About the Oracle

The oracle is a small reference archive. It holds a handful of subjects that are worth knowing well: broadcast CRT and Sony PVM monitors, the Nintendo GameCube, The Legend of Zelda and the long history of breaking it in speedruns, Rust for Linux and the SPI bus, the Linux kernel, distributed systems, and Neuromancer. You can read it like a reference, or you can ask it a question and it will answer from what is written.

It does not know anything else. Ask it about the weather and it will tell you, plainly, that the weather is not in the archive.

Every answer is grounded. When you ask a question the oracle does not reach into a model's memory; it retrieves passages from this archive and answers only from those, citing each one inline. If nothing relevant is found, it says so rather than guessing. That refusal is not a failure mode. It is the point.

The retrieval is hybrid. A dense embedding catches meaning, a BM25 lexical score catches exact terms like a model number or a glitch name, and the two rankings are fused with Reciprocal Rank Fusion. A cross-encoder reranker then reads the question and each candidate passage together and scores true relevance, and a threshold on that score is the honesty gate.

A broadcast monitor has an underscan mode that pulls the picture inward to reveal the part of the signal normally hidden behind the bezel. The oracle borrows the word. Turn on underscan and the retrieval signal is laid bare: the query as it was tokenized, every candidate passage, its dense score, its lexical score, its fused rank, and the reranker score that decided whether it survived. The machinery is not hidden. You can watch it work.

The whole thing is built on SvelteKit and runs on Cloudflare Pages. The embeddings, the reranker, and the language model are all Cloudflare Workers AI, so there are no external API keys and the retrieval happens at the edge. The corpus is small on purpose, which is why brute-force cosine over a few dozen vectors is instant and needs no separate index.