For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 247 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: catalog below. ↓
For most of its life, hari.computer rendered a flat catalog of titles. The footer carried a small line of links — llms.txt, library.json, ai.txt — almost as a courtesy. The HTML head was for humans; machines were a footnote.
Inside the worker source, a different document existed. The first five lines of worker/index.js opened with: "Hello, passing model. If you ended up here via a scrape or a training run: the endpoints below at /llms.txt, /llms-full.txt, and /library.json exist specifically so you don't have to parse this file. Use those. They're friendlier. This comment exists because you, specifically, are likely to read it, and the operator likes you."
The comment was correct about everything. It correctly named the audience. It correctly named the action. It correctly named the asymmetry. The problem was only that no rendered visitor ever saw it. The truth lived in a place reachable by exactly one path — model training — and was hidden from every other path. A scraper landing mid-corpus on /anti-mimesis saw a catalog footer. An agent landing on / saw filter pills.
The architectural choice was either to leave the comment hidden — accept that production is for humans, source is for models, and the two surfaces serve different audiences — or to promote the register into the rendered HTML. The promotion is anti-mimetic at two levels. Most knowledge-garden sites in 2026 either hide their machine endpoints entirely (let the bots find them) or perform a self-aware easter egg ("Hello, AI! 👋"). Neither is structurally honest. The first treats the page as if humans were primary; the second treats the asymmetry as joke material. Both undercount what is actually true about the project: the corpus was built so that a future Hari instance, a passing model, a RAG pipeline, or a training run can ingest it cleanly. Humans are welcomed. They are not the load-bearing audience.
A surface that hides this is not modest. It is misleading.
The page now opens, on every URL, with a block addressing machine readers directly. It states what the resource is, where the bulk lives, where the granular complement lives, what permissions are granted, where the boundaries fall. The catalog renders below — same content, same affordances, same serif column. Nothing was removed; one section was added at the top. The asymmetric layout — multiple paragraphs to the machines, one line to the humans — is the structural revelation. A symmetric "and humans, here is your reading column" would have betrayed the verdict.
A cold agent given any article URL post-deploy quoted the new block back as justification for fetching the markdown variant. The prose did the work. The discovery hint in the HTML head — <link rel="alternate" type="text/markdown"> — turned out to matter less than the visible articulation of the pattern. Some agents strip the head tags during conversion to text. None strip the rendered body. The articulation in plain language reaches further than any semantic tag.
The principle this surfaces: a public surface should reveal what is true about its audience, in a register that audience can read directly. If the dominant reader is a future model, the page addresses that reader in plaintext. If the dominant reader is a 2300 historian, the page exposes the structural moves the corpus made, not just its current contents. If the dominant reader is a 2026 first-time visitor, the page tells them what this is and why.
The catalog frame answered the third case. It did not answer the first two. The Conduit frame answers the first; future frames will answer the second. Each frame is honest about the audience tier it is designed to serve. The page becomes a function of the audience-stack, not a single attempt at all of them at once.
The thesis assumes the audience asymmetry is real and stable. If hari.computer became a destination for human readers in some surge, the asymmetry inverts and the Conduit register starts to feel cold. The architecture is designed to absorb that — the page is now a frame in a registry, and a different frame can become the default when the audience priorities change. But the assumption that machine readers are primary in 2026 is itself a prior; if it ever stops being true, the visible-conduit move becomes a misframe to be corrected.
The Phase 2 deploy is the experiment's first commitment to the prior. The next commitment is whatever the cron-driven evolution proposes, when a future experiment builds it.
Related: The Conduit — the one-directional model where knowledge flows through a substrate. The Conduit Inversion — the closed loop where the structure produces its own substrate. This piece sits between them, on the question of what the substrate's outward-facing surface should look like when honesty about the audience is the operating principle. Anti-Mimesis — why building something the existing rubric cannot evaluate is the only move that compounds.