GAD Editorial · Published · · English study · Chinese summary · Study dated 2026-10-02
A BIM model can contain the geometry needed for a question while omitting the decision that gives that geometry meaning. A client may have changed a target in a meeting; a design team may have agreed which spaces to include in a comparison. An assistant must locate that decision and apply it to the correct model revision.
IFCMemoryBench studies this combination through 143 multi-session tasks across 19 projects. The preprint was submitted on 13 July 2026. In its main deployment-realistic ingestion condition, the strongest tested system reaches 32.4% answer accuracy. The benchmark is synthetic and human-validated, and the paper shows that results depend partly on the agent using the stored information. The number should stay attached to those experimental conditions. Original paper · Submission history
Give each decision a usable identity
GAD proposes a modest decision register alongside the normal model record. Start with a decision identifier, its source, the person responsible and the date it took effect. Describe the building, level, zone or element set it applies to. Record the unit and any relevant measurement convention. If the decision changes, retain the earlier entry with a clear link to its replacement.
For an illustrative area study, distinguish the measured area in a particular model from the agreed target and the rule for which spaces belong in the total. Those are different pieces of evidence. Combining them in one uncited sentence makes it harder to discover whether an incorrect answer came from the model query, the target or the inclusion rule.
Build questions with known answers
Choose a small set of realistic queries whose results a project team can verify manually. Include one that requires only the model, one that needs an external decision and one that combines both. Add a deliberately superseded decision, a missing source and an ambiguous scope. This is a proposed office test, not part of the published benchmark results.
Ask the assistant to show the model revision and the decision source used in its answer. Have a reviewer follow each reference. A citation earns little credit if it points to a relevant meeting without supporting the particular value or instruction applied.
When the record is incomplete, the desired output may be a precise question rather than a total. Define that behaviour before evaluating the system. Otherwise a confident guess can appear more successful than an appropriate request for clarification.
Diagnose the failed step
Record whether a failure involved preserving the original decision, finding it later or applying it correctly. Keep calculation errors separate. This gives the team a practical remedy: a better source record, a clearer retrieval path or a revised query procedure.
Repeat the same questions after a controlled model revision and a changed client instruction. Check that the response updates only where the new evidence requires it. The useful outcome is a traceable answer that survives project change, with uncertainty visible at the point of use. This article is a documentation review; it neither tests a deployed assistant nor authorises changes to a project model.