Model Zero
--:--:--Cairo

Model ZeroIn research

A model that reasons like an applied statistician.

Model Zero is the language model we are building to power Misura: a specialist fine-tune of an open-weight model, taught the decisions that matter in applied statistics. Which method fits the data, which assumptions to check, and what to do when one fails.

Model Zero is a plan with its training material collected. No model has been trained yet, and we will not publish performance claims until there are evaluations to back them.

Talk to us about Model Zero See Misura, where it will run

Why build a model at all

General-purpose models are good at writing code and poor at statistical judgment. On StatQA, a benchmark of 11,623 hypothesis-testing tasks, the best general model scored 64.83%, and its most common mistake was choosing a method whose assumptions do not hold. StatQA, NeurIPS 2024

Misura already checks a model's work with code. Model Zero is meant to make those checks fire less often, and to leave us independent of any single model vendor for cost and behaviour.

Do the two treatment groups differ in recovery time?

A general-purpose model

Thinking…

Model Zero, the goal

    Thinking…

    Illustration of the idea, not a measured result.

    How it is made

    A specialist fine-tune, not pretraining from scratch. The methodology cards that constrain Misura at runtime are the same material that teaches the model. Click a step to stop on it.

    Methods. Misura’s 74 methodology cards are the curriculum: method choice, assumption checks, failure routing and cross-software equivalents.

    Click a stage to pause on it.

    One card, many lessons

    Each methodology card already says when to use a method, what could go wrong and how to recover. Every part of it becomes a different kind of worked example, so the model learns the reasoning around a method, not just its syntax.

    • When to useCorrect method, correct run“Outcome is continuous, so fit OLS.”
    • AlternativesWhy this test, not that one“Not logistic: the outcome is not binary.”
    • AssumptionsAssumption violated, then remedied“Residuals fan out, so use robust errors.”
    • Failure routingRecover from a failed run“Singular matrix: drop the collinear term.”
    • R · Stata · SPSS · SASSame analysis, five packages“lm(y ~ x) ↔ regress y x”
    • Out of scopeRejection and redirection“That claims causation. Here is what the data supports.”
    Planned: 74 cards, about 78,000 reasoning examples.

    Only what survives is kept

    A teacher model writes the examples, and a teacher can be wrong. So each one has to pass the same kind of checks Misura applies at runtime: the code runs, the method suits the stated assumptions, and the required diagnostics are there. What fails is thrown away.

    • example_01···
    • example_02···
    • example_03···
    • example_04···
    • example_05···
    • example_06···

    Illustration: real examples are checked the same way, then sampled by hand.

    What it will be taught

    Most of the data teaches reasoning: choosing, checking, repairing and refusing. A smaller share teaches how researchers write results up, and the rest teaches it to push back and to admit uncertainty. Published papers are only seed material. Every training example is newly written.

    • Correct method, correct run22,200
    • Assumption violated, then remedied14,800
    • Why this test, not that one14,800
    • Same analysis across five packages11,100
    • Multi-step prerequisite chains7,400
    • Rejection and redirection7,400
    • Results in journal register8,000
    • Interpretation narratives6,400
    • Reporting assumption checks5,600
    • Effect-size reporting4,000
    • Adversarial prompts with corrections5,000
    • Expressing uncertainty5,000
    About 112K synthetic examples in total. A plan, not a result.

    What it is meant to do

    • Method selection

      Picks the right test or model for the design and data, and says why not the alternatives.

    • Assumption handling

      Knows which diagnostics a method needs, reads them, and switches to a remedy such as robust errors or a non-parametric test.

    • Sequencing

      Runs prerequisites in order: reliability before scale scores, stationarity before ARIMA.

    • Failure routing

      When a run fails or a diagnostic trips, takes the next step the method prescribes.

    • Researcher-register writing

      Reports results the way journals do, with the statistic, the effect size and what it means.

    • Five-software equivalence

      Expresses the same analysis in Python, R, Stata, SPSS and SAS.

    • Refusal and redirection

      Declines inappropriate requests and proposes a valid alternative.

    • Calibrated uncertainty

      Says plainly when the data cannot support a conclusion.

    • Cites literature

      Justifies choices with published sources, not internal documentation.

    Where it stands

    The realistic target is a highly competent applied statistician, not a PhD. We will publish results, including the ones that disappoint, once there is something to measure.

    • 74 hash-pinned methodology cards with machine-readable profilesDone
    • About 370 openly licensed study seeds, each logged with licence, DOI and checksumDone
    • A deterministic verification layer in Misura that checks any model’s workDone
    • Example generation and validation pipelineNext
    • First fine-tuning runNext
    • Held-out evaluation against baselinesNext