An experiment selector can make a poor decision before the model ever runs: an untested configuration is exported as a failed print, or an interrupted machine run is given the same label as a completed material experiment.
That distinction matters when successful outcomes are rare and physical tests are expensive. Washington State University highlighted a GRCop-42 additive-manufacturing study on August 24, 2026. The AAAI paper itself was published on March 14. Its abstract describes selecting small batches with a surrogate model, then validating them through directed energy deposition (DED). The research joins WSU computer-science and manufacturing researchers with Aryan Deshwal at the University of Minnesota; NASA developed the alloy.
The data contract proposed below is inspired by that workflow. It is not the researchers' software, a reproduced experiment, or a claim about their internal dataset.
Separate a candidate from an observation
A candidate describes a configuration you might test. An observation describes what actually happened. Creating the first must not silently create the second.
For a small system, consider three records:
- Candidate: an immutable configuration ID, process and equipment references, material-lot reference, geometry revision, parameter values and units.
- Selection: the candidate ID, selection round, model revision, evidence cutoff and a short selection rationale.
- Observation: the run ID, candidate ID, execution state, assessment state, assessment protocol revision and links to retained measurements.
Do not initialize a new candidate with a default outcome of failure. A missing observation means that no outcome is available. It does not supply a negative training example.
Execution and assessment answer different questions
Execution asks whether the physical test completed. Assessment asks what a completed test established under a particular protocol.
An interrupted powder feed may tell you something useful about the process, but it is not automatically equivalent to a completed print that fails a dimensional requirement. Retain the event and its reason. Let the experiment protocol decide whether it belongs in a particular model's training view.
Likewise, a completed run that awaits inspection is neither accepted nor rejected yet. A practical set of assessment states is pending, assessed and invalidated. An assessed record then carries the relevant result and protocol reference. This avoids overloading one Boolean with several meanings.
Freeze the evidence available to each selection
Suppose round 12 selected three candidates using observations available at noon. One result arrives at 15:00, another remains pending, and the third is invalidated because the inspection file belonged to a different specimen.
A later export should not make it appear that round 12 had those results when it chose its candidates. Keep the evidence cutoff and the included observation versions. Append a correction when a record changes; do not rewrite the historical decision as though the corrected evidence had always existed.
This also makes offline evaluation more credible. Replaying a selector with observations that became available only after its selection would leak future information into the comparison.
Preserve the reason for exploring
The WSU announcement describes balancing promising candidates with uncertain regions. A low-confidence candidate can therefore be intentional. A dashboard that labels every uncertain choice as a bad recommendation would misrepresent the experiment's purpose.
Store a rationale such as exploring an under-observed region or checking a promising neighborhood. Those labels should explain the selection objective, not assert that the model has proved a physical outcome. The suggested labels here are product-design examples, not documented fields in the published system.
Keep performance claims attached to their test scope
A headline laser-power value is not a transferable recipe. DED is not interchangeable with laser powder bed fusion, and equipment, material, geometry and evaluation conditions matter.
NASA's separate 2022 GRCop-42 presentation cautions against extrapolating standard bulk-specimen properties directly to thin walls. That is background on geometry dependence, not an independent replication of the WSU result. A useful application should make that distinction visible instead of treating every successful coupon as approval for every part.
For acceptance checks on this proposed contract, verify that an untested candidate cannot appear in the assessed training view, a pending inspection cannot become a negative label by default, and a corrected result cannot alter the frozen evidence for a past selection. These checks test the integrity of the decision history. They do not validate the model's ability to choose a printable alloy configuration.
The important boundary is modest: the model proposes the next physical question; a traceable experiment provides the answer. Keeping those two events separate gives engineers a record they can inspect when a recommendation succeeds, fails or remains unresolved.
Sources and disclosure
- WSU announcement, August 24, 2026
- AAAI paper abstract and publication metadata, March 14, 2026
- NASA NTRS presentation on GRCop-42 size effects, 2022
AI disclosure: This article was drafted and edited using AI, including editorial checks by AI agents. The data contract and example are proposed designs; no manufacturing experiment or implementation benchmark was performed for this article.
Top comments (0)