Model Works

lastmod 2026-10-29created 2026-10-2713 edits

We build domain world models.

Model Works models learn from recorded histories of how your systems, processes, assets, or cases change over time. Using the changing state, relevant conditions, actions where there are any, and observed outcomes, they build lightweight predictive models that forecast what happens next. They forecast trajectories and outcomes over useful prediction horizons and, where the data supports it, explore within validated bounds how different actions change the outcome.

John Wainwright (computer scientist)

• John Wainwright (computer scientist) - Wikipedia

Model Works demo pre-briefing

I searched the web for technical documentation, papers, and benchmarks on Mirinae Technologies and Model Works, but found none. model-works.ai also has only a one-line description, "domain world models". So this demo is effectively the first technical evidence we will see.

Pre-contract checks

• The weights become ours, but the runtime program is licensed, and we do not receive the tool for retraining (Workbench).
• So if the data distribution changes, we are likely to depend on the after-sales service (Model Care).
• This version omits Chapter V, "Discussion of application direction" (p.32~35, partner-specific content).

Background: what to expect going in

Two branches of world models

• Foundation world models: Video and physical-world generation and prediction models such as Google DeepMind Genie 3, NVIDIA Cosmos, and Meta V-JEPA 2. They are huge and general-purpose.
• Domain world models: The side Mirinae is promoting. Given the customer's structured time-series data and the condition of tens of millions of parameters or fewer, it is presumably not a video-generating model but a time-series model that takes actions as conditions and predicts state changes (something close to the Dreamer family of latent state models).
  • So the demo is more likely to be a dashboard showing What-if comparison curves than flashy video like Genie.

Four technical issues to judge by

1. Counterfactual prediction and confounding: This is the most important. Historical data contains only the outcomes of the actions actually taken. For example, if patients in worse condition received stronger treatment, the model may learn "the situation in which that treatment was given" rather than "the treatment effect". The deck's appendix (p.36) also admits this limitation itself.
2. Error accumulation in long-horizon prediction: Even if one-step prediction is accurate, rolling out multiple steps in sequence (rollout) accumulates error.
3. Uncertainty display: Check whether it confidently outputs predictions even for actions or conditions with almost no data.
4. Baseline: Check whether it is actually better than simple methods such as XGBoost or existing simulators. This is exactly PoC step 3, "agreeing on a performance baseline".

Material to skim in advance (in order of priority)

1. Ha & Schmidhuber, World Models (2018), worldmodels.github.io: 30 minutes is enough to grasp the concept -> See Recurrent World Models Facilitate Policy Evolution.
2. Janner et al., When to Trust Your Model: Model-Based Policy Optimization (MBPO, 2019): Covers issue 2 (rollout error accumulation).
3. Bica et al., Counterfactual Recurrent Network (ICLR 2020) or Melnychuk et al., Causal Transformer (ICML 2022): Covers issue 1 (confounding). The structure is nearly the same as the deck's medical case.
4. Hafner et al., DreamerV3 (2023): A representative example of a latent state world model.

Things to check

• Whether the data is real, public, or synthetic
• How many steps ahead the predictions are shown
• Whether confidence intervals are shown
• Whether there is a point of comparison (baseline)
• Whether the demo is live or recorded
• Whether the LLM makes the predictions or only explains them
• When collecting data, how do you collect data that exists visibly but has no means of being collected?
  • What if video data is collected?
    • For data collection using video -> from the perspective of data collection, where is the boundary between the two, "domain world model" ~ "foundation world model"?

Question

"방금 보여주신 행동별 결과 비교에서, 과거에 실제로 실행하지 않은 행동에 대한 예측이 맞다는 건 어떻게 검증하셨나요? 실제 결과와 맞춰본 사례가 있다면 기존 방식 대비 오차가 어느 정도였는지 듣고 싶습니다."

• A question that directly asks whether the deck's core value, "compare outcomes before acting" (p.7, 11, 13), holds up.
• Because it asks in terms of the conditions the deck appendix (p.36) set for itself ("substantive differences in past actions", "excluded if the data cannot support causality"), it does not sound aggressive.
• The answer leads directly into the PoC evaluation design (baseline agreement).

Even if they answer like the last warning sign, it is not very unusual for an early-stage company. In that case, following up with "Then how would you design that validation in the PoC?" leads to a substantive conversation.

JEPA

JEPA architecture

Documentations

Model Works Design Partner Deck

See also

• World Models
• ComStock - Data the U.S. Department of Energy provides to AI researchers
• HyperFlow AI

Favorite site

• Model Works — We build domain world models