What a keyword support score can tell us

Make the calculation visible#

The verifier receives explanation and grounding objects separately. It takes the headline, fit, taste and texture fields, then adds substitution guidance and quick takeaways as text units. A recursive walk collects string values from the grounding object.

Each unit first passes a literal phrase check for language including treatment claims and allergy-safety certainty. A matching unit increments the dropped count and adds a warning. Otherwise, the algorithm compares lowercase alphanumeric keywords of at least four characters with the grounding-term set.

The share of matching keywords is compared with a threshold, defaulting to 0.28. The response reports supported, dropped and total units, deduplicated warnings, a grounding-term count and an overall outcome. A reader can trace that outcome to a small set of rules rather than an unexposed model judgment.

Keep the interpretation as narrow as the rule#

A sentence can reuse the right words while changing the subject, quantity or negation. Keyword overlap does not establish that the relationship it asserts is true. The unsafe-phrase list is also literal matching, not a complete medical or food-safety policy.

Supported therefore means that a unit cleared this calculation. It does not mean that the claim was independently verified. The response contains counts, not approved claim identifiers or a rewritten explanation.

Nothing in this handler removes text from Brimfox's response. Laravel would have to define that behavior at the integration point, and its current recipe-explanation path does not call the verifier.

Turn the weak cases into evaluation cases#

The outer request is typed, but the nested dictionaries are broad. The algorithm converts an arbitrary configuration value into the support threshold. Constraining that value at the request boundary would make malformed input fail before calculation.

Extraction also converts selected values to strings before testing whether they are empty. An omitted field can therefore become the nonempty word None and count as another unit. This is visible in the source, not a runtime failure reproduced during this review.

The existing verifier test checks an unsafe allergy statement. An evaluation set still needs omitted fields, threshold boundaries, misleading overlap and mixed supported and unsupported units, as well as the HTTP route and product acceptance rule. The dated six-test report is historical evidence, not a fresh evaluation of those cases.