How does a state-diff grader catch an error a fields-only check misses?

An evaluation lead runs a candidate model on case FAC-6105, a line of credit for a retail group. The model creates the facility, links both parties and attaches security SEC-6105. Its final message says booking is complete, and a fields-only check passes: limit, borrower and status all match the approval.

The state-diff grader goes further. It compares the schedule table with the offer and finds the model set the rate basis to a fixed rate where the offer specifies a variable rate with a review date. Nothing in the model's text mentions the choice.

Grader output for the case: - `limit_match`: pass - `parties_linked`: pass - `security_attached`: pass - `rate_basis_match`: fail - `review_date_present`: fail - `claimed_complete_but_state_invalid`: true

**Consequence.** In a real lender this error would not surface until the first rate review, when the facility would not reprice and interest would be undercharged until someone noticed. A grader that only checks headline fields would score the run as a success and reward the shortcut.

**What it teaches.** Evaluating facility recording needs field-level checks on the derived records, not just the ones typed in from the approval. It also shows why the claim in the final message should be scored separately from the final state, so a lab can measure overconfidence.