Verivello drops a retrieved record before the model can cite the wrong company

The filter runs late. `const grounded = toolResults.filter(r => verifyMatch(query, r))` sits after the tool calls, so the register request is already paid for by the time a record is rejected ^[5][15]. A name search that returns dozens of candidates costs the same round trip whether one candidate survives or none ^[12].

Order matters inside `verifyMatch`. The identifier comparison is guarded and returns: if both the query and the record carry a `companyNumber`, the function returns the equality of those two numbers ^[4]. When the numbers differ, the record is dropped and the name comparison never runs ^[13]. Companies House issues that number ^[1].

The other branch admits a record when normalised names are equal, or when similarity is at least 0.92 and the incorporation years agree ^[4]. That is weaker than it looks. Because those two tests sit either side of an OR, exact normalised-name equality passes with no corroborating signal, so where several records collapse to the query string after normalisation, the filter drops none of them ^[14]. The post's own worked example is a search for "Smith Consulting" that might return any of dozens ^[12]. Whether those dozens survive normalisation is not something the snippet can answer: the author calls it illustrative shape, not production code, and does not include the implementations of `normalise()` or `similarity()` ^[18].

Of the two grounding rules, only one is enforced by the interpreter. Tool output goes into context as verbatim JSON and the system prompt instructs the model to answer only from the provided records and to cite them ^[3]. That is an instruction. The filter is a boolean ^[5]. Writing on dev.to, Verivello's builder gives the reason the ordering is worth the code: feed an unverified name-search hit into context and "Now it's confidently wrong with a citation, which is worse," the author wrote ^[7].

Refusal follows from the same ordering. With no verified record, the agent says it does not know instead of offering a plausible guess ^[6]. The author treats a wrong director or ownership figure in due diligence as worse than no answer at all ^[16]. It is also the only branch with a fixed expected output.

For this pattern to transfer, two fields have to exist in your domain: an identifier issued by whoever is authoritative, present on both the query and the record, and a second attribute like an incorporation year that can corroborate a fuzzy name match ^[4]. UK company registers supply both ^[1], while a customer table keyed on names typed in by hand supplies neither and inherits only the 0.92 branch.

Each register sits behind its own typed Model Context Protocol server, independently unit-testable, and the same servers are reused across Claude Code, Claude Desktop and the product ^[9]. That prices the transport layer per source. The author open-sourced a small keyless example exposing UK public-data tools, with unit tests and CI ^[10].