High-fidelity synthetic data — with measured proof on every run.
Point Synthesis at your own CSV or Parquet. The learned engine learns the joint distribution and emits synthetic data with source-relative fidelity and privacy measurements on the delivered rows.
One learned engine, three governed operating modes, and one request-to-evidence authority. Quality, Automatic, and Fast change control policy and latency—not the model hidden behind the product.
Whether the generated schema actually covers the concepts you asked for. PASS means every extracted concept is represented.
How readable, deterministic, and policy-clean the synthetic rows are after every post-render repair has run.
Cross-field invariants the engine evaluates against every generated row (arithmetic totals, derived numerics, enum coherence).
Judge the delivered dataset by its own evidence.
Every Synthesis run measures the actual delivered rows against the customer source. Per-column similarity, bivariate fidelity, privacy risk, and the release decision are emitted into the sealed reports instead of relying on a single headline benchmark.
The artifact, utility report, privacy report, release decision, and evidence bundle are bound to the same job authority.
Choose quality, governance, or speed—not a hidden model.
Every mode runs the same learned Synthesis engine and emits the same artifact, quality, privacy, and evidence authority. The difference is who owns the training controls and the authorised latency ceiling.
Quality
The highest-quality governed profile, with source-relative measurements on every delivered dataset.
- Customer-authorized advanced training controls
- Held-out convergence can stop before the ceiling
- Deterministic: same seed + request → same bytes
- Every artefact binds into the evidence chain
Automatic
The server owns training controls and stops from measured held-out convergence.
- One learned Synthesis engine
- Dataset-shape and capacity-aware compute
- Exact quote before execution
- Quality and privacy measured on delivered rows
Fast
A bounded lower-latency profile of the same learned Synthesis engine.
- Server-owned training ceiling
- No hidden statistical fallback
- Same artifact and evidence authority
- Measured quality trade-off remains visible
One call. One quote. Verifiable output.
Hand the client an owned dataset, row count, mode, seed, and maximum authorised charge. The server binds the request, runs Synthesis, and emits the customer artifact plus quality, privacy, and evidence records.
- Deterministic: same request + seed → byte-identical output
- Fail-closed: privacy risk over threshold → job aborts before emitting
- Offline-verifiable: the verifier CLI re-runs the chain locally
- Three public modes, one learned engine and one evidence contract
from radmah_sdk import RadMahClient, SynthesisReleasePolicy
client = RadMahClient(api_key=os.environ["RADMAH_API_KEY"])
release_policy = SynthesisReleasePolicy(
minimum_overall_fidelity=0.75,
minimum_column_fidelity=0.75,
minimum_bivariate_fidelity=0.60,
maximum_membership_advantage=0.20,
maximum_linkage_risk=0.05,
require_zero_exact_copies=True,
)
dataset_id = "owned-dataset-uuid"
quote = client.estimate_cost(
kind="synthesize", dataset_id=dataset_id,
rows=1_000_000, mode="quality", compute="auto",
)
job = client.create_job(
kind="synthesize", dataset_id=dataset_id,
rows=1_000_000, seed=42, mode="quality",
compute="auto",
release_policy=release_policy,
max_credits=quote["credits_required"],
)
job = job.wait() # terminal status or explicit failure
quality = client.get_utility_report(job.id)
privacy = client.get_privacy_report(job.id)
bundle = client.download_evidence_bundle(job.id)
assert client.verify_job(job.id)["verified"]Quality metrics — emitted, not claimed.
Every synthetic dataset ships with quantitative fidelity and privacy measurements inside the evidence bundle. No subjective claims — the bundle proves the numbers and the verifier confirms the chain offline.
Distributional similarity
Per-column similarity between real and synthetic marginals is measured on every run. The full number set is written into the sealed utility report for the delivered dataset.
Correlation preservation
The pairwise correlation matrix is compared real vs synthetic. Every run passes the correlation gate and ships — or aborts fail-closed. Exact numbers land in the utility report.
Privacy risk metrics
Nearest-neighbour distance, membership-inference, attribute-inference and disclosure-risk metrics are measured and written into the privacy report. Zero PII by construction — synthetic rows are sampled, never copied.
Sealed evidence bundle
Every run emits the contract snapshot, reproducibility record, constraint/utility/privacy reports, run telemetry, artefact index and a chain seal. The verifier CLI replays the chain offline.
Audit-ready from day one.
Cryptographic evidence bundles, privacy-risk metrics and deterministic reproducibility give your compliance team what a DPIA, an ISO 27001 data-sharing control, or a SOC 2 audit needs to see. (RadMah states control alignment — it does not claim certifications it hasn't completed.)
Cryptographic evidence bundles
A signed, multi-artefact bundle covering the sealed contract, determinism proof, constraint/utility/privacy reports, run telemetry, evidence record and seal. Hashes chain every artefact; any change breaks the chain and the verifier refuses to certify.
GDPR Article 89 alignment
Synthetic data that doesn't relate to an identifiable person falls outside the material scope of GDPR Articles 5–15. The bundles document the process and full privacy-risk metrics, supporting Article 89 research-exemption claims and DPIA submissions.
Audit-ready provenance chain
Bundles provide the traceability, reproducibility and integrity verification auditors require. Every generation is deterministic — same contract + seed → byte-identical output — and the audit trail is cryptographically immutable.
Zero PII by construction
Records are sampled from the learned joint, never copied from source rows. Privacy reports measure disclosure and inference risk; when risk exceeds your configurable thresholds, the job fails-closed before emitting data.
Cross-border-transfer enabler
Synthetic output lets teams in different jurisdictions work on faithful data while source data stays in sovereign storage — removing the need for SCCs on the synthetic artefact.
Deterministic reproducibility
A sealed contract plus a seed produces byte-identical output on any host, any time. The reproducibility record captures the exact RNG state at every checkpoint.
Bring a CSV. Leave with an evidence bundle.
A 30-minute working session: upload (or we mock) a representative dataset, we run it through the flagship engine, and you keep the sealed evidence bundle plus the utility report.
Mock Data
Deterministic schema-driven fabrication when statistical fidelity isn't the goal.
Agentic Data Scientist
A governed agent plans Synthesis with exact review and approval gates.
Healthcare FHIR
HL7 FHIR R4 bundles with LOINC, RxNorm, ICD-10-CM.
Evidence chain
How every run is sealed end-to-end, and how you verify a bundle offline.