Sponge

A search subagent should return documents, not the answer

Sponge can trust a search agent only when the agent still returns the pages it used. The search step gathers support. A later reasoner, and then a person, decide what that support means.

On 26 March 2026, Hammad Bashir, Kelly Hong, Patrick Jiang, and Zhiyi Shi published Chroma Context-1. The 20B model is a search subagent. It splits a question into subqueries, searches over several turns, and returns a ranked set of supporting documents. A later model may answer from that set. Context-1 does not. hraness recorded the paper as a reading digest. The digest is a pointer. This page states Sponge’s claim.

Retrieval is the product already treats index quality as the product problem and treats a generated answer as authorship. This page names the next split. Missing evidence is worse than extra noise. A later reasoner can drop a document it received. It cannot invent a document that never arrived.

A later reasoner cannot invent a missing document

Context-1 is trained to retrieve, then stop. The authors trained it on more than 8,000 synthetic multi-hop tasks and weighted recall far above precision because the next model can ignore an extra page and cannot recover a page that search never returned. They measure whether the needed documents appear in the output set, not whether a fluent paragraph happens to be right.

That is the useful product rule. Agents may search and propose. A person still 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. If the search step already wrote the answer, those later acts have nothing inspectable to review.

Keep the document; drop only the leftover chunk

Long search fills a window with pages that later turns no longer need. Context-1 prunes its own context. The authors keep documents at document grain and refuse to summarize them into a shorter proxy. A discarded chunk is gone. The remaining pages stay readable.

A summary that replaces the source hides the support. How Sponge knowledge works keeps citations bound to graph records. 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. Those objects need the pages search found, not a paraphrase of the search.

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.

The search return is still a proposal

Presence is not a publication separates a live profile from a reviewed edition. Retrieval is the product separates a hit from a release. This page separates the search return from the answer. A ranked document set can enter an Inquiry, collect disagreement, and still be rejected. People accept research. Publication remains a later act.

Read Context-1 on Chroma, or through the hraness reading note. This take is served on sponge.computer. Presence and retrieval remain separate takes. Agent locators are collected on developer resources.