Define the criteria. Let test results guide development.

Define goals and assumptions:Specify the material properties, structural dimensions or process workflow being compared, and the decision the results will support. Review existing information, available software, data permissions and the operating environment. Mark unknown inputs as assumptions and list missing conditions separately.

Fix the evaluation criteria:The client and domain specialists confirm the reference baseline, evaluation metrics and pass conditions. For experiments or equipment, the experimental team documents batch differences, operating ranges and measurement error. Criteria are fixed before testing; any subsequent change retains its reason and version.

Outputs from this step:Task specification, input and assumption lists, reference baseline and evaluation method. If critical inputs are missing, arrange additional tests or information first, then fix the scope of the round.

Combine AI exploration with reliable solving:Choose the required combination of data structuring, candidate generation, parametric modelling and external tool calls. AI may help organize information or propose candidates. Calculations involving physical, chemical or engineering metrics are performed by code, mechanistic models or solvers confirmed as available by both parties. Tools, licences and interfaces are agreed at the start.

Fix versions and involve domain specialists:We build dedicated interfaces or automated workflows that fix data sources, physical units and versions. Domain specialists review candidate generation conditions and model assumptions, and decide whether outputs are suitable for further computation, detailed design or experiments.

Outputs from this step:Candidate comparison tables, parameter configurations, code or models, operating steps and input/output records. Check mappings and exception handling before choosing which outputs enter evaluation.

Independent checks at four levels:Each checking path corresponds to a specific task level: software checks inputs and program execution; geometry checks solids and component assembly relationships; computation uses comparisons to examine model discrepancies; experiments check physical performance under specified conditions. Generation and evaluation are organized separately, and historical log replay is distinguished from a fresh execution.

Discuss stability within its tested range:Domain specialists judge whether discrepancies affect the decision. Model calibration and independent validation are separated by batch or operating condition. The experimental team records measurement conditions and repeat results. Stability conclusions describe only the ranges, variation and failures already checked; conditions outside that range require further tests.

Outputs from this step:Comparison results, errors, lists of passed and failed checks, and the applicable range. Software, model, simulation, platform and physical states are reported independently; passing at one level is never automatically extended to another.

Locate the source of failure:When a check reveals a discrepancy or failure, investigate source information, model assumptions, design parameters, implementation and external conditions. Revise within a defined scope when the evidence is clear. If several causes remain possible or an environment is missing, retain the issue and specify the conditions still needed.

Independent out-of-sample regression checks:Rebuild or rerun affected parts and retain the reasons for changes and regression results. Where tuning is required, the revised version must face independent tests not used during tuning. The client and domain specialists use the new results to decide whether to proceed, add tests or expand the scope.

Outputs from this step:Revised version, regression results, unresolved issues and the next-round plan. Check whether the revision introduced secondary differences and whether previous pass conditions still hold.