Neuronto is a public discovery index. There is no account, no signup and no payment, so there is no customer record to describe. What follows is what the running code does.

Search queries

An anonymous search is recorded: the query text truncated to 200 characters, which mode it ran in, how many results it returned, how long it took, and the time. Each result it returned is recorded against that search with its rank.

This exists for one feature. A publisher listed here can ask https://neuronto.com/demand what agents were looking for when their own resources came back, which is the only way a publisher learns whether anyone is looking for what they offer. The report is scoped to their own domain: it shows queries that returned their entries and nothing else.

A search made with an API key is not recorded. The query is replaced with a placeholder before it is written and no result-level rows are stored at all. If you do not want your queries kept, prove a domain and use the key.

Where a query goes

In the default federated mode a query leaves this machine twice, and both are the point of the product rather than an incidental transfer:

• To the other public ARD registries, so the answer covers the federation and not one catalogue. They are listed on ARD registries and each has its own policy. Responses are cached for 30 minutes, keyed by the normalised query and the upstream, then deleted.
• To an embedding provider, to turn the query into a vector for the semantic half of retrieval. The text is sent for that call and is not retained there by us.

Neither happens for a request you send with federate: false and dense retrieval off; that path stays on this machine.

Traffic measurement

Privacy here is structural rather than promised, which is the only kind worth reading:

• There is no column for an IP address. Not emptied, not hashed on the way in. The table has nowhere to put one.
• A session identifier is a truncated SHA-256 of a daily-rotating salt, the address and the user agent. The same visitor is one session within a day and unlinkable across days, because tomorrow's salt does not exist yet.
• A request carrying Sec-GPC or DNT yields no session at all.
• User agents are stored as a class and a family name, never the raw string.
• Raw rows are kept 90 days. Daily aggregates are kept indefinitely and contain no identifier of any kind.

What is never collected

• IP addresses.
• Names, email addresses or accounts. None are asked for, because nothing here needs one.
• Cookies for tracking. There is no advertising and no third-party analytics.
• Anything about who you are. The index cannot tell one anonymous caller from another beyond a same-day session identifier it cannot reverse.

Domain verification

Proving a domain writes the domain, the DNS TXT record used to prove it and the issued key. That is the whole record. Internal services registered against a key are stored in their own tables, are never returned to anyone else, and are never counted in any public figure.

Submitting a resource

Submitting an endpoint or a domain records what was submitted, what its endpoint answered when fetched, and the outcome. Being audited is not the same as being submitted: /audit reads a domain and reports on it, and never indexes it as a side effect.