The same five checks run as software on every upload, which is how a hand review becomes a pipeline without lowering the bar. They are also the checks the benchmark grader applies to an AI’s answer.
Method
How a fix is verified
Five checks on every record, written down and then turned into software: geometry diff, load checks, setup diff, hygiene scrub, tiering.
Sec_001 · The review
Five checks on every record, in this order
- 01Geometry diff
Before and after STEP files are compared. Only the changes in the change log may differ, and all of them must.
- 02Load checks
Loads, boundary conditions and materials are read from the solver input, not the write-up, and must match the spec.
- 03Setup diff
Before and after solver decks are diffed. Mesh, solver settings and conditions must be identical apart from geometry.
- 04Hygiene scrub
Customer names, drawing numbers and anything identifying are removed. Ownership is confirmed.
- 05Tiering
Each record is scored for difficulty: single part to assembly, single physics to multi-physics, with or without cost.
Sec_002 · The pipeline
From hand review to software
Fig_007
Automated review
Geometry diff, load checks, setup diff, hygiene scrub and tiering, run automatically on every upload.
Fig_008
Contributor portal
Engineers upload records, get instant feedback, fix and resubmit. Payment and IP paperwork tracked in one place.
Fig_009
Catalog
Search by industry, physics, tier and failure type. Slices ship with manifests and licences.
Sec_003 · Tiers
Harder records, harder tasks
Fig_010
[ Tier score ]
Tier 1Single part · single physics · one limit
Tier 2Single part · rejected options logged
Tier 3Assembly · contacts and stack-ups
Tier 4Multi-physics · coupled limits
Tier 5Production · optimised for volume, still passes
Tiers are how a slice is defined and how the benchmark gets harder. The design engineer is trained up the same ladder: single parts first, then assemblies, then coupled physics, then production at volume.