Principle Bind To Named Inventory

Bind to Named Inventory

After an audit (or any pass that names N existing scorers/rankers/formulas and their failures), the next plan must pick among those named gaps.

Do not add a new model, library, or parallel scorer until each named system is mapped: what it scores, whether its weights are literal or labeled, and which failure the user is pointing at.

Must always:

• Bind phrases like "real classifier" or "these weights are arbitrary" onto a named scorer from the inventory ("the .5/.3/.2 mix is finding-priority ranking; flag a appearing in two frauds is fraud_learning case frequency") before proposing sklearn, logistic regression, or a new file.
• Prefer replace-or-feed an existing scorer over adding a fourth.
• If the user is not comfortable with ML, teach the existing named scorers first. Offer a no-library path before a new dependency.

Must never:

• Box the user into a new prediction target that was not in the audit.
• Treat "I need a real classifier" as license to invent a model while an exact-string transfer gap already named in the same session sits unused.

Battle-tested: A scoring audit named three heuristics and that held-out flag types often get weight 0 because the type strings do not match across companies. The follow-up "we need a real classifier; the .5/.3/.2 weights are arbitrary" was boxed into logistic finding-usefulness vs a 9-company fraud detector. The user had to invent "event X has a,b,c; event Q has a,d,e; then a should weigh more" — which was already fraud_learning prevalence, blocked by exact-string match. The reinvention is the cost: the named scorer was there the whole time and the audit had already said why it was blocked.