Self-published technical report (UPTR-202601) by Juho Ojala, CTO and
co-founder of Upright Oy,
15 June 2026. Source:
sources/impact-measurement/2026-09-10-upright-leena-2.md.
U.S. patent pending on the post-training technique described.
Announces Leena-2, Upright’s vertically-trained LLM for impact-size inference, successor to the 2018 BERT-based “Leena-Leena”. Base model: a 14-billion-parameter open-weights dense model from Mistral AI, chosen after screening for dense models in the 10–20B range.
⟨Inference: of the three Upright texts this is the one to read closely. It is the clearest external statement of the problem the EIC exists to solve, arrived at independently and from the opposite direction.⟩
The problem as stated
Impact-size comparisons “have no ground-truth answers”. They bundle counterfactual reasoning, causal attribution and value systems, so “no one has a general way to define impact, or its size, that holds across everything”. Per the earlier CoCoCo evaluation, general-purpose models are “locally plausible and globally incoherent” — their magnitude judgements fail transitivity and flip with framing. The report insists this is “structural, not a fixable prompt issue”: nothing in pre-training forces thousands of separate magnitude judgements to agree with each other.
The mechanism: two multiplicative sources of coherence
1. Coherence trained in. A training loop:
- A strong model, with web search and instructed to ground answers in scientific research, runs a large number of pairwise impact-size comparisons.
- Those answers are mutually inconsistent, so the system “mathematically solve[s] for the implied impact sizes: the single set of sizes most coherent with all of the pairwise answers at once”.
- Those implied sizes become the “correct answers” generating training data.
- The model is post-trained on that data; the loop repeats.
In the report’s own words, “the model is pulled toward the most coherent version of its own research-grounded judgments.”
2. Coherence computed at inference. Published estimates are never single outputs. Per product, the production harness runs thousands of pairwise comparisons — within the product’s own impacts and across products — then solves the whole system for one coherent set of sizes. Each product is scored against the Universal Impact Structure, a proprietary hierarchy of ~800 impacts.
The two are called multiplicative rather than additive because (2) is only affordable at that inference volume with a small vertically-trained model, and only converges if the model’s comparisons are already coherent enough.
Results
- Significant improvement over the base model on held-out transitive coherence.
- Unplanned generalisation: the post-trained model consistently beats its own base model on a hardened version of the Allen Institute for AI Fermi estimation benchmark, “despite never being trained on Fermi problems”. Working explanation offered: coherent impact comparison exercises decomposition, unit sanity, counterfactual discipline and uncertainty handling, which transfer.
- Described as “early results from modest training runs”.
The alignment claim in the closing section
Citing Anthropic’s Claude constitution (January) — that aligned AI cannot be built from rule-following alone — the report argues that “a model that cannot coherently size what happens cannot reliably judge what to do”, and that the same machinery serves two markets: “steering models, and steering money”.
⟨Inference: steering is the literal etymology of cybernetics (kybernētēs, the helmsman) and the training loop is a textbook second-order control loop — a system observing its own output and correcting on it. Upright uses neither the word nor the tradition; the terminology sweep of 2026-09-10 found no hit for Ashby, Beer, Meadows, feedback loop, requisite variety or viable system anywhere on the site.⟩
The structural weakness
⟨Inference, and the most important reservation on this page: the loop has no exogenous error signal. It is closed on the model’s own output. The report says ground truth exists but is sparse, and that the model is “anchored by real data and science-grounded research where they exist” — but nowhere quantifies the weight or coverage of that anchor. A control loop closed on its own output converges to a fixed point of its own priors, not to the world; whether Leena-2 is converging on the structure of reality or on the structure of its base model’s biases is not answerable from this text. Goodhart is never named.
This is also where the complementarity with CCTI sits. Upright has reasoning without an anchor; WP-A’s attested-data layer produces an anchor without reasoning. Attested measurement is precisely the exogenous signal such a loop lacks.⟩
Do not cite
The “Meaning & Joy” figures (pizza at 34.6 ¢ of sensory pleasure per dollar of revenue, gardening lessons at 172.8, and so on) are labelled in the source itself as coming from experimental harness runs, not released platform scores, pending a validation pass — offered only as “evidence that the machinery has range”.
Related
- Upright Project
- Coherence as a truth substitute — the concept, in full
- Epistemic Integrity Certificate — the primitive this material argues for from outside
- Evaluating LLMs (2025) · Consequences as an alignment problem (2026)
- AI Advantage
- CCTI — where this material is read as a candidate WP-C evaluation use case ⟨link added by the 2026-09-14 lint; the citation was one-directional⟩