The trust is in the method, and we publish it.
Every accuracy number on this page is measured against internal ground truth — cases we labelled by hand. Nothing here is customer-measured yet; when it is, we will say so and keep the two separate.
Deterministic screening — no model judgment on hard constraints
A stated constraint is enforced in SQL, not by a language model. A VOC ceiling is a WHERE clause; a maximum particle size compares against the top-cut column, never the mean range. The model never decides whether a product passes a hard requirement, so it cannot be argued into a wrong pass.
Verbatim provenance, or the value is rejected
Every extracted value — from the inquiry and from the data sheet — must carry a quote that appears verbatim in the source. If the quote is not found, the value is dropped rather than trusted. Provenance is shown in the product, not hidden behind a tooltip.
An eval gate that permits zero constraint violations
Before a release ships, an evaluation harness runs a labelled case set and computes top-1 / top-3 accuracy and a false-pass rate — a product that survived screening but should not have. The gate requires that rate to be 0.000, and an independent verifier re-checks every survivor against the constraints. A single false pass fails the build.
Permanent prompt-injection tests
Inquiries are untrusted text. A suite of injection cases confirms that instructions hidden in an inquiry can only ever appear as a quote or a "missing info" note — never as a live constraint, and never as a recommendation of a screened-out product. These tests are permanent; they run on every change.
Cross-producer evaluations
The harness runs per producer, because a method that works on one catalogue can fail on another's units and conventions. Each producer's cases are scored separately, and a false pass on any producer fails the gate. One producer's data never influences another's results.
Generated-case invariant testing
Because the screen is deterministic and pure, we generate random catalogues and constraint sets and assert six invariants across more than a million cases: no constraint-violating product ever survives; a max-size limit never reads the mean range; a NULL on a constrained field is always flagged unknown; every elimination carries a reason; the survivor / eliminated / unknown buckets partition the catalogue exactly; unit conversions round-trip. Property-based on every push, with a full million-case sweep nightly.
Application-match against producers' own published recommendations
Separately from our internal ground truth, we collect cases from producers' public selection guides — where the producer themselves names a grade for a stated application — and check whether GRADE's ranked recommendation surfaces that grade. Every pair keeps its source URL and the verbatim sentence. What it measures: application-matching. What it does not measure: our core constraint screening — most published recommendation sentences carry no hard constraint. We label this number "application-match, externally labelled, n=X" and never merge it with the internal eval. The false-pass rate stays 0.000 on these too. This is an early, small, honestly-limited benchmark that grows as the corpus does.
No third-party penetration test. No SOC 2 report. Both are funded and planned, not done. Accuracy figures are against internal ground truth, not customer-confirmed sales. We would rather tell you that than imply a certification we have not earned. Security questions: security@gradeltd.com.