Sponge

Retrieval is the product

Sponge can publish only what it can still find. The product problem is the quality of retrieval: whether a source remains addressable, attributable, and distinct from the fluency that later surrounds it. Another generated page does not solve that.

In August 2026, Vass Bednar wrote in The Walrus that Google Search is dying because the corpus beneath it is collapsing. Sunset times invented by AI summaries are the comic case. The serious case is a web where original pages rot, get deleted, or sit behind a generated answer that never sends the reader to the source. hraness recorded that essay as a reading note. The note is useful. Reprinting it is not the job of this page.

Bednar’s claim is institutional. Retrieval is public infrastructure, and volunteer archives cannot hold it up alone. Sponge’s claim is product-shaped. A research workspace that uses AI to search still has to keep the index honest. Publication is a later act.

A generated answer is authorship

A German court held Google liable for a false AI overview after the overview linked two publishers to scam practices. The court treated the rewrite as authorship, not as a pointer into the public record. That is the same kind Sponge already names. An agent may search and propose. A person accepts research into a private graph. A public /k locator still requires a second review, eligible rights, an exact supported edition, explicit publication enablement, and a deliberate release.

When a summary replaces the click, the product has published. It has also hidden the support. Wikipedia becomes, in Bednar’s phrase, “the infrastructure of its own demise” when models ingest the page and keep the visit. Sponge’s public contract runs the other way. Citations stay bound to graph records, and a released edition keeps the evidence-bearing path beneath the prose.

Retrieval fails when the source disappears

Bednar writes that the corpus is collapsing in real time. Disney deleted the FiveThirtyEight archive after the remaining staff were laid off in March 2025. Newsrooms now block the Wayback Machine because an archived page can feed a model. Planted Reddit posts try to poison the summaries that never disclose their diet. A better query cannot recover a page that no longer exists, or a page the interface will not show.

Sponge treats that as an index problem before it is an editorial one. Lossless document payloads stay immutable. The control plane stores exact references. A public locator binds an identity to a revision. The about page calls the durable object the path: identities that survive renaming, statements that retain context, assertions that remain attributable, and editions that preserve their own history. How Sponge knowledge works states the live boundary. None of that is a substitute for finding the source in the first place.

Keep the hit and the edition apart

Presence is not a publication already separates a live profile from a reviewed edition. This page separates a hit from a release. A question can stay private. An Inquiry can accumulate evidence, disagreement, and rejected proposals. Agents may search, structure, compare, and challenge. People accept, reject, and publish.

hraness/kb keeps the same split for agent memory. Markdown and Git stay authoritative. Indexes and graph views are replaceable. Search finds files. It does not become the record. Sponge applies that lesson to product knowledge. The index has to be good enough that a person can inspect what was found. The published object is still an edition, not the ranking, the embedding, or the fluent paragraph that cited nothing.

Read Bednar in The Walrus, or through the hraness reading note. The kb product states the retrieval contract those notes assume. This take is served on sponge.computer. Presence remains a separate take. Agent locators are collected on developer resources.