Data residency & compliance — inference and data stay where you put them

Startups in regulated markets — India under the DPDP Act, 2023, China under PIPL / the Data Security Law / the Cybersecurity Law, and Europe under the GDPR and the incoming EU AI Act — face the same structural pressure: personal data (and, in China, “important data”) should be processed on infrastructure the operator controls, and cross-border transfer is constrained. A hosted, cross-border AI API is the awkward shape here. fak is the opposite shape by construction.

Fence — this is not legal advice, and fak is not a certified compliance product. fak is a technical enforcement boundary you deploy on infrastructure you choose. It gives you the controls (locality, deny-by-default, a tamper-evident log) that a data-protection program is built from; whether your overall system is compliant is a question for your counsel and your DPO, not for this page.

The mechanism, not a marketing claim

fak’s residency story is a direct consequence of four properties it already ships. None of these are new for this doc — they are the same properties described in the README and llms.txt, read through a residency lens.

1. Self-host-first. fak is one static Go binary whose whole external dependency set is two golang.org/x extended-standard-library modules. Put it in front of a local model (fak guard --gguf …, or fak serve fronting Ollama / vLLM / SGLang / llama.cpp on your own hardware) or a domestic provider, and the inference path never leaves infrastructure you control. Same artifact on a laptop and in a fleet — you add flags, not components or third-party services.
2. Fail-closed residency across backends. The boundary’s default is to refuse, not to forward-and-hope. An effect the policy did not allow is one that never leaves the box — there is no “leaked to an external service by default” path to reason about.
3. Default-deny capability floor (structural, not a classifier). Which tools may run and which tool results may re-enter model context is decided by the policy on the call path, in-process. A model cannot request an effect the capability was never wired for. Structure beats a recognizer a prompt can argue past — the relevant property when the thing you are containing is untrusted model output over personal data.
4. Tamper-evident audit surface. Every decision is recorded with an X-Trace-Id correlation id — an auditable “who asked for what, and what the kernel decided” trail, the kind of record a data-protection review asks for. See the gateway API reference and trajectory observability.

Mapping to the three regimes

India — Digital Personal Data Protection Act, 2023

The DPDP Act pushes fiduciaries toward processing personal data on controlled infrastructure and being able to demonstrate the safeguards around it. fak’s role:

• Locality: run the model and tool execution on your own Indian-region infrastructure; fak is the in-process gate, not another data processor in the chain.
• Purpose/So-what limitation as capability: the allow-list is the enforced list of what the agent may do with the data — a technical expression of data-minimization.
• Auditability: the per-call decision log is evidence for a review, not a promise.

China — PIPL / Data Security Law / Cybersecurity Law

The Chinese regime constrains cross-border transfer of personal information and “important data” and expects local security controls. fak’s role:

• In-country processing: the self-host boundary keeps PI and important data on domestic infrastructure; fak fronts the domestic models teams already run — Qwen and GLM are proven bit-exact in the in-kernel reference engine, and DeepSeek / Yi / Baichuan / Kimi and any open-weights model are fronted over the OpenAI-compatible wire — so residency does not cost you a model switch.
• Structural containment: default-deny means a cross-border effect must be explicitly wired to exist at all, which is a stronger control than after-the-fact detection.
• Records: the audit trail supports the “demonstrate your controls” expectation.

Europe — GDPR and the EU AI Act

The GDPR pushes controllers toward data-minimization and purpose limitation (Article 5(1)), demonstrable safeguards (the Article 5(2) accountability principle), and constrained third-country transfers (Chapter V). The EU AI Act adds, for high-risk systems, a mandatory tamper-evident logging obligation (Article 12, enforceable August 2, 2026). fak’s role:

• Locality and transfers: run the model and tool execution on your own EU-region infrastructure — or fully local with fak guard --gguf — and the inference path simply has no third-country transfer on it. The Chapter V question dissolves for that path instead of needing a transfer mechanism; fak is the in-process gate, not another processor in the chain.
• Data-minimization as capability: the default-deny allow-list is the enforced list of what the agent may do with the data — a technical expression of Article 5(1)(c) minimization, decided on the call path rather than promised in a policy document.
• Accountability and Article 12: every decision lands in an append-only, SHA-256-hash-chained journal that fak audit verify re-validates offline. The full requirement-by-requirement mapping to the AI Act’s logging obligation is in EU AI Act Article 12 conformance — the mechanism shipped for fak’s own integrity story and happens to be the regulated artifact.

See the localized front doors for the in-language version of this pitch: हिन्दी · 简体中文 · Deutsch · français.

What this does not claim

• It does not claim fak makes you compliant. It gives you locality, deny-by-default, and an audit trail; the program around them is yours.
• It does not claim a certification, a data-processing agreement, or a legal review.
• It does not change any benchmark or capability claim — residency is a deployment posture of the same boundary, achieved with flags (--policy, --gguf/local upstream, --require-key-env), not a separate product.

The honest summary: fak is the technical control surface a data-residency posture is built on — self-host the model, deny by default, log every decision, keep the data on the box — for exactly the markets where that posture is becoming mandatory. The go-to-market context is in the emerging-market adoption note.