PartsCleanse AI Engine Validation Benchmark

Enterprise trust posture

Proof controls buyers expect before they upload operational data.

Source purge Uploaded catalog files are deleted after report generation; only summary metrics and Open Findings remain.

No ERP write-back The diagnostic creates evidence for review. It never changes, deletes, merges, or overwrites ERP records.

Local currency Reports display money in the user's selected or country-derived currency, while USD remains the base audit calculation.

Audit trail Report ownership, access, quota, and feedback events are retained for governed review.

Session downloads Excel, Word, PDF, and CSV downloads are available only in the active generation session.

Open Findings Browser findings remain available without retaining the original source catalog rows.

Validation surfaces

What the engine is expected to survive before buyer trust is earned.

50K scale runtime A 50,000-SKU benchmark run is used to validate that the product can process enterprise-sized catalogs without turning into a consulting queue.

Dimension traps Adversarial catalogs include same-description records that differ only by size, pressure, or rating so false positives are visibly penalized.

Material traps Carbon steel, stainless, duplex, bronze, alloy, and specialty material families are treated as critical discriminators, not disposable tokens.

Part-type controls A valve, gasket, bearing, seal, motor, filter, and fastener can share numbers and materials but remain different commercial objects.

Currency normalization Uploaded unit costs can be interpreted in the submitted currency and converted for report display without changing the audit base.

Human review posture The engine creates confidence tiers and review queues. It does not authorize automatic ERP merges or write-backs.

Benchmark matrix

Public test categories and what they prove.

Category	Test pattern	Buyer evidence
Scale	50,000 SKU synthetic benchmark	Runtime, grouping stability, report generation
False positives	Dimension and material trap catalogs	Unsafe look-alike matches suppressed
Recall	Real-world abbreviation chaos catalogs	SAP-style shorthand and verbose descriptions matched
Governance	Tier 1 / Tier 2 / Tier 3 outputs	Human review sequence preserved
Data handling	Source purge after report generation	No source catalog retention

Devil's advocate control

The benchmark is intentionally adversarial.

PartsCleanse AI should not be rewarded for finding only easy duplicates. The harder product problem is rejecting unsafe matches: different sizes, pressure classes, materials, part types, model numbers, and commercial units that share most of their description text.

Buyer test recommendation: include known duplicates and known non-duplicates in your pilot file. A credible diagnostic must show both recall and restraint.

FAQ

Validation questions technical buyers should ask.

Does the benchmark prove every customer catalog will be perfect?

No. It proves the engine has been tested against scale, false-positive traps, abbreviation chaos, and report generation boundaries. Customer data quality still affects final interpretation.

Why publish false-positive controls?

Because unsafe duplicate consolidation is the main risk in industrial MRO catalogs. The product must prove it can say no when descriptions look similar but parts are not interchangeable.

Does AI2COE train on customer uploads?

No. Uploaded catalog files are processed for the run and purged after report generation. The retained data is limited to Open Findings, summary metrics, quota, feedback, report ownership, and audit metadata.

What should a buyer test first?

Upload a small exported catalog with known duplicate and known non-duplicate examples. The best pilot includes both true positives and adversarial look-alikes.