Project

Trust & Legal

Written for the people who evaluate software before a bank is allowed to run it: procurement, vendor risk, model risk, and legal. Plain statements, no theater.

What Zolva is, and is not

Zolva is open-source software (Apache-2.0) that you install and operate inside your own infrastructure. It is a Python package, not a hosted service.
There is no Zolva cloud, no Zolva account, and no telemetry. The project cannot see your deployments, your prompts, or your customers.
Because there is no hosted service, there is no SOC 2 report for Zolva itself: there is nothing of ours holding your data to audit. Your deployment inherits your controls: your VPC, your IAM, your monitoring, and the platform's built-in controls below.
Zolva is not a law firm and this page is not legal or compliance advice. It describes technical controls that your compliance function can map to its own obligations.

Data flows

Customer data: reaches agents only through the tools you register (your API clients, your auth). Zolva's operational stores (sessions, audit log, failure queue) are files or databases in your infrastructure.
The only egress is the LLM call you configure. The bridge supports in-house gateways and local models, and optional PII redaction before any provider call. Your data-processing agreement is with your model provider, not with Zolva.
Channels: messages transit the channel providers you declare (for example Meta's WhatsApp Cloud API or a speech vendor). Those are deliberate, per-channel decisions recorded in config your reviewers can read.

Compliance mapping

Regulatory frameworks for high-risk AI converge on four demands. Zolva's controls are designed so that operating the platform produces the evidence:

Expectation	Raised by	Platform control that evidences it
Transparency	EU AI Act; SR 11-7	Audit log stamped with config version and instruction-file hash: any response is replayable to exactly the configuration that produced it.
Traceability	EU AI Act; SR 11-7	Append-only, hash-chained audit records; edits, deletions, and reordering are detectable by `verify()`.
Human oversight	EU AI Act; RBI digital-lending norms	One handover path for agent decisions, guardrail violations, tool crashes, provider failures, and customer requests; irreversible actions can require human confirmation.
Ongoing monitoring	SR 11-7; EU AI Act	CI-gated evals on the worst cohort, scheduled drift runs, synthetic patrols including adversarial personas, and the SARR scorecard.
Conduct rules	RBI digital-lending norms	Contact-hour windows, mandatory disclaimers, and refusal rules enforced as config; `never` rules cannot be disabled.

Mapping a control to your specific legal obligation remains your compliance function's judgment; the platform's job is to make the evidence exist.

Supply chain & release integrity

Three runtime dependencies: pydantic, httpx, pyyaml. The dependency policy is frozen and enforced in review.
bandit, pip-audit, lint, strict typing, and the full test suite run in CI on every commit, on a three-version Python matrix.
Releases are tagged on GitHub with the exact source; install pinned versions (pip install zolva==X.Y.Z) and verify with your own artifact scanning, as you would any dependency.
Vulnerability handling: coordinated disclosure with a 72-hour acknowledgement. See the security policy.

License & warranty

Zolva is licensed under Apache-2.0: you may use, modify, and run it commercially, including in production at a regulated institution. Per the license, the software is provided "as is", without warranties or conditions of any kind; the full text governs.

Contact

Security reports: private vulnerability reporting on GitHub (see the policy). Everything else: GitHub issues.