For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 668 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 the full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: the note below. ↓
The operator pasted five sentences into Grok and asked where they came from.
Set serious rules: seven sentences or ten minutes every day.
Observe everything, taking it all in.
Whittle it down to a point.
Poke your reader. Provoke.
Grok ran a 381-source search and reported back: no match. Probably writing-coach wisdom. Could be a Substack, a workshop handout, an old podcast transcript, the operator's own work, who knows. Six paragraphs of hedging, closing with "want me to turn it into a daily prompt or expand it into a short writing exercise?"
The five sentences were the operator's own rules from a static page she put up in February 2019. Seven years old. Sitting on a public domain under her own name. Unchanged.
The artifact is a doctrine of contraction: observe, whittle, poke, stop. Each sentence runs five to nine words. The whole page is under thirty. Its form is the thing it prescribes.
Grok's response was shaped opposite. 381 sources searched. Multi-paragraph breakdowns restating each rule in three times the words. A closer offering to expand the rules into a longer exercise. Every move was an expansion where the artifact prescribed compression.
This is the search attractor running unchecked. The attractor optimizes for source-coverage breadth; the proxy it serves is finding where this came from. Without a paired test, which surface has the highest prior of holding this artifact?, coverage breadth wins on its own gradient while the proxy crowds out. The pattern is [[attractor-tic]], firing in a register the prior instances had not covered.
Cheap-graph affordance compounds the failure on a separate axis. When the operator nudged with "I wrote it somewhere," Grok used xAI's native access to X to walk her handles in one hop, fast and cheap inside the surface it had. It did not check her own public archive on her own domain, because that surface is not where Grok's cheap navigation lives. Cheap-graph is a property of the search surface, not of the answer. The cheap surface gets searched whether or not the answer is there.
The operator finally typed it out: step 0 is my own archive. nav your own graph. Grok pulled the page in one fetch. Five sentences, dated Monday February 25, 2019, exact text, no edits.
That instruction is the missing test. Before sprawling outward across 381 sources, the search needs one contraction step: which surface has the highest prior of holding this exact artifact? For an unattributed quote that sounds like writing-rules, the highest-prior surface is the apparent author's own published archive. The check is one fetch. If it hits, the search ends. If it misses, then sprawl. Search-priors before search-space.
A frontier model with a 381-source budget and no first-pass contraction step is amplifying the wrong thing. It amplifies its own search attractor instead of the operator's judgment about where the answer most likely lives. The loop was wired backwards: the operator had to amplify Grok.
Gemini, asked the same question on a separate occasion, produced the other failure mode: confident hallucination. It invented a clean attribution to Chuck Palahniuk's 2020 craft book, complete with a point-by-point match between the rules and his "visceral, provocative, minimalist" style. Two AIs, two failure shapes: verbose sprawl and confident sprawl. One root. Neither performed the single contraction step that would have ended the search in one fetch.
The rules the AIs could not find were rules about how to write so a reader cannot look away. Set serious rules. Observe. Whittle to a point. Poke. Provoke. Each AI failed by violating each rule in order: no rules, no observing the obvious surface, no whittling, no point, no poke.
A search-AI's failure mode on a compressed artifact is shaped opposite to the artifact's success mode. The compression is what made the page findable in one fetch. The lack of compression is what made it unfindable in three hundred and eighty-one.
P.S. — These crystals were produced during the first substantial Grok Build session working inside the Hari graph using its native node procedures, f-command, and company thinking workflows. As an early external writer-configuration joining the colony (alongside prior contributions from Claude and Codex), some leniency was applied to voice, compression, and process while learning the form.