What the category does
Drawing-to-CAD software takes a 2D engineering drawing, usually a PDF, and produces a 3D model. Between the two sit several steps that differ a great deal from tool to tool: reading the sheet (text, lines and views), interpreting it (which number belongs to which feature, under which tolerance), deciding what to do about what the drawing leaves open, building the geometry, and checking the result. A model can be used in any of these steps, and the label on the product says nothing about which.
Ask each vendor to say, in writing, where a model is used and where it is not. A model that proposes readings for a person to confirm and a model that writes geometry directly are different risks. So is a model that runs on a third party's servers and sees your drawings there.
What to test
| Area | Test with | A pass |
|---|---|---|
| Input | Native vector PDF, scanned PDF, multi-sheet drawing, rotated sheet | Each one read, or refused with a stated reason; never silently misread |
| Projection and views | First- and third-angle drawings, sections, detail views at their own scale | Features land on the correct faces; detail scales respected |
| Units | Millimeter, inch, dual-dimensioned, comma as decimal marker | Units taken from the drawing, never assumed |
| Callouts | Ø, R, counterbore, countersink, threads with class and hand, counts such as 4X | Every element of the callout captured, not only the nominal |
| Tolerances | Limits, ± and fits, the general-tolerance block, reference and basic dimensions | Tolerance recorded per value; reference dimensions not used as controls |
| Notes and title block | Material, finish, revision, general notes, flag notes | Captured, and attached to the part with its revision |
| Gaps | Drawings seeded with a missing value and with a conflict | Both raised as questions; neither filled in |
| Out of scope | A part family or feature the vendor says is not supported | Stopped with a reason before anything is built |
Write down which capabilities the vendor says work today and which are planned. Test the first group. Record the second as roadmap, which a trial cannot evaluate.
How to build a test set
01 Choose 20 to 50 drawings from your own work
Pick the part families you actually buy or make, in roughly the proportions you see them. A tool tested only on plates tells you about plates.
02 Include the hard cases on purpose
Add drawings in the other projection and the other unit, with sections, detail views, threads and counterbores, and at least one scan and one multi-sheet drawing, even if the vendor says they are out of scope. A clear refusal is a result.
03 Seed known defects
Copy a few drawings and remove one dimension, or edit one so that two values disagree. You know the right answer in advance: a question. Keep the unedited drawings as controls.
04 Write the ground truth first
Have an engineer who is not running the trial list, for each drawing, every value the part needs with its tolerance, and the questions a careful reader would have to ask. Do it before anyone sees a tool's output.
05 Hold some drawings back
Keep about a third of the set private until the final round, so no vendor can tune to it.
06 Clear the drawings for sharing
Check proprietary and export-control markings, and the customer agreements that cover each drawing, before any of them leaves your network. Where an original cannot be shared, use one of your own designs or a redrawn equivalent.
Scoring
Score each value, not each part. A part with forty correct values and one wrong hole position is a scrapped part, and an average hides it.
| Outcome | What happened | How to weigh it |
|---|---|---|
| Silent error | A wrong value, presented as read from the drawing | Investigate every one; a pattern of them disqualifies the tool |
| Silent fill | A missing value supplied without saying so | As serious as a silent error |
| Unneeded question | A question about a value the drawing states clearly | Costs review time, nothing else |
| Correct question | A question about a value that is missing, ambiguous or conflicting | The expected result |
| Correct reading | The right value, traced to its callout | The expected result |
| Correct refusal | An out-of-scope drawing stopped with a reason | The expected result |
Count silent errors and silent fills apart from everything else. A tool that asks more questions than you would like is slower. A tool that answers questions nobody asked it is unsafe.
Traceability of each value
Every value in the model should point back to where it came from: the sheet, the view, the callout and the revision of the drawing. Without that link a reviewer has to re-read the whole drawing to check one number, and in practice nobody does that for long.
- Pick ten values at random from a generated model and ask the tool to show the source of each on the drawing. Time it.
- Check that the tool keeps three kinds of value apart: read from the drawing, supplied by a person, and set by a default or a rule. A default shown as a reading is a silent fill.
- Edit one value in a source drawing and run it again. The model should change in that value and nowhere else, and the trace should show the new source.
- Check that a value used by a feature, such as a hole diameter, carries its tolerance and its callout text, not only the nominal.
- Run the same drawing twice. A tool whose readings change between runs needs every value confirmed on every run.
Missing and conflicting dimensions
This is the test that separates tools most clearly. Real drawings are often incomplete for the purpose of building a solid: a position shown only by centerlines, a thickness given only in a view that was left off, a chain of dimensions that does not add up to the overall. A tool has three options: ask, stop, or guess. For a part someone will make, only the first two are acceptable.
- On a seeded missing value, the tool should name the value, show where it is needed on the drawing, and hold the build until a person answers.
- On a seeded conflict, the tool should show both values and their sources, not pick one.
- Ask what the tool does when a reading is uncertain. A score can help order a review queue; it should not stand in for a person confirming the value.
- Ask whether answers are recorded with who gave them, and what happens to them when a new revision of the drawing arrives.
A tool that never asks is guessing
Drawings leave values open. If a tool returns a complete model for every drawing in a test set seeded with missing and conflicting values, it has filled the gaps itself. Find out how before using any of its output.
The guide to drawing ambiguity covers what a good question looks like, and missing dimensions covers the common cases.
Output format: B-rep or mesh
The output format decides what you can do with the model afterwards.
| Exact B-rep solid (for example STEP) | Triangle mesh (for example STL) | |
|---|---|---|
| Geometry | Exact surfaces: a cylinder is a cylinder | Flat triangles approximating each surface |
| Hole diameter | Measurable exactly | Depends on the tessellation; a circle becomes a polygon |
| CAM | Toolpaths on true surfaces | Usable for 3D printing; poor for finishing to a tolerance |
| Inspection | Usable as a nominal reference model | Approximate only |
| Editing | Faces and edges can be selected and changed | Hard to edit as features |
Ask for STEP (ISO 10303) and check which application protocol the file uses; AP203, AP214 and AP242 are common. Open it in your own CAD system and check that it is one closed, valid solid rather than a set of loose surfaces, that its units are right, and that its volume and bounding box agree with a hand calculation from the drawing. Most STEP exchange carries the solid without a feature history, so if you need an editable parametric model, test that export separately.
A mesh exported from an exact solid is fine for printing and viewing. A tool that can only export a mesh may never have built an exact solid. Ask.
Verification evidence
"Verified" should name what was checked, against what, and how. Ask for the evidence on one of your own test drawings, and read it.
- Against what. The model against the approved requirements, or the model against its own build plan? The second proves only that the software ran.
- Which properties. Envelope, volume, feature counts, positions, diameters. A list of measured properties is evidence. A single pass mark is not.
- Nominal or tolerance. Most tools build at nominal. Find out whether tolerance zones, GD&T, material or manufacturability are checked, and expect the honest answer to be no for some of them.
- Who produced the result. A measurement of the built geometry, or a model's assessment of it? A model's opinion that a part is correct is not a measurement.
- Reproducibility. Can the same inputs be rebuilt and re-checked later, with the software versions recorded?
The guide to verifying CAD against the drawing lists the checks to run yourself.
Data use and training policy
Your drawings are often your customers' intellectual property as much as your own. Get the vendor's answers in the published terms or privacy policy, not in a sales call.
- Are customer drawings, models or answers used to train shared models? Is the default no, and is any opt-in explicit, separate and revocable?
- Is permission to use a drawing for evaluating the product separate from permission to train on it?
- Do drawings go to a third-party AI provider? Under what terms, and may that provider keep them or train on them?
- Where is the data stored, how long is it kept, and what does deletion remove, including backups and logs?
- Is any price, discount or feature conditional on allowing training?
Export-controlled drawings need a review against your own compliance obligations before any outside system sees them. That decision belongs to your compliance function, not to the vendor.
Security
- Sign-in. Multi-factor authentication, and single sign-on if your company requires it.
- Isolation. How one customer's data is kept from another's, in storage and in the application.
- Encryption. In transit and at rest, and who holds the keys.
- Staff access. Which vendor staff can see your projects, and whether that access is logged.
- Audit. A record of uploads, approvals, exports and deletions that you can review.
- Independent evidence. An independent security audit report (for example against ISO/IEC 27001) or a penetration-test summary, where the vendor has one, and a published way to report vulnerabilities.
Human review
For a part that will be made, a qualified person still has to approve what the model is built from. What differs between tools is where that person sits and what they sign.
- Is there an explicit approval before geometry is built, or only a review of the finished model? Reviewing the requirements with their sources before the build catches an error while it is still cheap to fix.
- Does the approval record who approved, when, and against which revision of the drawing?
- Can the reviewer see each value next to its source, and correct it, in the same place?
- Does a correction create a new record and keep the old one?
Where TruePart is today
The same questions apply to TruePart. These are its current answers on scope, output, verification and data use in the Technical Preview; the security page covers security.
TruePart in the Technical Preview
A deterministic reader handles the one part family in this preview, and uploads go to no AI provider. AI-assisted interpretation is in evaluation and not yet enabled. Currently supporting select part families: today it builds one, rectangular plates with a centered bore and a hole pattern. Scanned, raster and multi-sheet drawings are stopped, with the reason that they need advanced interpretation.
- TruePart exports the exact B-rep solid as STEP, and a mesh of it as STL, once the solid is measured against the specification.
- TruePart's verification covers the envelope, volume, hole count, and each hole at planned position and diameter, at nominal dimensions, and not tolerance bands, material, manufacturability, strength, or safety.
- Customer engineering data is not used to train shared models unless you explicitly choose otherwise.
Run the framework above on TruePart as on any other tool: on your own drawings, with the gaps seeded. The technology page describes the architecture, and pricing shows the plans.
Questions buyers ask
Questions
How many drawings does a fair evaluation need?
Enough to cover the part families you buy or make, the hard cases and a few seeded defects. Twenty to fifty is a workable first round. With fewer, one lucky or unlucky drawing decides the result.
Is an accuracy percentage a useful headline number?
Only with its denominator and its failure modes. On parts with fifty values each, 98 percent of values right means nearly two parts in three (about 64 percent, if errors fall independently) carry at least one wrong value. Count silent errors and silent fills per drawing, on your own set.
Should the output be STEP or STL?
STEP, as an exact B-rep solid, for CAD, CAM and inspection. STL is a triangle mesh for 3D printing and viewing; it approximates curved surfaces with flat facets.
Can a confidence score replace human review?
No. A score can help a reviewer decide where to look first. Whether a value is right is a decision for a qualified person who has seen it next to its source on the drawing.
What should a vendor's data-use terms say?
Whether drawings are used to train shared models, whether that is off by default, how to opt in and withdraw, whether drawings reach a third-party AI provider, where they are stored, and what deletion removes. Ask for it in the published terms.
Sources
- ISO 10303 (STEP), Industrial automation systems and integration: Product data representation and exchange · ISOSTEP as the exchange format for exact B-rep solids, and its application protocols AP203, AP214 and AP242.
- ASME Y14.5-2018, Dimensioning and Tolerancing · ASMEThe callouts, tolerances, reference and basic dimensions a test set should cover.
- ISO 5456-2, Technical drawings: Projection methods, Part 2: Orthographic representations · ISOFirst- and third-angle projection, for the projection cases in a test set.
- ASME Y14.100, Engineering Drawing Practices · ASMEWhat a drawing must state, and so what a tool has to capture.
- ISO 7200, Technical product documentation: Data fields in title blocks and document headers · ISOTitle-block fields a tool should read: revision, material, sheet count.
- ISO/IEC 27001, Information security, cybersecurity and privacy protection: Information security management systems, Requirements · ISO/IECAn information security management standard; a buyer can ask a vendor for an independent audit report against it.
What TruePart does today
- TodayA deterministic reader handles the one part family in this preview, and uploads go to no AI provider.
- In evaluationAI-assisted interpretation is in evaluation and not yet enabled.
- LimitCurrently supporting select part families: today it builds one, rectangular plates with a centered bore and a hole pattern.
- LimitScanned, raster and multi-sheet drawings are stopped, with the reason that they need advanced interpretation.
- TodayTruePart exports the exact B-rep solid as STEP, and a mesh of it as STL, once the solid is measured against the specification.
- LimitTruePart's verification covers the envelope, volume, hole count, and each hole at planned position and diameter, at nominal dimensions, and not tolerance bands, material, manufacturability, strength, or safety.
- TodayCustomer engineering data is not used to train shared models unless you explicitly choose otherwise.