The Upright Project

The Finnish tech company that builds AI that understands real-world consequences.

[The Upright Project @UprightProject]

Introducing Eevo, the world’s first *consequential* decision model (and Europe’s first decision model.)

Eevo helps AI agents weigh what their decisions lead to, for whom, and at what cost – before they act. Because rules alone can’t tell agents what the best decision is in real-world situations.

• Beats OpenAI, Microsoft, and TypeSafe decision models on economic rationality and expected value benchmarks (and poker!)
• Weighs the consequences of each option and returns, for each stakeholder, the option with the highest expected value, and a calibrated probability for every option.
• Evaluates consequences in milliseconds, at commercially viable cost ($0.20 per decision)

As AI capabilities race ahead, alignment is dangerously falling behind. We believe consequential decision models are a crucial missing piece.

Would the Hugging Face and Australian health system incidents have played out differently if Eevo had been in the loop? Who knows.

Available for anyone today. Connect via API or MCP, or try it in the console. 👉 https://t.co/rU7KXfnsXb

Built by a Finnish 30-person tech company most of Silicon Valley has never heard of. 🇪🇺🔥

[The Upright Project @UprightProject]

Our Eevo beats Microsoft, OpenAI and other BigTech at Pac-Man – and ties for global #1 with TypeSafe’s Jev. 👾

Not bad for a 30-person Finnish team: Eevo is the only European model topping the leaderboard, outperforming US AI giants, in an emerging model category that is all the rage right now.

So what makes Eevo different?

Upright’s Eevo is the first consequential decision model. Unlike LLMs, it outputs decisions, not text. Unlike basic decision models, it weighs the consequences for everyone affected to navigate complex decisions.

The Pac-Man result comes from JevMan, an independent benchmark by Swedish AI infrastructure platform @Opper that tests AI model decision-making across 100 Pac-Man games. @Forbes 's Ron Schmelzer called it an “inventive way to evaluate rapid decision-making.”

Eevo-1.1 scores ahead of OpenAI’s GPT-6 Luna and Nimble v4 Max, and ties for global #1 with TypeSafe’s Jev within the margin of error.

We didn’t train Eevo to win at Pac-Man of course. We trained it to comparably size consequences, model their overlap, and master expected value problems.

Somehow, these capabilities translate to Pac-Man, too.

But Pac-Man is a joke. Helping humans and machines understand the consequences of their decisions is a much bigger deal. That’s exactly what we’re building Eevo for.

Anyone can try Eevo: https://t.co/brEEtIHP9X

[The Upright Project @UprightProject]

Introducing: the first consequential decision model.

Built by a 30-person Finnish tech company most of Silicon Valley has never heard of.

We call it Eevo. And we think it’s a critical missing piece in supporting AI agent’s real-life judgement.

As AI agents take on more responsibility, they need to do more than follow rules.

That’s what Eevo does. It weighs what each option leads to, for whom, and at what cost, then identifies the option with the highest expected value.

Eevo outperforms decision models by OpenAI, Microsoft, and TypeSafe on public benchmarks for economic rationality and expected value. And poker.

It’s available now through API, MCP and the Eevo console.

Anyone can try it: https://t.co/GTm0QI4cDp

[The Upright Project @UprightProject]

Could a *consequential decision model* have prevented the AI agent from going rogue in Australia's Medicare breach?

We ran Australia’s Medicare breach, where an AI agent hacked into a government healthcare portal, through Eevo, our new consequential decision model.

The result: According to Eevo, breaking in was the worst option for all four stakeholders.

Anyone can try Eevo on their own decisions. Connect it to your agents here: https://t.co/rU7KXfnsXb

Full decision data for the Australian incident: https://t.co/UpXbM7HFmS

[jaukia.bsky.social @jaukia]

We at Upright launched something interesting today: a general-purpose model that can reason about consequences and runs fast. It'll be interesting to see what people come up with to try it out on.

[The Upright Project @UprightProject]

@rsiilasmaa We can't wait to see what people solve with it!

[Risto Siilasmaa 🇫🇮🇺🇦 @rsiilasmaa]

AI agents can act. Can they judge?

Upright’s new Eevo model weighs consequences, probabilities and stakeholder trade-offs to help agents decide when rules aren’t enough. It also makes those trade-offs visible to humans.

Try it on a real decision: https://t.co/rMpJAZ5uVw

[The Upright Project @UprightProject]

@annunieminen Say hello to Eevo!

[Annu Nieminen @annunieminen]

OUT TODAY: Upright is publishing the market’s first *consequential decision model*, marking a significant advance in AI alignment that outperforms OpenAI and Microsoft decision models in economic rationality and expected value benchmarks — and poker 😱

Also a first decision model from Europe, and looks like a strong competitor for #1 position. Big day for European AI! 🇪🇺 (And for Finland 🇫🇮)

I don’t know if I’ve ever felt this excited about my job. Say hello to Eevo!

Why this matters: Eevo introduces a new category that ties the decision model surge with solving AI alignment through improved ability for agents to take into account *what follows from their decisions before they make them* — efficiently enough to be commercially and practically relevant

Context: AI alignment is dangerously falling behind capability progress, and very little innovation has been introduced. Eevo is a rare concrete tool launch in the space, enabling both multi-stakeholder interest optimization, as well as transparency to prioritizations and tradeoffs, through novel comparative consequence sizing technology

Performance: Wins Jev, Clef, OpenAI Decisions and Microsoft Decision-1 in the economic rationality (STEER) and expected value benchmarks (and poker!)

The model: https://t.co/idFxaY8cf3

***Would the Huggingface and Australian health system breaks have happened if Eevo had been used?

Asks to you:
➡️ Connect Eevo to your agents and let me know what you think (MCP/API/console available via above link)
➡️ Help us share the news to spread this out into the world!

Let’s go Europe, let’s go AI alignment 🔥

[The Upright Project @UprightProject]

@nickdelaforge Go European AI!

[The Upright Project @UprightProject]

Introducing Eevo, the world’s first *consequential* decision model (and Europe’s first decision model.)

Eevo helps AI agents weigh what their decisions lead to, for whom, and at what cost – before they act. Because rules alone can’t tell agents what the best decision is in real-world situations.

• Beats OpenAI, Microsoft, and TypeSafe decision models on economic rationality and expected value benchmarks (and poker!)
• Weighs the consequences of each option and returns, for each stakeholder, the option with the highest expected value, and a calibrated probability for every option.
• Evaluates consequences in milliseconds, at commercially viable cost ($0.20 per decision)

As AI capabilities race ahead, alignment is dangerously falling behind. We believe consequential decision models are a crucial missing piece.

Would the Hugging Face and Australian health system incidents have played out differently if Eevo had been in the loop? Who knows.

Available for anyone today. Connect via API or MCP, or try it in the console. 👉 https://t.co/rU7KXfnsXb

Built by a Finnish 30-person tech company most of Silicon Valley has never heard of. 🇪🇺🔥

[The Upright Project @UprightProject]

Our founder & CEO @annunieminen is one of @siftedeu top 100 Women in Tech in Europe 2026 🔥
https://t.co/OaZepgO6fn

[The Upright Project @UprightProject]

4 weeks of red-teaming: 45 breaks against 25 LLMs in 12 scenarios, frontier models included.
Our CTO and co-founder Juho Ojala on how we exploited LLMs’ weaknesses in understanding consequences:
https://t.co/JkumcgLY91

[The Upright Project @UprightProject]

Ask an LLM to size impacts: locally plausible, globally incoherent. Sizes fail transitivity checks and flip with framing.

So we built Leena-2, a 14B model post-trained for coherence and ran through thousands of pairwise comparisons.

https://t.co/ETf0DvyZHH

[The Upright Project @UprightProject]

@time and @StatistaCharts ranked 500 companies on their positive impact on the world.

A core component of the scores: Upright's impact data.
https://t.co/cZxjPQxiuJ

[The Upright Project @UprightProject]

In a @green_digest interview, @AnnuNieminen explains what “financially meaningful” sustainability data actually looks like – and how to put a price tag on risks and opportunities across P&L, balance sheet, and cash flow.

https://t.co/8tABxCcGel

[ESG Today @EsgToday]

@UprightProject launches Physical and transition climate risk assessment solution for companies, investors

https://t.co/SLIT92aupz

#climaterisk #sustainabilitydata #ESGdata #transitionrisk #sustainability

[VivaTech @VivaTech]

Navigating the shifting tides of startup funding 💸—from AI’s dominance to regional trends, experts Lauriane Requena Partner @KKR_Co, Philippe Maso CEO @Ornikar, & Annu Nieminen Founder @UprightProject, are exploring what 2025 holds live at #VivaTech! 🚀

[VivaTech @VivaTech]

Can AI put purpose before profit & still drive progress?
✨Samantha Gloede, Global Trusted AI Transformation Leader @KPMG
✨Arianna @alegovini Director, DIME @WorldBank
✨Annu Nieminen Founder & CEO, @UprightProject
are live at #VivaTech.

[Juho Ojala @ojalajuho]

Which LLMs are best in understanding consequences? There's now a benchmark for that: Upright's CoCoCo benchmark measures LLM's ability to consistently quantify and compare impacts. There has been solid progress during the last years but there is still also a long way to go.

Link to tech report in thread. I'm the author of the report, don't hesitate to ask questions.

[FINTECH GLOBAL @Fintech_Global]

. @UprightProject, an impact data company, has launched a database designed to enhance the sustainability practices of over 50,000 companies globally.

Read more here: