Steel data provenance
How should shape data, identifiers, units, versions, transformations, missing values, and source limitations be recorded so a downstream result can be reproduced?
Read the data field guideQSD Research
Research for a more verifiable steel workflow.
QSD Research connects structural steel data, detailing practice, automation testing, and human accountability. The program begins with transparent methods and public technical assets; quantitative field findings will appear only when eligible evidence exists.
Research program
Four connected lines of inquiry.
The aim is not to produce more AI language. It is to make sources, transformations, exceptions, reviewer decisions, and limits easier to inspect across the steel information chain.
Detailing QA and revision intelligence
Which bounded comparison and completeness checks are suitable for automation, how should known issues be seeded, and which misses require escalation?
Review the evaluation methodHuman-accountable AI
How can AI assist retrieval, comparison, classification, and explanation while preserving citations, uncertainty, role boundaries, and qualified approval?
Read where AI helps and failsInteroperable steel information
How do models, drawings, reports, revisions, CNC data, and open exchange formats carry different scopes of authority through detailing, fabrication, and erection?
Explore BIM, automation, data, and AIPublication policy
Rules for a useful research record.
Methods and boundaries matter as much as results. These principles apply to QSD-authored benchmarks, field studies, technical datasets, and case studies.
State the evidence class
Label demonstrations, controlled tests, internal benchmarks, permissioned case studies, independent reproductions, and third-party sources distinctly.
Version the inputs
Identify the source, edition, units, transformation, software or ruleset version, date, and exclusions required to reproduce the result.
Report difficult cases
Include misses, incorrect flags, indeterminate outputs, unsupported conditions, and reviewer overrides instead of publishing successes alone.
Protect project information
Do not publish confidential drawings, model content, business data, or identifiable results without clear authority and agreed presentation.
Keep authority visible
Distinguish educational material and assisted workflow output from engineering decisions, fabrication release, safety planning, and contractual authority.
Correct the public record
Date material changes, preserve a clear version trail where practical, and correct significant errors without silently rewriting the finding.
Current research record
Public material available today.
This initial collection provides method, terminology, and inspectable examples. It is not presented as peer-reviewed academic research or representative field-performance evidence.
Proof & Automation Lab
The evaluation protocol, metric definitions, human-review gates, verified technical baseline, and unfilled future evidence register.
Open the labChoosing trustworthy steel data
A source-to-output framework for provenance, identity, units, missing values, version control, and validation contracts.
Read the guideWhere AI helps and fails
Task selection, approved sources, prohibited data, test cases, citations, human review, exception logs, and controlled expansion.
Read the guideThe steel information chain
How decisions and deliverables move among design, detailing, fabrication, erection, inspection, and project information systems.
Read the articleTry a controlled example
Reproduce a result before drawing a conclusion
These exercises use public references or your own non-confidential test files. They illustrate how to establish expected results and record exceptions; they are not completed client studies or evidence of a productivity gain.
A nominal-weight check
In Material Takeoff, enter four W18X35 members at 20 feet each. The reference designation represents 35 pounds per foot, so the expected nominal total is 4 × 20 × 35 = 2,800 pounds. Change two members to 21 feet: the expected total becomes 2,870 pounds. Compare the result and exported rows, including quantities and units, with the worked revision example.
Record the source edition and tool release with the calculation. This checks the selected data and arithmetic; it does not establish delivered weight, connection allowances, availability, or structural suitability.
A known drawing-search limitation
Prepare a small PDF containing selectable text with a known mark, then prepare an image-only version of the same visible page. Search both in Search Drawing PDFs. Check the text PDF's result against the page itself, and record that the image-only copy needs a different process because the tool does not perform OCR.
Keep the files, search term, browser version, expected matches, actual matches, and unsupported cases together. A correct result on this sample does not establish complete coverage of an unrelated drawing set.
For a repeatable record, use the published evaluation method and its metric definitions. Report a failure with the exact tool version and a shareable reproduction through Contact. The editorial disclosure explains how published material is prepared and how source limitations are handled.
Open research roadmap
Planned work without invented completion dates.
Priorities will move as qualified reviewers, suitable data, design partners, and product boundaries become available.
Controlled test library
Build non-confidential cases with known expected results for selected comparison, completeness, and data-validation tasks.
Versioned benchmark release
Publish the task boundary, cases that can be shared, scoring, misses, false alerts, supported coverage, and maintainer environment.
Permissioned design-partner study
Establish a pre-pilot baseline, measure the same bounded work, retain qualified approval, and publish only agreed information.
State of AI in steel detailing
Develop an industry study only after the questions, respondent criteria, sampling limits, privacy approach, and analysis plan can be published.
Participate without surrendering the truth
Help turn one real workflow into measured evidence.
QSD is seeking bounded, reviewable problems rather than permission to make sweeping claims. Confidentiality, publication rights, success criteria, and human responsibility should be agreed before a pilot begins.