MisuraPre-release
A statistician that shows its work.
Misura is a statistical analysis agent for researchers. Load a dataset, describe the question in English or Arabic, and Misura plans the analysis, runs it on your data, and reports only numbers that came from code that actually ran.
- Does tutoring raise exam scores, controlling for prior GPA?
Every number, computed.
A 21-second look at the idea behind Misura.
Read the narration
Every answer starts with a question. Ask it in plain language. Misura doesn't guess. It computes. One model. One number. Fixed as evidence, never rewritten. Then the paper is written around it. Misura: every number, computed.
From question to report, in seven steps
Follow one analysis through Misura. The yellow marker shows where it is; click any step to stop on it.
Ask. Load your data and ask in plain English or Arabic. Misura may ask a few clarifying questions.
Click a stage to pause on it.
Why you can trust the numbers
A language model can write plausible statistics. Misura does not rely on it to be right: the system, not the model, decides whether an analysis was done properly and whether a number may be printed. Four rules, each enforced in code.
A written method comes first
Every analysis must follow a methodology card the agent has read in full. 74 cards across 11 families cover the assumptions to check, the diagnostics to run, and what to do when one fails.
run_python(purpose="analysis")- Method card cited
- Card read in full (a search summary does not count)
The engine refuses the run, not the model's good intentions. Every number is traced
Statistics are captured from the fitted model objects themselves. A figure in the answer that cannot be matched to an executed result is replaced with “[not computed]”, never printed.
Executed results
- coefficient2.31
- p-value.004
- R².87
- sample size212
The model's draft answer
Each tutoring hour raised scores by 2.31 points (p = .004), explaining .87 of the variance across 212 students, with an effect size of d = 1.40.
Checking 2.31 … found in executed results
Illustration with made-up numbers. An audit stands before the report
Missing diagnostics are blocking findings. The formal report is held back until they are resolved, or explicitly waived by the researcher.
- Linearity
- Residual normality
- Heteroskedasticity
- Influential points
- Multicollinearity
Formal report: lockedBlocking findings hold the report back. Anyone can re-run it
Each run leaves a hashed, read-only record. Export a replication package whose verify script re-runs every analysis and checks each result against its recorded hash.
- run_01 · descriptives
a3f9…c21e— - run_02 · regression
7be0…44d9— - run_03 · robustness
19c4…e07a—
Hashes shown are examples. - run_01 · descriptives
What it does
Your data, your formats
CSV, Excel, SPSS, Stata, SAS, Parquet and JSON. Survey structure such as Likert blocks is detected on load.
Plain-language questions
Ask in English or Arabic. Misura asks a few clarifying questions, proposes a plan, and waits for your approval before it runs anything.
Sandboxed analysis
Generated Python runs in a constrained sandbox with no network access and no way to write outside its output directory.
Literature that exists
Searches scholarly sources and cites only what it actually retrieved, with references rebuilt from the retrieved records.
Formal reports
APA, Vancouver, Chicago, Harvard or IEEE, exported as PDF, DOCX, HTML or TeX, with right-to-left Arabic support.
Local-first desktop app
Datasets, the database and outputs stay on your machine. Model and literature calls go to remote services.
Built on research
Misura comes out of applied statistics research, and Model Zero publishes that work openly in The Null Hypothesis: essays on the mathematics, measurement and incentives behind data and AI.
Read the research