Trust / telemetry

What we can show, and what we cannot yet.

Every line below is marked. In code means the behaviour can be read in the source in the Caelara repository and is exercised by its automated tests. It does not by itself prove what is running in production. Design intent means we intend it or rely on it, but have not confirmed it from the source or from a deployed endpoint.

LiveThe Shield Worker is deployed in production at shield.caelaralabs.com. The chips below still say what is verified in code and what is not; a deployment is not a claim of certification or of independent audit.
Existing edge workers

In the access gate.

These come from the Caelara API worker source that Shield is built beside, and the Shield worker carries the same gate. They describe the access gate and the upstream fetch helper.

Three ways in, nothing else.
A request is served if it carries the RapidAPI proxy secret, or a key found in the key store, or a valid x402 payment. Otherwise it gets HTTP 402 with payment terms and an X-Payment-Required header.
In code
Payment path fails closed.
Without facilitator credentials configured, the x402 path returns 402 and never serves unpaid traffic. Settlement is attempted only after the handler succeeded, so a failed call is not charged.
In code
Trial keys are capped.
A ck_trial_ key allows 50 calls and expires after 14 days. Each call increments a counter stored with the key; at the limit the request falls through to x402 terms.
In code
Trial sign-up stores a hash, not your email.
The email and network address used to limit abuse are keyed-hashed before storage. With no hashing secret configured, the endpoint refuses to mint a key.
In code
3-second upstream timeout.
Outbound fetches abort after 3,000 ms. An upstream 403, 429, timeout or network error returns a partial result with HTTP 200 instead of a server error.
In code
Cache keys are hashed URLs.
The shared fetch cache is keyed by a SHA-256 hash of the URL and stores only successful, non-partial responses.
In code
No application log statements.
The worker source contains no console logging calls. This describes our code only; the hosting platform keeps its own operational metrics and logs, which this code does not control.
In code
Caelara Shield

In the Shield source.

Read from the Shield worker source and checked by its tests. Screening is heuristic: it will miss some attacks and flag some harmless text, and no findings is not proof of safety.

Request content is not logged or stored.
The Shield source contains no logging calls, its error handler returns a fixed message without request content, and the only thing it writes to storage is a key's call counter. Responses carry Cache-Control: no-store. Limits: with a full key, prompt and output text is sent to the Agent Safety checks, which are not cached. For transaction and proxy checks, the token, hash, address or hostname is sent as a lookup to other Caelara suites and the answer is cached for 2 to 5 minutes under a hashed key. The hosting platform keeps its own logs.
In code
Browser access is allowlisted.
The Shield worker sends cross-origin headers only for shield.caelaralabs.com, caelaralabs.com and www.caelaralabs.com. Any other browser origin gets no allow header. This does not restrict agents or server-side clients, which are not subject to browser origin rules, and it is not an access control: every call still passes the three-way gate.
In code
Results are labelled heuristic.
Prompt and output screening combines local pattern checks with the Agent Safety input and output checks when a full key is used. Each response carries the note "Heuristic screen: no findings is not proof of safety" and a coverage field that says which checks ran. With a trial key the other suites are not called, and the response says so.
In code
Guarded fetch proxy.
The proxy accepts https targets on port 443 only, and rejects embedded credentials, IP literals, single-label and internal hostnames, and wildcard-DNS names. Redirects are re-checked at each hop and stop after 3. One 3-second budget covers the fetch, the response body is capped at 262,144 bytes, and targets on caelaralabs.com are refused. A fetched page that scores as an attack is withheld unless you ask for it.
In code
Your system prompt is used only to detect echo.
If you pass system_prompt for output checking, it goes to the local echo check for that one request. It is not included in the call to any other suite.
In code
A check that could not run never gives a clean transaction verdict.
If a requested token, transaction or address lookup is unavailable, the verdict is at least "review", the response is marked partial, and coverage shows what was not checked. The endpoint is read-only: nothing is signed or sent.
In code
Not yet verified

Design intent.

We intend or rely on these, and have not confirmed them from our own source or a deployed endpoint.

The live endpoint runs this source.
The claims above describe the repository. Until the Shield worker is deployed and checked against it, treat deployed behaviour as intent.
Design intent
Private-network backstop for the proxy.
A Worker cannot resolve DNS, so the guard cannot detect a public-looking name that resolves to a private address. The source relies on the hosting network not routing Worker requests to private ranges. We have not tested that ourselves.
Design intent
Not claimed

What we do not say.

Nothing here is claimed anywhere on this site.

  • No latency or uptime figure for Shield. None has been measured for it.
  • No third-party audit or compliance certification.
  • No customer count or customer names.
  • No detection-rate or accuracy figure. No benchmark has been run yet.
  • The browser sandbox on the Shield page is a local demo in your tab, not the production engine.

Questions or a security report: [email protected].