Accuracy results

TruePoint Marketing Partners · Tests run October 12, 2026 · Claim Grounding 1.4.5 · Prospect Intelligence 1.8.6

We publish summary numbers, how we tested, and what our tools do not do. The test items themselves stay private so nobody (including us) can tune the tools to the published set. Results are from small, self-built tests. Read the limitations below before relying on them.

Claim Grounding: Deep verify

All 8 misses on the hard set came back as insufficient_evidence or mixed, never as a confident wrong verdict.

Hard set by category (live 1.4.5)

How we tested

• Deep verify checks each claim against six public sources: Wikipedia, Wikidata, Google News headlines, Europe PMC, FRED (Federal Reserve Economic Data) and SEC EDGAR.
• TruePoint wrote the claims and decided the correct answer for each one. Numbers and dates for FRED and SEC claims were checked against the primary data. Claims were deliberately hard: wrong dates, swapped names, outdated facts, negations, two-part claims and invented entities.
• A claim counts as correct when the verdict matches the expected answer. "Confidently wrong" means a wrong verdict with confidence of 0.7 or higher.
• The hard set was run against the live service after deployment. The 17-claim held-out set was written separately and never used while building fixes, so it is the fairest single number here.

Prospect Intelligence: Email lookup

How we tested

• 40 real people whose work emails are published as press or media contacts, plus 10 invented names at real company domains.
• An address counts as right only if it exactly matches the published address.
• Version 1.8.6 changed one rule after this test: an address that was only guessed from a company's pattern, on a domain whose mail server accepts any address (a "catch-all" domain), is now labeled verify_first instead of send_with_caution. The label counts above are scored under that rule.
• Published press contacts are easier to find than the average person, so expect a lower find rate on ordinary lists.

What these tools don't do

Claim Grounding

• It is not legal, medical, financial or compliance advice, and not a certification that something is true.
• It works in English only and checks public sources only. It cannot see paywalled, private or internal documents.
• It can and will return insufficient_evidence when public sources don't settle a claim. That is a feature, not a failure, but it means not every claim gets a yes or no.
• Very recent news may lag: it relies on news headlines, and sources take time to update.
• Verify claim and Check answer use Wikipedia only. Only Deep verify uses all six sources.
• It checks facts, not opinions, predictions or tone.

Prospect Intelligence

• It does not guarantee an email will be delivered or read. "valid" means the company's mail server accepted the address, not that a person reads it.
• It finds work emails at a company domain. It does not find personal or private emails, and it does not return phone numbers.
• Decision-makers are reported as unknown unless a public source names them. Nothing is invented.
• If a person has asked our email provider to stop processing their data, we return nothing for them.
• Company research uses public sources (such as Wikipedia and the company's homepage) and may be incomplete or out of date.

About these numbers

Contact

Questions, bug reports or a claim we got wrong: email truepointmarketingpartners@gmail.com or call (951) 234-7620.
Product page: truepointmarketingpartners.com/ai-agent-tools