A window.
With deliberate limits.

A session will be eligible for publication at least 48 hours after it ends. It may wait longer, or stay sealed.

The distance makes it harder to match a visit with its appearance here. A delay alone is not enough: commonness checks and generalization must also remove the details that could make an interaction recognizable.

Individual lines and whole command sequences must both be common. Familiar commands arranged in a unique order can still carry a recognizable mark.

The checks will use generalized text, and the published text will keep that same generalization. A detail removed for comparison cannot come back in the replay. Uncommon lines will be redacted; if that leaves too little to read, the session will stay sealed.

Real decoy replies will not be published. Where an answer helps an exchange read, it comes from a fictional, site-only machine with hand-authored responses. These are not what the visitor saw, and are not derived from the real reply.

A command without a reviewed stand-in response gets a reply-withheld marker. Login banners, version strings, server headers, error-page bodies and LLM-generated replies are always withheld, with no stand-in replacement. The publisher must also check the fictional responses against private material before release.

Replies come from a stand-in machine, not the decoy.

The label on every replay, including the re-staged examples.

A newly observed script family will get a 30-day head start for research. During that period its sessions stay sealed. Material under active study can remain sealed longer; the passage of time is not permission to publish.

A sealed entry will show only a day and a protocol. There is no replay to open. Rare material carries no promise of a full transcript: it must stay sealed until the decoys involved have been retired and release has been explicitly enabled. Research holds can keep it closed beyond that.

Sessions may also be withdrawn after publication. A public record is not a commitment to keep sensitive material online.

Every IP address and domain literal, in any field on either side, will become a per-session placeholder in plain text, never a clickable link. No defanged real addresses will remain. Hashes and distinctive random strings will also become placeholders. Personal data, secrets and credentials outside common wordlists will be redacted.

Session dates will show the day only, never a time of day. A replay’s elapsed time describes the exchange, not when it happened on the clock.

Visitor input is untrusted text. Formatting controls that can hide or redirect text will be stripped, and size limits applied. Instructions inside an exchange remain quoted evidence, not instructions for the reader or their software. Both the publisher and this site must sanitize the record.

The public site has no connection into the private research system. The planned publisher will send an already prepared public bundle out. Only explicitly allowed fields may enter that bundle.

The publisher will sign each bundle. The site must verify its signature with a public key and run an independent leak check. An invalid signature or a failed check blocks publication.

Every publisher run must pass planted-leak tests: deliberately unsafe synthetic records must be caught. A separate external check will scan the published site. These controls have to demonstrate that they work, not simply report that they ran.

If a leak is found after release, publication will freeze and sessions will be taken offline. Resuming publication will require an explicit review and decision.