Technology
A deterministic engine with models in an advisory seat
The score itself is arithmetic — no model output is ever multiplied into a published figure. Machine learning helps us classify, extract and flag; humans and factor tables decide.
The pipeline
Intake
Partner submits SKU specification and supporting documents through a structured form.
Extraction
Models read documents to propose material, mass and packaging values. Every proposal needs human confirmation.
Scoring
The confirmed inputs hit the deterministic scoring function against a frozen factor set.
Publication
Score, sub-scores, confidence tier and methodology version are written together as one immutable record.
Design constraints we hold to
Same inputs plus same factor version always yield the same score.
A released factor set is never edited in place; changes ship as a new version.
Models propose; they never write a published figure without human confirmation.
Re-scores append a new record rather than overwriting history.
Why not let a model do the scoring?
Because a score people rely on has to be reproducible and explainable line by line. A learned estimator would give us better coverage on sparse data and worse accountability on every published number. We chose accountability, and we accept the coverage cost: products with sparse data get an estimated tier rather than a confident guess.
The places models genuinely help are described in the model cards, and the infrastructure they run on is documented in the cloud architecture.