Why I'm building Hyphae: provenance over prediction (and the 3-line baseline that tied it)

Hyphae is a CPU-only Rust system that answers by quoting stored fragments verbatim from a SHA-256 hash-chained journal.

Benchmarks showed a trivial echo baseline matched more complex LLM and retrieval setups on correctness and grounding.

That result reframed the project: the real value is verifiable provenance, not better prediction.

Byte-identical quotation plus hash chaining makes answers auditable, and an external Ed25519 signature anchors the journal head.

The article argues grounded AI should be source-bound and independently verifiable, not just fluent.