If a compliance review, a PII/PHI policy, or a "we can't copy that into staging" conversation has stalled something you needed to ship, this page exists so you can see exactly what SpiceGrinder does and doesn't do with data.
There's no code path in SpiceGrinder that reads or transmits your production data, no telemetry, and no external service in the loop. Here's what that means in practice.
No proprietary format. Every model is a readable XML or JSON file you can diff, version-control, and hand to an auditor.
> spicegrinder generate Reading model.json SEED=482913 > spicegrinder validate model/customers.xml
model.json entities: customers: orders: addresses: seed: 482913
SpiceGrinder never imports, ingests, or connects to your production systems — there's nothing to leak.
Same model file, same seed, byte-identical output every time — runs fully air-gapped, with no outbound connectivity required.
1111111111111111111111111111 1111111111000000111111111111 1111111100011110001111111111 1111111000111111000111111111 1111111000111111000111111111 1111111100000000001111111111 1111110000001100000011111111 1111100000001100000001111111 1111110000000000000011111111 1111111111111111111111111111
SpiceGrinder runs entirely on infrastructure you control — your laptop, your CI runner, your own server. There is no telemetry, no phone-home license check, and no background sync. What you run is what runs; there is no external service in the loop.
SpiceGrinder can optionally log audit, performance, and diagnostic events for your own troubleshooting. Every one of those streams is off by default, and when enabled, writes local, plain-text NDJSON files under your own home directory. None of it is ever sent anywhere.
SpiceGrinder isn't an AI model guessing at what your data "probably" looks like. The same model file and seed produce byte-identical output every time, on every machine. There's no sampling drift, no hidden training data, and nothing to audit beyond the plain-text model in front of you.
If you have agentic workflows, SpiceGrinder fits in as a set of tools your agents can use. Its clean, interface-driven design and plain-text artifacts make it easy for agents to understand context without digging through your business rules. We also plan to ship sample skills. Whether to use AI agents at all, and what access they get, is always up to you.
Because SpiceGrinder has no external dependency to function, it runs equally well fully air-gapped, inside a locked-down VPC, or on a laptop with no network access at all. Nothing about generation requires outbound connectivity.
SpiceGrinder is not independently certified against SOC 2, HIPAA, or any other formal compliance framework, and running it doesn't make your broader environment compliant by itself. What it does do is remove one specific, common blocker: needing to move real PII/PHI into test, demo, or training environments in the first place.
If you're evaluating SpiceGrinder for a regulated environment and need more detail than this page covers — architecture questions, a security questionnaire, deployment specifics — reach out and we'll be happy to discuss it with you.
