agent-paranoid-android

Agent Paranoid Android

Deterministic synthetic test data from CSV and database metadata, without copying source rows. Use the CLI or Python library to profile structure, review a DatasetSpec, generate reproducible datasets, and validate the result.

Stable 1.6.0 is in preparation and is not yet published. Until verified publication, stable 1.5.0 remains the published recommended release:

python3 -m pip install "agent-paranoid-android==1.5.0"

After verified publication: Stable 1.6.0 is the recommended release. The installation commands below apply only after that publication. This promotion retains the accepted RC5 runtime; RC5 acceptance and RC5 publication evidence remain historical records, not verification of the new stable commit. Read the documentation for full workflows.

See the source/output matrix for the common reviewed DatasetSpec, field types, settings and business-rule contract.

Install And Try It

Python 3.11 or newer is required.

After verified 1.6.0 publication, install the exact stable release:

python3 -m pip install "agent-paranoid-android==1.6.0"
test-data-agent demo --output out/demo

The offline demo uses a bundled fictional fixture and needs no checkout, network, database, or AI provider. A successful run reports:

Generated synthetic dataset: out/demo | rows: customers=12 | seed: 20260801 | validation: passed | source rows copied: no

The output contains synthetic CSV data, a safe profile, the effective DatasetSpec, a validation report, and a generation manifest. The demo has no relationship or business-rule evidence, so it makes no claim about preserving those properties.

Start With Your Data

CSV and JSON workflows are included in the base package. Install optional database, Parquet, MCP, or AI support only when you need it; see Installation.

Guarantees And Limits

• An explicit seed makes generation reproducible under the recorded package, dependency, locale, and serializer environment.
• Source rows are never copied into generated output. Exact source literals are replaced by default; a field-scoped allowlist can retain only reviewed, bounded, non-sensitive business enums or constants in approved local destinations.
• Database profiling is read-only, allowlisted, and resource-bounded. It does not expose arbitrary unrestricted SQL.
• AI providers are optional. They receive safe metadata, cannot approve a specification, and cannot generate data directly.
• Human review remains required for ambiguous identifiers, free text, inferred relationships, and organization-specific privacy policy.

The project does not certify statistical anonymity, protection from every re-identification attack, or cross-environment byte identity.

Stable 1.6.0 includes the accepted RC1–RC5 generation, validation and business-rule changes. Read the migration notes before regenerating existing fixtures.

Choose A Guide

• Review The Output
• Related Tables
• Safety Model
• CLI Command Index
• Troubleshooting

Development

The accepted RC5 runtime includes saved-profile transformation CLI/MCP workflows; the prospective stable 1.6.0 retains them. See the CLI contract and acceptance limits. The separate transformation/session acceptance campaign remains open. Registration grants no dataset permission: preservation still requires exact local approval. Ordinary generation remains source-free; agents cannot issue that approval.

python3 -m pip install "uv==0.11.23"
uv sync --frozen --all-extras --no-install-project
uv sync --frozen --all-extras --no-editable --no-build-isolation
uv run --no-sync scripts/check_release.sh

See Contributing, Support, Governance, Code Of Conduct, Security Policy, Changelog, and License. Releases use PyPI Trusted Publishing with verified wheels and source distributions, checksums, SBOMs, and GitHub attestations.

AI-Assisted Development

AI-assisted changes require human review and tests; never send production data, raw PII, credentials, or tokens to AI. The name nods to Radiohead's "Paranoid Android"; this project is unaffiliated.

Agents can use the portable project workflow map and selective-transformation guide. These skills work offline with installed command help; they do not grant permissions or enable unfinished features. Agents that do not discover .agents/skills automatically can load either SKILL.md directly. Both are also included in the installed wheel under test_data_agent/skills; find that directory with:

python -c 'from importlib.resources import files; print(files("test_data_agent").joinpath("skills"))'

Point the agent at the selected skill directory using its own configuration; installing this package does not silently register skills with other agents.

Details for reviewed decimal inference, rule auditing and numeric profile limits: RC4 feedback corrections.

Business rules can travel in DatasetSpec.business_rules across Python, CLI, MCP, approved agent workspaces and every output format. DECIMAL formulas and aggregate mappings use exact arithmetic with one target-scale HALF_UP rounding. See portable business rules.