 3mm tall, especially on drawings rendered at low DPI, fall below the AI's effective resolution threshold. Symbol and modifier detection both degrade. This is a fundamental physics constraint: there are not enough pixels to disambiguate similar symbols.

Why This Matters for FAI and Quoting

Most of this chapter has been about how AI parses GD&T at a technical level. The reason this technical detail matters is operational: GD&T parsing reliability directly determines whether AI extraction is usable for the two highest-value workflows in a precision shop, which are quoting and first article inspection.

Quoting

When an estimator quotes a precision part, the GD&T frames on the drawing tell them what manufacturing process is required. A position tolerance of 0.05 mm to a multi-datum reference frame implies a process capability: this feature has to be machined on equipment that can hold that tolerance, with a CMM inspection step, with the right datum simulation hardware. A 0.5 mm position tolerance on the same feature implies a much looser process and a much lower cost.

Missing GD&T during extraction means missing process implications during quoting. The estimator quotes the part as a standard machining job, the shop accepts the work, then the quality engineer discovers the GD&T requirements after production starts, and the margin is gone.

First Article Inspection

Every GD&T frame on the drawing becomes one or more rows on AS9102 Form 3. The Form 3 row specifies the requirement, the acceptance criteria, the inspection method, the actual measured value, and the pass/fail determination. If the GD&T frame was parsed wrong, the Form 3 row is wrong, and the FAI submission may either fail customer review or, worse, produce parts that ship to the customer with wrong inspection records.

For Boeing, Airbus, Lockheed, RTX, and Northrop, GD&T documentation on Form 3 is one of the most-audited elements of the submission. Customer quality teams have seen every kind of error. They look for them specifically. A wrong modifier, a wrong datum priority, or a missing layer of a composite frame is a guaranteed rejection at most aerospace primes.

The Review Step Is Not Optional for GD&T

Across the entire technology stack, the consistent message is that AI plus human review beats AI alone, and beats human alone, on time and accuracy. For GD&T specifically, the review step is not optional. It is structurally required.

Even at the strongest single-layer accuracy numbers (96 percent on symbol recognition), a drawing with 30 GD&T frames produces an average of one to two frames per drawing where at least one layer is wrong. Combine that with the cost of a single wrong frame (often thousands of dollars in rework), and the math is clear: the workflow has to include human verification of GD&T extraction.

The good news is that human verification of pre-extracted GD&T is fast. A QE who reads 30 frames from scratch takes roughly 90 minutes. A QE who reviews 30 pre-extracted frames, with 5 to 7 flagged as low-confidence, takes roughly 8 minutes. The AI compresses the cognitive work by an order of magnitude. The review is what makes the compressed output reliable.

For Mavlon specifically, every GD&T frame extracted carries a confidence score for each of its five layers. Frames where any single layer is below a configurable threshold are flagged for the QE's review. Frames where all five layers are high-confidence are passed through. The QE focuses their attention exactly where it matters, which is the small subset of frames where AI judgement was uncertain.

How GD&T Parsing Will Evolve

The technology is improving. Here is the realistic trajectory for the next 18 to 24 months, based on what is shipping in current research and what is going into production systems now.

Composite frame accuracy will improve fastest. This is the weakest area today and the area receiving the most attention. Specialized training on composite frame patterns, particularly from aerospace drawings, is producing meaningful accuracy gains. Expect composite frame parsing to move from 75 percent to 85 percent within 18 months.

Modifier detection will close to 95 percent. Improvements in vision model resolution and in fine-tuned symbol recognition heads are addressing the modifier gap. The current 88 percent will likely sit around 94 to 95 percent by end of 2027.

OEM-specific fine-tuned models will emerge. Boeing PVS, Airbus AIPS, RTX/PWA, and Lockheed standards each have their own layered conventions. Vendors who collect enough drawings under each standard can fine-tune dedicated models that outperform generic systems on that specific OEM's drawings. Mavlon is one of several teams working on this.

Real-time validation during design. Beyond extraction from existing drawings, AI will increasingly validate GD&T at the design stage. CAD plugins that check whether a feature control frame is internally consistent, that flag missing datum definitions, and that warn about ambiguous specifications are starting to appear. This will reduce the population of malformed GD&T frames in circulation over the next several years.

The review step will get faster. Better UX for verifying flagged extractions, side-by-side visualization of the drawing region and the parsed result, and progressive trust models where high-confidence extractions are silently approved will compress the review time further. The current 8 minutes for a 30-frame drawing will likely become 3 to 4 minutes.

GD&T extraction is not a solved problem. But it is a tractable problem, and the trajectory is clear. The shops that integrate AI extraction with disciplined review workflows today will compound an advantage that, by the time GD&T parsing is fully solved, will be very hard for laggards to close.