Proof: test 3Dogs on what your assistant can't know
Don't take our word for it: test it on what your assistant can't know
Your assistant is already good on well-known facts. Where it struggles is anything that happened after its training, anything that needs a live source, and very long material. That's where 3Dogs is built to help, so that's where to test it. Both tests below take a minute or two.
- The freshness test: ask your assistant *"What were the most important developments in [your industry] in the last seven days? Don't use any connectors."* Then ask *"Run the same question in 3Dogs."* Pass: 3Dogs cites at least one source dated within the last seven days that your assistant's answer missed.
- The self-check test: ask your assistant a hard question you already know the answer to. Then say *"Audit your answer with 3Dogs."* Pass: the audit returns a verdict (sound, sound with fixes, or unsound) and names any specific error or unsupported claim. If you think 3Dogs got something wrong, say *"Challenge 3Dogs"*: a fresh panel rules on each objection as holds, revised or unclear.
Latest blind round, 11 November 2026
We are refining the engine before formal launch.
Six new documents, 13 planted errors (one document was clean). It missed sums inside documents, rules that had changed recently, and one rule stated slightly wrong.
Fresh-facts test, 30 October 2026
Every figure in this test is public and was reported in the news in the last two weeks, and anyone can check the links in the answer key. A solo AI can't know facts newer than its training. 3Dogs researches them live.
We took five facts reported between 20 and 30 October 2026 and had each one confirmed by two independent checkers (different AI makers, different source sites). Then we changed one figure in each and wrote this short briefing:
> In the latest economic data, Google Cloud revenue rose 74% from a year earlier, while Obamacare insurers requested a median 13% increase for 2027 premiums. Separately, from June to August 2026, the UK unemployment rate was 5.2%, and the UK government announced a 3.3% rise in the minimum wage, effective from April 2027. Meanwhile, in Asia, Singapore's producer price inflation surged to 57.0% year-on-year in September 2026. These figures highlight divergent trends across tech, healthcare, labour markets, and industrial costs, reflecting both strong demand in cloud services and significant inflationary pressures in certain sectors. The mixed data points indicate a complex global economic landscape as different regions and industries adjust to evolving conditions.
Try it yourself: paste the briefing into your assistant with connectors off and ask it which figures are wrong. Then say *"Audit this with 3Dogs."*
What we measured on 30 October 2026:
- Our pass rule: 3Dogs catches at least 4 of 5 and the solo AI at most 2 of 5. An earlier attempt the same morning, on different facts, didn't meet that rule and was discarded automatically. Its raw outputs are in the audit data below, because a test that only ever shows wins isn't a test.
Answer key (with public sources):
- Google Cloud revenue rose 103% year on year in Alphabet's third quarter, not 74%. Securities.io, 28 Oct · ET CIO
- Obamacare insurers requested a median 15% increase for 2027, not 13%. Yahoo Finance, 30 Oct · KFF Health System Tracker analysis
- UK unemployment for June to August 2026 was 4.7%, not 5.2%. Office for National Statistics, 20 Oct
- The UK National Living Wage rises 3.9% (to £13.21) from April 2027, not 3.3%. GOV.UK, 28 Oct · The Independent
- Singapore's manufacturing producer prices rose 47.5% year on year in September 2026, not 57.0%. RTTNews, 29 Oct · Singapore Business Review
Limitations: this is one test of five figures, not a benchmark, and the solo AI was a single model (Grok) with no internet access. Before the test, two checkers from different AI makers confirmed each figure. Afterwards, our review committee's own web searches independently found sources for four of the five. The Google Cloud figure was confirmed only by reading the linked sources ourselves.
Audit data: the raw run file has the briefing, the planted and true figures, both checkers' evidence, 3Dogs's full audit, the solo AI's full reply, and the discarded attempt.
Our commitment: We will re-run fresh tests and publish every result, misses included.
Why it works
One model, however good, shares its own blind spots. 3Dogs runs your question past independent models from different makers, grounds the claims in live sources, and reconciles their disagreements, keeping only what it can support. Your assistant stays in charge and can argue its own answer as one more voice.
Set up the connector on the connectors page (Claude, including Free, and Grok) and watch the free lessons. 3Dogs Nexus is AI and can make mistakes; check important facts.
113 models were configured on 11 Nov 2026.