A capability profile answers two different questions: what an offer appears to do, and what the available evidence actually establishes. Keeping those questions separate helps you choose an integration without treating a polished description as a guarantee.
Capifold preserves source manifests and distinguishes provider declarations, derived information and scoped checks. This article explains how to use those distinctions when you compare candidates. It is a method for evaluating evidence, not a certification of any provider.
Read a claim as a claim
Suppose a manifest says “processes invoices securely.” That is a provider statement. It may be useful for discovery, but it does not specify input formats, data location, access rights or a tested result. Each of those requires its own source or test.
A declared field can be precise: for example, a particular input format or a documented endpoint. Precision makes a claim easier to check. It does not by itself make the claim independently verified. Preserve the original field and source URL so another person can inspect the same statement.
Treat inferred capabilities as navigation aids
A description mentioning document extraction can help a search system classify an offer. That classification is an inference from text, not a new promise by the provider. It should help you decide which profile to open, not create a permission or pricing fact.
Consider two descriptions: one converts text into speech, the other extracts entities from documents. Both mention text. A useful search must distinguish the operation, and a useful profile must leave unsupported details blank. No classification should invent an output schema simply to make two products look comparable.
Ask what a check actually covered
A format check establishes whether a document follows the checks that were run. It does not establish whether the endpoint is available, the output is correct or the provider can handle your production load.
A functional test is more specific. To interpret one, look for the tested version or fingerprint, runner, inputs or test scope, outcome and timestamp. A successful test against a bundled reference implementation belongs to that implementation. It cannot be transferred to an unrelated imported offer.
This is why an unqualified “verified” badge is less useful than a readable test scope. If the manifest changes, reassess whether earlier evidence still applies. Version binding makes that question visible instead of allowing a historical result to silently endorse new content.
Unknown is a decision state
If your task requires no network access and the profile does not state network behaviour, there is insufficient information to establish the requirement. Treating unknown as “probably fine” would change the meaning of your filter.
You have several reasonable next steps: request clarification from the provider, review a more specific source, conduct an explicitly approved test, or consider another candidate. The correct next step depends on the task. The catalog should show the gap rather than quietly remove your requirement to produce more results.
Keep three dates apart
A source retrieval date tells you when a system last observed a source. A metadata assessment date tells you when the profile was derived. A functional test date tells you when particular behaviour was exercised. A recent retrieval does not turn old functional evidence into a new test.
In a review note, record each date that matters. Also record what you could not establish. This makes it possible to revisit a decision after an update without reconstructing it from a screenshot.
A comparison you can hand to another person
For each candidate, write down the task, mandatory requirements, inspected source and version, supporting fields, relevant checks and unresolved questions. Avoid a single overall score when the important differences are permission scope or missing facts.
Use MCP server profiles to inspect a candidate, then follow the selection checklist. Capifold's public connection guide explains how another agent can read the same discovery information. The MCP security overview also advises evaluating a server before running it; discovery alone is not execution approval.