Core Concepts
The foundational building blocks of RadMah AI: how specifications are written, how data is generated, how provenance is sealed, and how billing works. Every concept on this page applies uniformly across all engines and plan tiers.
sealed contract
Every generation job starts with a sealed contract — a typed, machine-readable specification that fully describes what to generate. The platform never generates data without one; the sealed contract is the single source of truth for the entire pipeline.
ℹWhat a sealed contract captures
At a high level, a sealed contract captures the data shape you want, any business or physics constraints that must hold, which engine family will run, and the seed and row count for reproducibility. The platform fills in the details from your natural-language description and uploaded data — you do not write sealed contract by hand.
sealed contract is serialised deterministically and cryptographically sealed to produce a unique commitment. Once sealed, a sealed contract is immutable — no tool in the platform can modify it. Sealed contracts are stored and referenced by ID, enabling deterministic replay at any future point.
Engines
RadMah AI provides five generation engines. All plans — including Free — have access to every engine. Credits are the only gate; there are no engine-access restrictions.
| Engine | Technique | Best for |
|---|---|---|
| Mock | Deterministic, seed-reproducible generation from a natural-language description | Instant test data, prototyping, CI fixtures — no training data required |
| Synthesize | Our trained generative model, learned from your uploaded data | Privacy-safe replicas that preserve distributions, correlations, and statistical properties of real datasets |
| Constrained Synthesis | Business-rule enforced generation with hard-constraint sampling | Domains with strict regulatory or business rules that must hold on every generated row |
| Virtual SCADA | Physics-based industrial simulation with 6 live OT protocols | OT/ICS telemetry for SCADA testing, digital twin validation, and operator training |
| ICS Security | MITRE ATT&CK ICS attack-scenario generation with ground-truth labels | Cyber-range datasets for IDS/SOC training — pcapng, truth.ndjson, signals.parquet bundles |
✦AI-powered engine selection
You do not need to choose an engine yourself. The AI Orchestrator automatically routes your request to the right engine based on the sealed contract specification and your data.
Evidence Bundles
Every generation job produces a signed, cryptographically-chained evidence bundle, unconditionally. The bundle is identical across all engines, all plan tiers, and all dataset sizes. It cannot be disabled or reduced.
ℹWhat's in the bundle
A snapshot of the sealed job spec, input and output fingerprints, the generation log, a determinism proof, and reports for statistical fidelity, privacy risk, and constraint satisfaction — all bound under a single tamper-evident cryptographic seal.
The bundle's artefacts form a cryptographic chain. Any modification to any artefact breaks the seal. This provides tamper-evident provenance that is independently verifiable without access to RadMah AI infrastructure.
⚠Unconditional guarantee
Evidence bundles are not an add-on or premium feature. They ship with every job on every plan, including Free. If a job succeeds, the full bundle exists. No exceptions.
AI Orchestrator
The AI Orchestrator is the intelligence layer that translates natural-language requests into fully-typed sealed contract specifications and routes them to the correct engine. It consists of two components working together.
Broker
The single-turn front door. Handles straightforward requests that can be resolved in one pass: “Generate 50K rows of customer transactions” becomes a sealed sealed contract and a dispatched job in a single turn.
Agentic Data Scientist (ADS)
The multi-step autonomous brain. When a request requires analysis, training, iterative refinement, or multi-entity planning, the Broker delegates to ADS. ADS executes a full pipeline — analyze, train, generate, validate — across multiple jobs.
✦Automatic routing
The user never selects Broker or ADS directly. The system uses a 3-level auto-routing pipeline to determine whether a request is single-turn or multi-step, then delegates accordingly. The right engine, the right plan, and the right execution path are all selected automatically.
Credit System
ℹCore billing unit
1 credit = 1,000,000 generated rows. Credits are the unified billing unit across all engines. Credits are deducted only after successful job completion — never at submission time.
| Plan | Monthly Price | Credits / month | PAYG Rate | Concurrent Jobs |
|---|---|---|---|---|
| Free | $0 | 25 | Not available | 1 |
| RadMah AI Sovereign | $6,000 | 1,500 | $2.50 / credit | 10 |
| Enterprise | Custom | Custom | Negotiated | Dedicated |
Training costs
Training jobs (used by the Synthesize engine) consume credits based on a GPU-tier multiplier. The platform displays the training tier and estimated credit cost before you confirm any job.
| GPU Tier | Epochs per Credit |
|---|---|
| CPU / GPU-S | 10 |
| GPU-M | 5 |
| GPU-L | 3 |
Credits expire at the end of each billing month. PAYG credits purchased mid-month expire at the same month-end. All plans — including Free — get access to every engine, the full signed evidence bundle, the AI Orchestrator, and the Agentic Data Scientist.
Determinism
ℹBit-identical reproducibility
Same sealed job spec + same seed = bit-for-bit identical output, on any machine, at any future point. The determinism proof in the evidence bundle is independently verifiable without access to RadMah AI infrastructure.
Determinism is enforced by a cryptographically-strong, seed-reproducible random-number generator driven from the sealed job spec. A cross-platform cryptographic core provides canonical serialisation and hashing so results are identical across environments.
This guarantee enables audit replay, regulatory re-verification, and exact reproduction of any previously generated dataset — without needing to store the output data itself. The sealed job spec alone is sufficient.
Multi-Tenancy
RadMah AI is a multi-tenant platform with complete isolation between organizations. Every API request, database query, and storage operation is scoped to a tenant_id. There is no cross-tenant data access at any layer.
⚠Isolation is absolute
Each tenant has its own datasets, seals, jobs, evidence bundles, trained models, and billing state. No API endpoint, admin tool, or internal process can read or modify another tenant's data.
Tenant roles
| Role | Capabilities |
|---|---|
| Owner | Full control — billing, members, settings, all operations (one per tenant) |
| Admin | Manage members, datasets, jobs, and seals — no billing access |
| Member | Create and run jobs, upload datasets, manage own resources |
| Viewer | Read-only access to all tenant resources |