A Helsinki company selling quantified “net impact” data on companies and funds to investors, corporates and auditors. Not a competitor to SINE in any market — but the closest thing the wiki holds to a control group: an organisation attacking the same coordination failure SINE attacks, with the opposite institutional answer (centralise and enclose, rather than federate and govern in common), and one that has independently arrived at the problem the EIC exists to solve.
Ingested 2026-09-10 from three self-published texts by CTO Juho Ojala, at Martin Pompéry’s request, read specifically through a CCTI lens.
Company
⟨Background: the facts in this section come from live web research on
2026-09-10, not from a file in sources/. They are not primary-source
traceable and should be re-verified before being used externally.⟩
- Upright Oy, business ID 2848353-4, Annankatu 27, Helsinki. Founded 2017.
- Founders: Annu Nieminen (CEO) and Juho Ojala (CTO) — Ojala is the author of all three ingested texts.
- ~€5.27M raised in total; last disclosed round a €5M seed in December 2022 led by Planet A Ventures, with First Fellow Partners (Risto Siilasmaa) and Lightrock. No later round found.
- Headcount not published; third-party estimates cluster around 40.
- B2B data/SaaS. Buyers: private equity, private credit, asset managers and owners, corporates doing CSRD double materiality, advisors, universities. Self-reported users include Nasdaq, PwC, the European Investment Bank, Pictet, EQT, APG, Wharton and EDHEC — vendor claims, unverified.
The method
Four dimensions (people, planet, society, knowledge) across 19 impact categories, resolved against a proprietary Universal Impact Structure of ~800 impact topics. A macromodel estimates the impact of essentially all products and services globally; a company model combines that with a firm’s revenue mix. Inputs: public statistical databases (World Bank, IMF, WHO, OECD, Eurostat, IPCC) plus an NLP/LLM pipeline over a stated 200M+ scientific articles.
The machine-learning line is called Leena: a BERT model trained on 30,000 manually labelled papers in 2018, and since June 2026 Leena-2, a post-trained 14B Mistral base model. See Leena-2 (2026) for how it is trained and why that matters.
What the three texts actually argue
- Evaluating the ability of LLMs to quantify consequences (2025) — frontier LLMs are locally plausible and globally incoherent on impact magnitudes; and they systematically flatter companies and minimise environmental harms, attributed to training corpora full of corporate sustainability reports.
- Leena-2 (2026) — since these questions have no answer key, train and solve for coherence instead of correctness. See Coherence as a truth substitute.
- Consequences as a practical alignment problem (2026) — 45 prompt-injection breaks against 25 models; the exploitable surface is facts, not instructions.
The commons reading — a clean negative result
A terminology sweep of 41 pages on 2026-09-10 found zero occurrences of: commons, common-pool resource, Ostrom, tragedy of the commons, free rider, collective action, coordination failure, externality, data commons, data sharing. Also zero for the cybernetics vocabulary (Ashby, Beer, Meadows, feedback loop, requisite variety, viable system) despite the structure being thoroughly cybernetic — the Leena-2 closing line is literally “steering models, and steering money”.
What is there instead is an enclosure trajectory:
| Year | Position |
|---|---|
| 2018 | ”Why we are opening our model to public exploring and crowdsourcing” |
| 2022 | Open impact-data platform; €5M raised “to make net impact data openly available for all” |
| 2024 | Open database on 35,000+ funds |
| 2026 | U.S. patent pending on the post-training technique; harness closed; model whitepaper behind a lead form (the v2 PDF link 404s) |
⟨Inference: the inputs are a knowledge commons — 200M+ scientific articles and six public statistical databases — and the outputs are enclosed. The one asset that is commons-shaped by its nature, the Universal Impact Structure (a taxonomy of ~800 impacts, precisely the kind of artefact a standards body would govern collectively), is proprietary. Read against Data Commons and PACT, Upright is the same diagnosis with the opposite prescription: Ostrom’s rejected first option (privatisation) applied to a knowledge commons. That contrast is the page’s main strategic value — it is the sharpest available statement of what SINE is not doing and why.⟩
⟨Inference, offered as the more generous reading: the Universal Impact Structure is best understood as a requisite-variety play — the implicit argument is that the economy cannot be steered because the controller (ESG metrics) has far less variety than the system it steers, and ~800 impact topics is an attempt to raise controller variety to match. Upright never says this. It is a better argument than the one they make, and the frame to use if SINE ever engages them.⟩
Which failure are they addressing
Coordination failure, not bargaining failure: nobody is withholding information in a negotiation, there is simply no shared measure, so capital cannot be directed. That places them on the PACT/iLEAP side of SINE’s split, not the CCTI side — which is why the genuine overlap with CCTI is at the level of primitives (attested provenance, epistemic-integrity claims) and not at the level of use cases. Worth being precise about before anyone pitches them one.
Strategic relevance to SINE
⟨Inference throughout this section.⟩
- Strongest external validation of WP-C’s premise found so far. A commercial system whose entire product is a claim about the integrity of a reasoning process, with no mechanism to make that claim checkable by a counterparty. That is the EIC’s market, described by someone who does not know the EIC exists.
- A candidate WP-C use case. Transitive coherence over a solved system of pairwise judgements is a mathematical property, provable in zero knowledge without revealing the judgements. Proprietary inputs, a skeptical paying audience, a crisp statement to prove. The CCTI proposal still had no evaluation use case selected as of the 2026-06-17 draft.
- A citable outside argument for WP-A. If agents are broken by attacker-supplied facts rather than instructions, provenance is the only defence — see the 2026 alignment post.
- A warning about the greenwashing loop relevant to PACT and iLEAP: corporate self-reporting → training corpus → models sold to assess corporate sustainability → more reporting. Their own finding, their own words, unnamed as a loop.
- A methodological precedent for AI Advantage: they invented a domain-specific metric because none existed, published a result unflattering to the frontier labs, and reported an unplanned generalisation. That is what M4’s “publishable either way” commitment looks like in practice.
Evidentiary health warning
None of this is peer-reviewed. All three texts are self-published, two with self-assigned report numbers (UPTR-202501, UPTR-202601). The benchmark is their own, the harness is closed, the whitepaper is gated. A search on 2026-09-10 found no independent academic evaluation or critical journalism on the net impact model — all third-party coverage located (CFA Institute 2019, IPE 2020, MIT Technology Review 2018, funding trade press) is descriptive or promotional. Treat the method claims as a vendor’s account of its own system.
Sources
sources/impact-measurement/2026-09-10-upright-evaluating-llms.md— full text, retrieved 2026-09-10sources/impact-measurement/2026-09-10-upright-leena-2.md— full text, retrieved 2026-09-10sources/impact-measurement/2026-09-10-upright-consequences-alignment-problem.md— full text, retrieved 2026-09-10- Company facts, the 80-post blog inventory, the 41-page terminology sweep and the third-party-coverage search are live web research of 2026-09-10 and are not filed in
sources/— marked ⟨Background⟩ above where used.