misata

Most synthetic-data tools learn from a real dataset and imitate it. Misata works the other way: you declare the outcome you want: "monthly revenue rises from \$50k to \$200k," "fraud is 3% in Q1 rising to 8% by Q4," "every customer's `total_spent` equals the sum of their orders", and Misata generates individual rows whose aggregates hit those targets exactly, with full referential integrity, from no source data at all.

This is *outcome-conformant generation*. The mechanism is formalised in an arXiv preprint (2606.08736): a closed-form method that satisfies declared aggregates to \$0.00 error, where off-the-shelf imitation synthesisers trained on the same data miss by 74–86%. Every run can also emit an Oracle report, a proof bundle covering referential integrity, constraints, temporal consistency, and reproducibility.

It generates from a plain-English description, a YAML schema, or an existing database schema. No machine-learning model is required. No real data is needed.

Built for:

• Known-answer testing: declare the KPI, generate the data, then assert your dbt, Spark, or SQL transform returns exactly that number. A pipeline test with a ground truth, before any real data exists
• Database seeding: fill dev and staging environments with production-like data
• Integration tests: relational fixtures with FK integrity across every table
• Demos and prototypes: realistic numbers, names, and distributions, no PII
• BI and dashboard development: data shaped like your real domain before launch
• Statistical method validation: longitudinal, grouped, and multi-site datasets that pass mixed-effects models, ICC tests, and autocorrelation checks

Declare or mimic: two ways in

Misata works in two modes, and the difference is the whole point:

• Declare (the default, no data required). You state the schema and the outcomes you want, exact revenue curves, fraud rates, rollups, constraints, and Misata generates rows from scratch that conform to them. Use this when you do not have real data, or when you need a *known answer* to test a pipeline, dashboard, or demo against.
• Mimic (when you already have data). Point `misata.mimic()` at a real CSV and get a synthetic twin that matches its distributions and correlations but contains none of the original rows, with `fidelity_report` and `privacy_report` to measure the result. Use this for privacy-safe copies of data you already hold.

Most synthetic-data tools only do the second, learning from a real dataset and imitating it. Misata leads with the first: you declare the answer, then generate the data around it.

Research

Misata's exact-aggregate engine is backed by an arXiv preprint:

> Declarative Outcome-Conformant Synthesis: Exact, Closed-Form Specification Satisfaction and a Conformance Benchmark

> Muhammed Rasin, arXiv:2606.08736 (2026)

> https://arxiv.org/abs/2606.08736v1

The paper formalises the core claim: when you declare `"SaaS MRR from $50k in January to $200k in December"`, Misata generates individual transactions whose monthly totals match the declared curve to exactly $0.00 error, not approximately, but provably, via a closed-form Gamma conditional-sum mechanism (Lukacs' characterisation). Off-the-shelf imitation synthesisers trained on the very same data miss the declared monthly aggregate by 74–86%; Misata reaches exactly 0.

The paper also introduces SpecBench: the first benchmark measuring conformance to analytical outcomes for cold-start relational synthesis. Misata is the reference implementation.

@article{rasin2026declarative,
  title   = {Declarative Outcome-Conformant Synthesis: Exact, Closed-Form
             Specification Satisfaction and a Conformance Benchmark},
  author  = {Rasin, Muhammed},
  year    = {2026},
  url     = {https://arxiv.org/abs/2606.08736v1}
}

Install

pip install misata

Optional extras:

pip install "misata[llm]"        # multi-provider LLM schema generation
pip install "misata[documents]"  # PDF output via weasyprint
pip install "misata[advanced]"   # SDV/CTGAN statistical synthesis
pip install "misata[mcp]"        # MCP server, expose Misata to Claude, Cursor, and other AI agents
pip install "misata[evalpack]"   # evalpacks: verified eval databases for data agents (DuckDB)

Use it from a coding agent

Misata ships an Agent Skill, so Claude Code and anything else that reads

`SKILL.md` knows which entry point fits which request and what is worth

declaring:

/plugin marketplace add rasinmuhammed/misata
/plugin install misata@misata

The skill drives the CLI, so `pip install misata` is still required. There is

also an MCP server (`pip install "misata[mcp]"`) and a Claude Desktop extension

in `mcpb/`.

Use Misata from Claude / Cursor / Windsurf (MCP)

Misata ships a built-in Model Context Protocol server with a clear division of labour: the AI agent designs the schema, Misata guarantees the math. Agents are good at knowing that a veterinary clinic needs a `species` column; Misata is good at making 50 000 rows where every foreign key resolves, every roll-up reconciles to the cent, and the same seed reproduces byte-identical output. The primary tool, `generate_from_schema`, accepts the agent's schema dict and returns the data plus an integrity proof: per-relationship orphan counts the agent can show you.

1. Install:

pip install "misata[mcp]"

2. Add to Claude Desktop (`~/Library/Application Support/Claude/claude_desktop_config.json`):

{
  "mcpServers": {
    "misata": {
      "command": "misata-mcp"
    }
  }
}

Restart Claude Desktop. Then just ask:

> *"Generate a fintech dataset with 1 000 customers, payments, and a 2% fraud rate."*

> *"Design a clinical-trials database (sites, patients, visits, adverse events) and generate 100k rows."*

> *"I need SaaS data: MRR from $50k in January, doubled by December, with a Q3 slump."*

The agent designs whatever tables the request needs (any domain; it isn't limited to Misata's built-ins), calls Misata, writes CSVs to disk, and reports back with previews and the verified integrity summary. See the MCP guide for Cursor/Windsurf/Zed setup and all six available tools.

mcp-name: io.github.rasinmuhammed/misata

Quick start

misata generate \
  --story "Brazilian fintech with R$ payments, CPF verification, and 3% fraud" \
  --rows 1000 \
  --output-dir ./demo_data

# Writes CSVs plus:
# ./demo_data/oracle_report.json

import misata

# One sentence → multi-table DataFrame dict
tables = misata.generate("A SaaS company with 5k users, monthly subscriptions, and 20% churn")

print(tables["users"].head())
print(tables["subscriptions"].head())

# Or from the CLI
misata generate --story "A SaaS company with 5k users and 20% churn" --rows 5000

Misata Oracle

The Oracle report is Misata's proof layer. It separates hard guarantees from advisory realism checks so generated data can be trusted in CI, demos, notebooks, and research comparisons.

Guaranteed checks:

• referential integrity across configured relationships
• requested row-count fulfillment
• schema validation and configured constraints
• deterministic reproducibility when a seed is set

Advisory checks:

• quality score and plausibility warnings
• privacy heuristics
• schema-vs-output fidelity score
• locale/domain fit for countries, cities, phone prefixes, and national IDs
• data-card metadata

import misata

schema = misata.parse("Brazilian fintech with CPF verification", rows=1000)
tables = misata.generate_from_schema(schema)
oracle = misata.build_oracle_report(tables, schema, seed=schema.seed)

print(oracle["passed"])
print(oracle["advisory"]["locale_domain_fit"]["locale"])

Mimic mode: clone any CSV in one call

Point `misata.mimic()` at a real dataset and get a synthetic twin that matches every column's distributions but contains none of the original rows. No schema authoring, no config.

import pandas as pd
import misata

real = pd.read_csv("titanic.csv")
twin = misata.mimic(real, rows=2000, seed=42, table_name="passengers")["passengers"]

The profiler handles the columns that break other tools:

• Alphanumeric code columns (Ticket `"A/5 21171"`, Cabin `"C85"`, SKUs, reference numbers) are detected by their character-class shape and reproduced structurally, same shapes in the right proportions, entirely new values, zero verbatim leak from the source. They no longer fall through to prose text generation.
• Floats keep their cents. A Fare of `7.25` generates as `7.25`-shaped values. The profiler infers decimal places from the data; semantic quantization (charm pricing) never fires on mimicked columns.
• Distributions are fit from the data. Skewed-positive columns get lognormal; constant columns get a uniform stub; everything else gets normal. Categorical columns with fewer than 50 values carry their real frequencies.

# Verify: no verbatim rows can leak through
shared = [c for c in real.columns if c in twin.columns]
overlap = pd.merge(real[shared].astype(str), twin[shared].astype(str), how="inner")
assert len(overlap) == 0

Eight ways to generate data

1. Plain English, no config required

tables = misata.generate("A fintech startup with 10k customers, fraud rate 3%, and IBAN accounts")

Misata reads the story, infers domain (fintech), scale (10 000 rows), and column semantics (fraud flag, IBAN format), no schema authoring needed.

A sentence is read by a recogniser that handles a fixed set of phrasings. Anything it cannot turn into a declaration is named in a warning rather than dropped, so you always know what took effect. When a prompt needs to be exact, write it as a spec instead.

1b. A structured spec, parsed deterministically

Prose that declares tables, row counts, columns and rules is not a story, and guessing at it is the wrong tool. Misata detects that shape and parses it directly, so "exactly 4000" means exactly 4000. No model, no inference, no rewriting.

Table 1: accounts
Rows: exactly 600
Columns:
  account_id
  company_name
  plan
  seats
  signed_up_on
plan must only be:
Starter
Professional
Enterprise
seats must be 1 to 120
signed_up_on must be 2023-01-01 to 2023-12-31

Table 2: invoices
Rows: exactly 3200
Columns:
  invoice_id
  account_id
  amount
  issued_on
account_id must match values from accounts table
amount must be 120 to 8500
issued_on must be 2024-01-01 to 2024-12-31

Revenue curve on invoices.amount by issued_on:
Jan 180000
Feb 195000
Mar 210000

tables = misata.generate_from_schema(misata.parse(open("spec.txt").read()))

What the spec guarantees:

Rules can sit inside a table's block or in one section at the end; either way they attach by column name. A column ending `_on`, `_at`, `_date` or `_for` is generated as a date. Anything the parser cannot translate is listed back to you, never guessed at.

2. YAML schema-as-code, commit it to git

misata init           # scaffolds misata.yaml in the current directory
misata generate       # reads misata.yaml automatically

# misata.yaml
name: my-app
seed: 42

tables:
  users:
    rows: 1000
    columns:
      user_id: { type: int, unique: true }
      email:   { type: text, text_type: email }
      plan:    { type: categorical, choices: [free, pro, enterprise] }

  orders:
    rows: 5000
    columns:
      order_id: { type: int, unique: true }
      user_id:  { type: foreign_key }
      amount:   { type: float, min: 5.0, max: 500.0 }

relationships:
  - "users.user_id → orders.user_id"

constraints:
  - name: amount_above_cost
    table: orders
    type: inequality
    column_a: amount
    operator: ">"
    column_b: cost

schema = misata.load_yaml_schema("misata.yaml")
tables = misata.generate_from_schema(schema)

3. Seed an existing database directly

from misata import schema_from_db, generate_from_schema, seed_database

# Introspect the live schema: no manual column definitions
schema = schema_from_db("postgresql://user:pass@localhost/myapp")
tables = generate_from_schema(schema)

# Seed it back: insert order respects FK dependencies automatically
report = seed_database(tables, "postgresql://user:pass@localhost/myapp_dev")
# SeedReport: seeded 6 tables, 47,300 rows in 1.2s

# One-command workflow
misata init --db postgresql://user:pass@localhost/myapp   # writes misata.yaml
misata generate --db-url postgresql://user:pass@localhost/myapp_dev --db-create

SQLAlchemy models are supported too:

from misata import seed_from_sqlalchemy_models
from myapp.models import Base

report = seed_from_sqlalchemy_models(Base, db_url="sqlite:///test.db", row_count=500, create_tables=True)

4. From a dbt project's own schema.yml

cd my-dbt-project && misata dbt-seed

No story, no config. Misata reads the properties YAML your project already

has and generates seed CSVs that satisfy it: `relationships` tests become

foreign keys with guaranteed integrity, `accepted_values` become the exact

category pools, `unique` and `not_null` become hard constraints, and

`data_type` plus column-name semantics decide the rest. Then:

dbt build   # seed + run + test — the tests you already wrote, passing on day zero

Both the legacy inline test syntax and the dbt 1.9+ `arguments:` nesting are

understood. Tests Misata can't translate (`dbt_utils.*`, custom generics) are

listed in the output rather than silently guessed at.

5. From a Prisma schema

cd my-app && misata prisma-seed

Reads the schema.prisma your app already maintains: `@relation` becomes foreign

keys with zero orphans, enums become the exact value pools, `@id` and `@unique`

are honoured, `@@id`/`@@unique` become composite uniqueness, and optional fields

may be null. CSVs land in `seed-data/` ready for your seed script.

6. Python dict schema

schema = misata.from_dict_schema({
    "customers": {
        "id":    {"type": "integer", "primary_key": True},
        "email": {"type": "email"},
        "plan":  {"type": "string", "enum": ["free", "pro", "enterprise"]},
    },
    "orders": {
        "id":          {"type": "integer", "primary_key": True},
        "customer_id": {"type": "integer", "foreign_key": {"table": "customers", "column": "id"}},
        "amount":      {"type": "float", "min": 1.0, "max": 999.0},
        "order_date":  {"type": "date"},
    },
}, row_count=5_000)

tables = misata.generate_from_schema(schema)

Declared outcome curves: add `__outcome_curves__` as a top-level key alongside the table definitions. Generated rows sum to every declared target exactly, to the cent:

import pandas as pd

schema = misata.from_dict_schema({
    "__outcome_curves__": [{
        "table": "orders",
        "column": "amount",
        "time_column": "order_date",
        "time_unit": "month",
        "value_mode": "absolute",
        "start_date": "2024-01-01",
        "avg_transaction_value": 120.0,
        "curve_points": [
            {"month": 1,  "target_value":  50_000.0},
            {"month": 6,  "target_value": 110_000.0},
            {"month": 12, "target_value": 200_000.0},
        ],
    }],
    "orders": {
        "__rows__": 5000,
        "order_id":   {"type": "integer", "primary_key": True},
        "amount":     {"type": "float", "min": 5, "max": 500},
        "order_date": {"type": "date"},
    },
}, seed=42)

tables = misata.generate_from_schema(schema)
monthly = (
    tables["orders"]
    .assign(m=pd.to_datetime(tables["orders"]["order_date"]).dt.month)
    .groupby("m")["amount"].sum()
)
assert abs(monthly[1]  -  50_000) =", "column_b": "enroll_date", "action": "cap"},
        ],
        "__correlations__": [
            # heavier patients tend to have higher blood pressure (r = 0.41)
            {"col_a": "bmi", "col_b": "systolic_bp", "r": 0.41},
        ],
        "patient_id":  {"type": "integer", "primary_key": True},
        "enroll_date": {"type": "date"},
        "visit_date":  {"type": "date"},
        "bmi":         {"type": "float", "min": 16, "max": 55},
        "systolic_bp": {"type": "float", "min": 90, "max": 200},
    },
})

`__rate_curves__` works the same way for per-period rate targets on boolean or categorical columns (fraud rates, churn flags, plan distributions).

7. LLM-assisted generation, richer semantics, optional

from misata import LLMSchemaGenerator

gen = LLMSchemaGenerator(provider="groq", model="llama-3.3-70b-versatile")  # free tier, fast & reliable
# gen = LLMSchemaGenerator(provider="anthropic")   # Claude
# gen = LLMSchemaGenerator(provider="ollama", model="llama3")  # fully local, no API key

schema = gen.generate_from_story(
    "A fraud detection dataset, 2% positive rate, FICO scores, transaction velocity features"
)
tables = misata.generate_from_schema(schema)

Requires `pip install "misata[llm]"` plus one of `GROQ_API_KEY`, `OPENAI_API_KEY`, `ANTHROPIC_API_KEY`, `GOOGLE_API_KEY`.

> Groq model tip: `llama-3.3-70b-versatile` is the reliable free-tier default. Larger models (e.g. `openai/gpt-oss-120b`) can return `413 Request too large` on Groq's free tier, so use them only on a paid tier. Whatever the model returns, generation never crashes on an imperfect schema: missing relationships, malformed probabilities, and out-of-range `time_unit`s are repaired automatically.

8. Incremental generation, grow a dataset without re-seeding

tables = misata.generate("A fintech company with 1000 customers", seed=1)

# Add 1 000 more rows: IDs auto-offset, FK integrity maintained across both batches
tables = misata.generate_more(tables, schema, n=1000, seed=2)
print(len(tables["customers"]))  # 2000

Realism that survives inspection

Synthetic data rarely fails on the big numbers; it fails on the small tells a reviewer spots in five seconds. Misata kills each tell with a specific, deterministic mechanism. No LLM is involved; everything is seeded and reproducible.

tables = misata.generate("A hospital with 300 patients, doctors and appointments", seed=7)
# patients:     Tae-yang Ahn (Male) · Valentina Esposito (Female) · pooja.kapoor@icloud.com
# appointments: 2023-03-08 14:00:00 · 2022-07-21 09:15:00: 15-min grid, business hours, 2% weekends

The dataset grades itself

Every coherence class above is also a detector. `story_audit` checks a generated

dataset against the full invariant catalog: FK orphans, cross-table temporal

causality, roll-up agreement, status gating, count and percent bounds, rare-flag

base rates, age against birth date, and more. Nothing incoherent ships silently.

tables = misata.generate_from_schema(schema, verify=True)   # warns on any finding

report = misata.story_audit(tables, schema)                  # or audit explicitly
print(report.summary())    # "Coherence: clean" or a scored list of findings

Every evalpack manifest embeds this verdict alongside its DuckDB answer

certificate, so a pack asserts both that its answers are right and that the

data telling the story is internally coherent.

Reproducibility and stability

• Within a version, generation is deterministic. The same schema, seed, and

misata version produce byte-identical tables. Evalpack manifests record the

version, seed, and a SHA-256 of the spec for exactly this reason.

• Across versions, RNG streams may change when generation improves (they

did in 0.8.1.29 and 0.8.2). Declared outcomes still hold: aggregates,

rates, identities, and integrity survive any upgrade; the individual rows may

differ. Pin the version when you need bit-identical regeneration.

• The public API is the documented top-level surface (`misata.generate`,

`generate_from_schema`, `story_audit`, `coherence_audit`, `build_evalpack`,

the schema classes, and the builders). Underscore-prefixed modules and

functions may change without notice.

Where the library is expected to fail is documented honestly, boundary by

boundary, in LIMITATIONS.md. Every entry there started as a

reproduced defect or a deliberate design refusal.

Unknown domains: composed, not confabulated

The 18 built-in domains are templates. For everything else, Misata refuses to fake understanding, and refuses to give up. A compositional synthesizer derives structure from your sentence: plural noun phrases become tables, "80 beekeepers" binds a row count, and a small archetype lattice (person / asset / place / event / document) provides honest structural columns and foreign-key wiring.

tables = misata.generate(
    "A beekeeping cooperative with 12 apiaries, 80 beekeepers, hives, inspections and honey harvests"
)
# beekeepers:  beekeeper_id, first_name, last_name, email, joined_at, status
# inspections: inspection_id, beekeeper_id, apiary_id, hive_id, inspection_date, status
# → full FK integrity, profiled timestamps, Zipfian statuses: from one sentence, no LLM

What it will *not* do is invent domain semantics: unknown entities get structural columns (reference codes, statuses, dates) and the detection report says exactly that, pointing to the two upgrade paths, a schema dict, or an LLM. The same gate also prevents confabulation: a story that only weakly matches a built-in template (one incidental keyword) is composed from its own entities instead of being forced into the wrong template.

Capsules: teach Misata a domain once

A capsule is one shareable JSON file of domain vocabularies (the species, treatments, and model names a domain calls things) with provenance for every list. Intelligence is spent once, at creation; generation stays deterministic, offline, and free.

# Mine a capsule from example data you already have: no LLM, no key
misata capsule create --domain veterinary --from-csv ./samples/ -o vet.capsule.json
misata capsule show vet.capsule.json

# Vocabularies override built-in pools for matching columns
tables = misata.generate("a veterinary clinic with patients and visits",
                         capsule="vet.capsule.json")

Capsules can also be written by an LLM once and reviewed before use (`capsule_from_llm`, BYO key; Groq's free tier works), or written by hand: it's JSON. Because a capsule is a file, it's a community artifact. Share it via git, a gist, or HF datasets.

Localisation

Misata automatically detects the country context from your story and generates statistically accurate data for that locale, the right names, salary distributions, national ID formats, currencies, postcodes, and company naming conventions.

# Locale is detected automatically: no extra flag needed
tables = misata.generate("German SaaS company in Berlin with 2k enterprise customers")
# → names from de_DE Faker pool, salary ~ lognormal(μ=10.71, σ=0.5) ≈ €45k median,
#   postcodes are 5-digit, company names end in GmbH/AG/UG

tables = misata.generate("Brazilian fintech with R$ payments and CPF verification, 50k users")
# → pt_BR names, salary median ~BRL 33.6k, national IDs match CPF format ###.###.###-##

tables = misata.generate("Indian startup in Bangalore with ₹ salary bands and Aadhaar KYC")
# → hi_IN names, salary median ~₹350k/yr, national IDs match Aadhaar 12-digit format

Force or override a locale explicitly:

schema = misata.parse("An ecommerce store with 10k orders")
tables = misata.generate_from_schema(schema)  # defaults to en_US

# CLI
misata generate --story "Ecommerce store" --locale ja_JP

15 built-in locales

Each pack carries real salary distributions (median and lognormal priors), age distributions, top-ranked cities, phone-number prefixes, postcode patterns, company suffixes, and VAT rates, sourced from OECD, World Bank, ILO, and national statistics offices (2023–24 data).

# Inspect a locale pack directly
pack = misata.get_locale_pack("de_DE")
print(pack.salary_median)       # 45000
print(pack.currency_symbol)     # €
print(pack.top_cities[:3])      # ['Berlin', 'Hamburg', 'Munich']
print(pack.company_suffixes)    # ['GmbH', 'AG', 'UG', 'KG', 'e.K.']

# Auto-detect from a story
locale = misata.detect_locale("South Korean company in Seoul with KRW salaries")
# → "ko_KR"

Constraints

Enforce business rules that survive every row of generation:

from misata.constraints import (
    InequalityConstraint,   # price > cost on every row
    ColumnRangeConstraint,  # min_price ", "cost")
df = c.apply(df)

Constraints can also be declared in `misata.yaml`, they run at generation time, not as a post-processing step.

Cross-table roll-ups

Make parent summary columns reconcile with child rows, so the data survives a `GROUP BY ... JOIN`. A `customers.total_spent` column generated independently of that customer's actual orders is a giveaway that data is fake; a roll-up computes it from the real child rows.

schema = misata.from_dict_schema({
    "name": "shop",
    "tables": {
        "customers": {
            "rows": 500,
            "columns": {
                "customer_id": {"type": "int", "unique": True},
                # total_spent = sum(orders.amount) per customer
                "total_spent": {"type": "float", "rollup": {
                    "from_table": "orders", "fk": "customer_id",
                    "agg": "sum", "column": "amount"}},
                # completed_spend = sum(amount) where status == "completed"
                "completed_spend": {"type": "float", "rollup": {
                    "from_table": "orders", "fk": "customer_id", "agg": "sum",
                    "column": "amount", "where": {"status": "completed"}}},
            },
        },
        "orders": {
            "rows": 3000,
            "columns": {
                "order_id": {"type": "int", "unique": True},
                "customer_id": {"type": "foreign_key", "references": "customers.customer_id"},
                "amount": {"type": "float", "distribution": "lognormal", "mu": 4, "sigma": 0.5, "min": 1},
                "status": {"type": "categorical", "choices": ["completed", "cancelled", "pending"]},
            },
        },
    },
})
tables = misata.generate_from_schema(schema)
# tables["customers"]["total_spent"] reconciles exactly with the orders table.

Aggregations: `sum`, `count`, `mean`, `max`, `min`. When a parent column name explicitly names a child table (`num_orders`, `total_orders`), the roll-up is inferred automatically with no declaration. Roll-ups survive the `misata.yaml` round-trip and run at generation time.

Statistical realism: data that passes method validation

Most synthetic data tools generate rows independently. That works for database seeding and pipeline tests. It breaks the moment the data needs to pass a statistical method: an autocorrelation test on repeated measurements, a mixed-effects model checking whether groups differ, or an audit that catches values outside plausible bounds.

Misata 0.8.1.0 adds a suite of features that close this gap. All are declared in the same plain dict schema and are reachable from MCP agents, Studio, and direct Python callers.

Stratified distribution profiles: different distributions per subgroup

A realistic A/B test dataset does not draw all users from one conversion distribution. The control group looks different from the treatment group. Use `profiles` to declare this precisely on any column:

schema = misata.from_dict_schema({
    "users": {
        "__rows__": 5000,
        "user_id": {"type": "integer", "primary_key": True},
        "cohort": {
            "type": "string",
            "enum": ["control", "variant_a", "variant_b"],
            "probabilities": [0.50, 0.25, 0.25],
        },
        "session_duration": {
            "type": "float",
            "distribution": "lognormal",
            "mean": 180.0, "std": 90.0,  # fallback for unmatched rows
            "profiles": [
                {"when": "cohort == 'control'",   "distribution": "lognormal", "mean": 180.0, "std": 90.0},
                {"when": "cohort == 'variant_a'", "distribution": "lognormal", "mean": 240.0, "std": 100.0},
                {"when": "cohort == 'variant_b'", "distribution": "lognormal", "mean": 310.0, "std": 120.0},
            ],
        },
    }
})

The `when` expression is evaluated as a pandas query against already-generated columns in the same batch. Rows that match no profile get the column's top-level distribution. Profiles can reference any column generated before the current one in declaration order.

Informative missingness: MAR and MNAR

Real-world datasets have non-random missing values. Misata models both mechanisms:

Missing At Random (MAR): The probability of a value being missing depends on an observed column. High-spending users are more likely to skip the optional income field.

"annual_income": {
    "type": "float",
    "nullable": True,
    "missing_if": {
        "predictor": "total_spend",
        "relationship": "higher_increases_probability",
        "base_rate": 0.05,
        "max_rate": 0.40,
        "mechanism": "MAR",
    },
}

Missing Not At Random (MNAR): The probability of a value being missing depends on the value itself. Very low satisfaction scores are the ones most likely to go unreported.

"satisfaction_score": {
    "type": "float",
    "distribution": "normal", "mean": 7.5, "std": 1.8,
    "nullable": True,
    "missing_if": {
        "predictor": "satisfaction_score",   # references its own column
        "mechanism": "MNAR",
        "relationship": "lower_increases_probability",
        "base_rate": 0.02,
        "max_rate": 0.50,
    },
}

Conditional nulls (`null_when`): Null a column whenever a boolean expression is true.

"cancellation_reason": {
    "type": "string",
    "enum": ["price", "competitor", "unused", "other"],
    "nullable": True,
    "null_when": "churned == False",
}

Exact incidence control: precise rates, not statistical approximations

A `boolean` column with `probability: 0.03` gives approximately 3% True values across many runs. If you need the dataset to contain exactly 3% (auditable against its own spec) use `exact_incidence`:

"is_fraud": {
    "type": "boolean",
    "exact_incidence": {
        "mode": "exact",
        "rate": 0.03,   # exactly floor(n * 0.03) rows are True
    },
}

Per-segment exact rates work the same way:

"converted": {
    "type": "boolean",
    "exact_incidence": {
        "mode": "exact",
        "group_by": "cohort",
        "rates": {"control": 0.12, "variant_a": 0.18, "variant_b": 0.24},
    },
}

The difference between "approximately 3% fraud" and "exactly 3% fraud" is the difference between a dataset that passes an audit and one that does not.

Within-entity time-series autocorrelation: longitudinal data that passes statistical tests

Without autocorrelation, a longitudinal dataset (user sessions, IoT readings, financial time series) is statistically identical to a cross-sectional one. Every time-series test (Ljung-Box, Durbin-Watson, autocorrelation plot) will immediately detect that rows are independent and the data is synthetic.

The `time_series` spec re-writes a column to have real within-entity autocorrelation:

"daily_revenue": {
    "type": "float",
    "distribution": "lognormal", "mean": 8500.0, "std": 3000.0,
    "time_series": {
        "entity_id": "store_id",      # one process per store
        "order_by":  "day_number",
        "model":     "AR1",           # AR1 | LINEAR_TREND | RANDOM_WALK | MEAN_REVERSION
        "phi":       0.72,            # autocorrelation coefficient (0 = independent, 1 = random walk)
        "noise_std": 800.0,
        "trend": {
            "slope_mean": 45.0,       # average daily growth per store
            "slope_std":  12.0,       # per-store growth variability
        },
    },
}

Four models are available:

Per-entity anchored distributions: separating within-entity and between-entity variation

When a child table's column should be anchored to its parent entity's value, use a formula in `distribution.mean`:

"stores": {
    "__rows__": 50,
    "store_id": {"type": "integer", "primary_key": True},
    "baseline_daily_revenue": {"type": "float", "distribution": "lognormal", "mean": 8500.0, "std": 3000.0},
},
"daily_sales": {
    "__rows__": 18250,   # 50 stores × 365 days
    "record_id": {"type": "integer", "primary_key": True},
    "store_id":  {"type": "integer", "foreign_key": {"table": "stores", "column": "store_id"}},
    "revenue": {
        "type": "float",
        "distribution": "normal",
        "mean": {"formula": "@stores.baseline_daily_revenue"},  # anchored to each store's baseline
        "std": 800.0,                                           # day-to-day noise
    },
}

The engine resolves the FK for every row and draws from that entity's personalised distribution. Between-store variation comes from the spread of `baseline_daily_revenue`; within-store day-to-day noise is `std: 800`. Generating all rows from one shared distribution (as every column-independent generator does) collapses between-entity and within-entity variance into a single number and fails every random-effects test.

Hierarchical ICC cluster effects: group structure that survives statistical tests

When rows are grouped under parent entities (stores, regions, branches), observations within the same group tend to look more alike than observations across groups. This within-group homogeneity (the intraclass correlation coefficient (ICC)) is a defining feature of grouped data. Without it, all groups look statistically identical.

`__cluster_effect__` is declared on the parent table and applies per-entity random intercepts to columns in the child table:

"regions": {
    "__rows__": 8,
    "__cluster_effect__": {
        "affects_table": "stores",
        "affects_columns": {
            "avg_order_value": {
                "icc": 0.22,         # target intraclass correlation
                "sd_total": 45.0,    # sd_between = sqrt(0.22) * 45 ≈ 21
            },
            "conversion_rate": {
                "sd_between": 0.04,  # supply sd_between directly
            },
        },
    },
    "region_id": {"type": "integer", "primary_key": True},
    "name": {"type": "string", "enum": ["North", "South", "East", "West", "Central", "NW", "NE", "SE"]},
}

One random intercept is drawn per parent entity from N(0, sd_between) and added to every child row in that group. The marginal distribution across all rows is preserved. Typical ICC values: 0.05–0.20 for store-level retail metrics, 0.10–0.30 for educational outcomes across schools, 0.15–0.40 for branch-level banking metrics.

Full correlation matrix: declare the complete covariance structure at once

For tables with many correlated columns, the matrix syntax is cleaner than a list of pairs:

"__correlations__": {
    "matrix": {
        "columns": ["session_duration", "pages_viewed", "revenue", "satisfaction"],
        "values": {
            "session_duration": [1.00, 0.71, 0.55, 0.32],
            "pages_viewed":     [0.71, 1.00, 0.48, 0.28],
            "revenue":          [0.55, 0.48, 1.00, 0.41],
            "satisfaction":     [0.32, 0.28, 0.41, 1.00],
        }
    }
}

The matrix is expanded into pairwise pairs and enforced via Iman-Conover rank reordering, which hits declared Pearson r values while preserving each column's marginal distribution exactly. Pairwise list syntax still works unchanged.

State machine terminal states: process-correct categorical columns

Any column that represents an entity's position in a process (customer lifecycle stage, order fulfilment state, subscription status) should follow a Markov chain, not a flat probability. `__state_machine__` generates the correct terminal distribution:

"orders": {
    "__state_machine__": {
        "state_column": "status",
        "initial_state": "placed",
        "transitions": {
            "placed":     {"confirmed": 0.95, "cancelled": 0.05},
            "confirmed":  {"shipped": 0.92,   "cancelled": 0.08},
            "shipped":    {"delivered": 0.97, "returned": 0.03},
        },
    },
    ...
}

States with no outgoing transitions are terminal. The engine traverses the chain per row until a terminal state is reached. Declared transition probabilities are preserved in expectation. Works alongside exact incidence, profiles, correlations, and time series in the same table.

Data validation: catch out-of-bounds values before they reach your pipeline

After generation, validate against declared domain bounds before the data reaches a model or a dashboard:

tables = misata.generate_from_schema(schema)

report = misata.validate_domain(tables, domain="financial")
print(report.summary())
# Domain validation (financial): 0 errors, 0 warnings.

assert report.passed

Built-in ranges for `financial` / `fintech`: price ≥ 0, discount 0–1, rate –1 to 100, salary ≥ 0. Column matching is by substring on the lowercased column name, `"unit_price"` matches the `price` rule.

Add custom ranges via the `custom_ranges` dict for any column type. Declare `"__domain__": "financial"` in the dict schema to attach the domain to the `SchemaConfig` for downstream tooling.

Export

# Columnar / analytical
misata.to_parquet(tables, "data/")
misata.to_arrow(tables, "data/")          # Apache Arrow IPC; requires pip install pyarrow
misata.to_duckdb(tables, "data/dataset.duckdb")

# Row-oriented
misata.to_jsonl(tables, "data/")
misata.to_sql(tables, "data/", dialect="postgresql")   # CREATE TABLE + INSERT statements
                                                        # dialects: ansi, postgresql, mysql

Reproducible incremental rows

Generate additional rows that append cleanly to an existing dataset without ID collisions:

# Day 1: generate the base dataset
schema = misata.from_dict_schema({...}, seed=1)
base = misata.generate_from_schema(schema)
for name, df in base.items():
    df.to_csv(f"./data/{name}.csv", index=False)

# Day 2: generate only new rows, PKs offset above existing max
new_rows = misata.generate_diff(
    schema,
    existing_dir="./data/",
    new_rows={"customers": 200, "orders": 1500},
    output_dir="./data/delta/",   # optional: write delta CSVs
)

`generate_diff` reads existing CSVs to find the maximum PK per table and generates new rows with PKs offset above that maximum. Use for streaming pipelines, day-over-day test fixtures, and any workflow where you need to extend a dataset without regenerating it from scratch.

Databricks and Apache Spark

Generate realistic, referentially-correct test data straight into Delta Lake: no

production data required. The `misata.spark` module bridges Misata's pandas output to

Spark/Delta on Databricks (Free Edition or full), AWS Glue, EMR, or any PySpark 3.3+ cluster.

import misata
from misata import spark as mspark

schema = misata.from_dict_schema({
    "customers":   {"__rows__": 500,  "id": {"type": "integer", "primary_key": True},
                    "email": {"type": "email"}, "country": {"type": "string", "text_type": "country"}},
    "orders":      {"__rows__": 2000, "id": {"type": "integer", "primary_key": True},
                    "customer_id": {"type": "integer",
                                    "foreign_key": {"table": "customers", "column": "id"}},
                    "total": {"type": "float", "distribution": "lognormal", "mu": 4.5, "sigma": 0.9}},
})

# One call: generate all tables (FK integrity guaranteed) and write to Delta
result = mspark.generate_to_delta(schema, spark, catalog="dev", database="bronze", mode="overwrite")
print(result.summary())
#   ✅ customers (500 rows) → dev.bronze.customers
#   ✅ orders   (2,000 rows) → dev.bronze.orders

What it does that `dbldatagen` can't: multiple related tables in one call, guaranteed

referential integrity, realistic distributions, and *outcome conformance*, declare an exact

aggregate or rate (e.g. "fraud is 1.8% in Jan ramping to 4.1% by Jun") and the data conforms,

giving downstream pipeline tests a known ground truth to assert against.

On Databricks serverless / Free Edition, install plain `misata` (PySpark is already on the

cluster, installing `misata[spark]` would stop a serverless session). On other environments:

`pip install misata[spark]`.

End-to-end tutorial: a complete fraud-detection medallion pipeline (Bronze → Silver → Gold)

tested entirely on synthetic data, with a CI-grade ground-truth assertion,

`examples/databricks/`. Full API reference: `docs/spark.md`.

Document generation

Render one document per row from any table, useful for demo datasets that need to look real end-to-end:

# Built-in templates: invoice, patient_report, transaction_receipt, user_profile
paths = misata.generate_documents(
    tables, "invoice", table="orders", output_dir="/tmp/invoices", format="html"
)
# format="pdf" requires: pip install "misata[documents]"

# Custom Jinja2 template
tmpl = "Order #{{ order_id }}Amount: ${{ amount }}"
paths = misata.generate_documents(tables, tmpl, table="orders", output_dir="/tmp/custom")

Quality and privacy analysis

bundle = misata.analyze_generation(tables, schema)   # runs privacy, fidelity, data_card

print(bundle.fidelity.overall_score)     # 0–100 statistical fidelity score vs. schema intent
print(bundle.fidelity.grade)             # letter grade for the same score
print(bundle.privacy.overall_risk_score) # heuristic PII / re-identification risk
print(bundle.data_card.tables)           # per-table row counts and metadata

Evalpacks: eval databases where the answer key cannot be wrong

Published text-to-SQL benchmarks are built by annotating question/answer pairs on top of an existing database, and that annotation step is where pervasive answer-key errors creep in. An evalpack inverts the order: the ground truth is the declared spec itself (outcome curves, rate curves, FK relationships), Misata generates a database that satisfies it, and every question shipped in the pack is then verified by executing its gold SQL against the written CSV files with DuckDB, an engine that shares no code with the generator. Questions whose observed answer does not exactly match the declared answer are dropped and recorded in the manifest. A wrong answer key is impossible by construction and double-checked by independent execution.

pip install "misata[evalpack]"
misata evalpack --config misata.yaml -o ./my_pack --seed 42

from misata.evalpack import build_evalpack

result = build_evalpack(schema, "my_pack")
assert result.all_verified

Each pack ships the tables as CSVs, `questions.jsonl`, a per-question verification certificate, a manifest with the spec hash and seed, and a standalone `verify.py` that anyone can re-run with nothing but `duckdb` installed. Use it to benchmark SQL agents, RAG-over-database systems, or any tool that claims to answer questions about data, against a database whose right answers are known before a single row exists.

Supported domains

18 built-in domain schemas, each generates a fully relational, multi-table dataset with realistic distributions, FK integrity, and domain-appropriate column semantics.

No keyword match → the compositional synthesizer builds a structural multi-table schema from your sentence's own entities (see *Unknown domains* above); stories with no entities at all fall back to a generic single table with smart column inference.

How it works

story / YAML / dict / DB introspection / MCP tool call
              ↓
        StoryParser  ·  compositional synthesizer  ·  locale detection  ·  load_yaml_schema  ·  schema_from_db
              ↓
        DetectionReport  (domain, confidence, near_misses, table_preview, warnings)
              ↓
        SchemaConfig  ←  validate_schema() catches issues before any rows are generated
              ↓
        DataSimulator
          ├─ topological sort (FK dependency order)
          ├─ domain priors  →  locale priors (salary, age, monetary)
          ├─ constraint engine (inequality, range, ratio, sum, unique)
          ├─ outcome curves (monthly targets from narrative control points)
          ├─ stratified profiles (per-subgroup distributions, pandas eval)
          ├─ AR1 / time-series autocorrelation (per entity, 4 models)
          ├─ state machine (Markov terminal states)
          ├─ ICC cluster effects (per-parent-entity random intercepts)
          ├─ Iman-Conover correlation engine (pairwise + full matrix)
          ├─ MAR / MNAR missingness (predictor-scaled and value-dependent)
          ├─ exact incidence (floor(n × rate), per-group rates)
          ├─ realism core (joint identities, temporal profiles, Zipf marginals,
          │                geo facts, grammar microtext, numeric quantization)
          └─ RealisticTextGenerator (capsules + Faker locale + vocabulary assets)
              ↓
        {table_name: DataFrame}
              ↓
        validate_domain  ·  seed_database  ·  to_parquet  ·  to_arrow
        to_duckdb  ·  to_sql  ·  to_jsonl  ·  generate_documents  ·  MCP CSV output

Domain priors: monetary columns get log-normal distributions. Categoricals use Zipf sampling. Blood types, country distributions, and salary bands reflect real-world statistics.

Locale priors: salary and age distributions are overridden with country-specific lognormal/normal parameters sourced from national statistics. `"Brazilian fintech"` in your story means salaries are sampled from the BRL distribution, not the USD one.

Outcome curves: natural-language narrative is parsed into exact monthly control points. Named events, quarters, and multipliers all work:

# All of these produce precise, shaped outcome curves:
misata.generate("SaaS mrr from $50k in Jan to $200k in Dec, with a Q3 slump")
misata.generate("Ecommerce orders, Black Friday spike, Christmas peak")
misata.generate("SaaS startup, MRR 10x growth over the year")
misata.generate("Fintech payments, strong Q4, dip in Q1")

Realism rules: `cost` is always less than `price`. `delivered_at` is always after `shipped_at`. `hire_date` is after `date_of_birth` + 18 years and never in the future. `tenure_years` is derived on the same row from `hire_date`. Email addresses derive from first and last name columns, names agree with declared genders, route distances agree with their cities, and review text agrees with its star rating.

What makes Misata different

Comparison reflects each tool's documented, out-of-the-box behavior as of late 2025; all

of these are capable libraries built for different goals, and a "No" means "not a built-in

feature," not "impossible."

**Faker** generates individual fake values, not relational, no schema, no statistical accuracy.

Synth excels at schema-as-code git workflows; limited distribution control.

syda uses an LLM for every row, semantically rich but expensive, slow, and requires an API key.

**SDV** learns from real data, a different problem (you need real data first).

**Gretel** is a cloud service that needs an API key and sends data off-premise; Misata runs locally.

Misata generates from intent, offline by default, seeds databases directly, and now brings country-accurate statistics to every column automatically.

Full head-to-head comparisons: Misata vs Faker, Misata vs SDV, Misata vs Gretel.

Performance

Measured on Apple M-series (single core, no GPU):

Contributing

git clone https://github.com/rasinmuhammed/misata
cd misata
pip install -e ".[dev]"
pytest tests/

1,088 tests, 0 failures. Issues and PRs welcome, github.com/rasinmuhammed/misata/issues