Eligibility Before Expensive Scoring

These examples and illustrative results are independently authored teaching materials, not measured model results.

Use case

Retrieved candidates go to scoring but teaching inputs include duplicate C1, disallowed C2 and C3 missing permission metadata. Obtain only authorized eligibility metadata rather than sending all bodies and hiding forbidden output later.

Mechanism

Define hard rules/metadata, verify scope/access/expiry/exact identity and remove provably ineligible items before scoring. Unknowns remain pending. Score soft relevance among eligible items and retain reasons separately; recheck changed content/scope.

Bad example

Send all candidate bodies to the model, then hide highly ranked C2 from the final answer and call its source unused.

Good example

Inspect permitted eligibility metadata first. Keep C1 once, exclude C2 before body scoring and leave C3 pending permission. Score eligible IDs only; record rejected/pending reasons and inputs. Resolve gaps within allowed scope, not access bypasses.

Why the change matters

Hiding output cannot retract content already sent for scoring. Eligibility first also avoids spending on unusable candidates while separating hard constraints from ranking preferences.

Observable expectation

The illustrative scoring set contains one C1, no C2 body and an explicit pending C3. Changed permission/expiry triggers recheck.

Compare reasons with actual scoring inputs, not final absence alone.

Limits

Dedup needs correct scoped identity; similar text is not necessarily one resource. Metadata can expire and access/license cannot be inferred from scores. Other source pipeline stages are not mandatory architecture for this method.