Route a request.
Choose the tool that fits, ask for a missing detail, or decide none fits.
EXAMPLE SETUP
“What's the weather in Seoul?”
HOBO/Preview
The honest decision model for apps and agents.
A clear choice.
Or back to you.
Give HOBO a question and a few options. It picks one, shows how sure it is, and hands uncertain decisions back to you.
A short Google form. We'll email when early access opens.
LESS GUESSWORK. MORE SAY.A LITTLE HELP, EVERY DAY
SMALL JOBS. A CLEAR NEXT STEP.
What could HOBO help with?
Check a reply.
See whether a reply answered the question or refused to answer.
EXAMPLE SETUP
“Did this reply answer the question?”
Check a claim.
Compare a claim with a passage. Does the evidence back it up?
EXAMPLE SETUP
“Does this passage support the claim?”
Illustrative setups, not model results. See real recorded decisions in the replay below.
A MOMENT WITH HOBO
A little HOBO.A little sunshine.
The official HOBO song. A little sunshine for your next decision.
Listen to the official HOBO songTHE CURRENT PREVIEW
HOBO, on real requests.
BFCL v4 live, 1,933 real tool-call requests, never trained on.
- Across all requestsRight answers
- 88.8%95% interval 87.3–90.2
- Across all requestsConfident mistakes
- 2.3%wrong, while 90%+ sure
- With a fitted handoff barDecisions handed back
- 11%to keep the rest 90% right
The handoff result uses a bar fitted on separate examples to aim for 90% right. The replay below uses a bar you choose. Neither guarantees accuracy on a new task.
Preview: it missed one test we registered before training it. See where HOBO struggles.
See how we testThis replay plays back HOBO's recorded decisions on 1,933 real tool-call requests from BFCL v4 live, which it never trained on. Over all of them, 88.8% of its answers are right, and a bar fitted to keep the rest 90% right hands back 11%. The scoreboard below fills in as the replay runs.
WATCH HOBO WORK
1,933 real requests.
One call each.
About this replay
HOBO's recorded decisions on requests it never trained on, replayed. It calls the tool, asks for what's missing, or hands it back to you.
HOBO's current requestOpen to follow the latest pick
ONE REQUEST AT A TIME Recorded replay
Scroll here and watch HOBO choose.
HOBO'S PICK
A little thought first.
This request's confidence
Your bar is marked on the track.
THE SCOREBOARD
Same requests. Two ways.
HOBO can ask first or hand it back. Read each row across.
| The request | Right tool | Wrong tool | Asks first | None fits | Hands back |
|---|
✓Right!OK, asking was better×Wrong↩Back to you
MISTAKES, SIDE BY SIDESame requests so far
HOBO
0
Always answers
0
Asking never waits on your bar. Confidence is not an accuracy guarantee.
REPLAY AT YOUR SELECTED BAR0% right of the 0% it decided
ROOM FOR A SECOND THOUGHT0 fewer mistakes than always answering
QUICK AND CHEAP
Small decisions should cost small change.
HOBO picks from your options instead of writing an answer, so there's nothing to pay for output.
- Answer time
- 0.13 sper answer on one GPU, typical request
- Input
- $0.05per million tokens, about 1¢ per 1,000 decisions
- Output
- $0it picks, it doesn't write
A LITTLE MORE UNDERSTANDING
Korean claims.
Tested, too.
Does the evidence support the claim? We put HOBO to the test in Korean, too.
Measured in Korean on claim checking: 81.9% on KLUE-NLI, a test we didn't write.
What this test tells usWHY HOBO EXISTS
Honest and fair.
That's the mission.
We're building HOBO to be an honest, unbiased decision machine. Clear about its confidence. Ready to hand back what it can't settle.
Evidence over prestige. A famous name, elite university, or impressive title shouldn't give a claim extra weight.
Less swayed by pressure. In the tested prompts, telling HOBO the user is sure of a wrong answer changes its pick 1.8% of the time, down from 7.9%.
That's the goal. This is still a preview, and we keep testing how close we are.
See how we testA WELL-KNOWN NAME
Here's the evidence.
A NAME YOU DON'T KNOW
Here's the evidence.
Same evidence.
Same standard.
Still learning. Out in the open. We publish how we test and track where our training data comes from.
The data story
A SMALL HELLO GOES A LONG WAY
Be here for
HOBO's first day.
Want early access? Sign up. We'll write back when it opens.
Get early access A short Google form. We'll email when early access opens. We use your answers only for early access.