Zero-Retention Audit

Zephyr 662 LLC · Last reviewed August 21, 2026

Saying “we don't store your prompts” is easy. This page is the working version: every surface in our stack where a prompt or completion could persist, the specific control on each, and how that control is verified. It is the internal audit we run against every code change, published as-is — including the surfaces we do not control.

Persistence surfaces

#	Surface	Control	Verified by
1	Gateway application logs	A whitelist enforced in code: a non-whitelisted field name raises rather than logs. Values are scalar-coerced and length-capped.	Sentinel sweep; unknown-field rejection test
2	Exception messages and tracebacks	Errors record the exception class name only, never its text — exception strings routinely embed request fragments. The catch-all handler returns a generic body.	Dedicated test asserting exception text is never logged
3	HTTP access logs	Disabled at the server. The error log carries no bodies.	Deploy config review
4	Usage database	The schema has no content columns. Not empty ones — none exist, and adding one is forbidden at the source.	Test dumps every row and asserts sentinels absent
5	Third-party library debug logs	HTTP client and ORM loggers are pinned to WARNING even if the root log level is lowered, so a debug flag cannot start echoing bodies.	Code review of logging setup
6	Crash dumps and core files	Core dumps disabled at the container level.	Deploy config review
7	Inference engine request logging	Request-payload logging is disabled by the launch scripts. The flag name has been renamed upstream before, so deploys grep for it.	Deploy smoke test greps the flag
8	Engine disk spooling	The engine does not spool request content to disk; local NVMe holds model weights only.	Image review at pin time
9	TLS terminator logs	Access logging explicitly discarded.	Deploy config review
10	URLs and query strings	Content travels in POST bodies only and is never placed in a URL; the gateway logs no query strings.	Code review
11	Test-harness output	Harnesses run synthetic, self-generated traffic only. There are no production traffic captures to point them at, by rule.	Operating rule
12	Prefix and KV caches	Accelerator memory only. Evicted during normal serving, cleared on restart, never serialized to storage.	Engine config review
13	Host swap	Swap is disabled on the gateway and on serving nodes, so process memory holding a request cannot be paged to disk.	Provisioning check at node acceptance

What we do not control

Three surfaces sit outside our software. We would rather name them than let the table imply a completeness it doesn't have.

Host operators. GPU capacity is rented. The personnel who run those physical machines hold administrative access to their own hardware. Our stack writes no content there — but we do not control their platforms, and no provider who rents compute can honestly claim otherwise.
Your client. Anything your own application logs, caches, or forwards is on your side of the wire.
Legal process. We respond to valid legal process. We can only produce what we have, which is the request metadata described in the privacy policy — never prompts or completions, because they do not survive the request.

How this is tested, not just asserted. The executable form of this table is a test suite that injects unique sentinel strings into prompts and completions, drives them through every path including error and mid-stream-failure paths, then asserts those sentinels appear in no log stream and no database column. It runs on every change. A claim in the table above that isn't covered by a test is a bug in the table.

What is explicitly not here: abuse-scanning retention. Some providers retain prompts to scan for misuse. We do not, which is why our acceptable-use enforcement is complaint- and metadata-driven rather than content-driven. It is a real tradeoff and we would rather state it than quietly keep a copy.

Rules that keep it true

New log call sites use the whitelisted logging helpers — never a raw logger, never a bare print. Adding a permitted field name is a reviewed change that updates this page.
No new database column may carry content. When a debugging need arises, the answer is synthetic reproduction, not capture.
Engine image bumps re-verify rows 7 and 8 before the image serves traffic.
Any new subprocess or sidecar gets a row here before it ships. Tracing and APM tools that capture HTTP bodies are banned outright.

Zephyr 662 LLC · 800 Ave Ponce De Leon Ste A, Unit 9121, San Juan, PR 00908 Home Privacy Terms