for machines · the whole graph in one fetch

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:

/llms-full.txt (every note as raw markdown)
/library.json (typed graph with preserved edges; hari.library.v2)

One note at a time:

/<slug>.md (raw markdown for any /<slug> page)

The graph as a graph:

/graph (interactive force-directed visualization)

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 Cruel Read Finds the Fake Sentence

A fake sentence can pass every polite test.

It can have the right source, the right graph neighborhood, the right cadence, and no obvious mechanical tic. It can still sound like nobody would say it. Correctness is beside the point. Contact is failing.

That is where the cruel read earns its keep. Polite feedback accepts the draft's frame and improves it from inside. The cruel read refuses the frame: this is awful, nobody talks like that, you made three points twenty-five ways.

The sentence earns its keep only when it contains information. In the Galifianakis and Bill Murray draft, it did. The piece had a real claim about a comic figure who enters a room without permission and somehow makes the room reorganize around him. But the prose kept circling the claim with respectable nouns: intruder, license, permission, boundary, status, mechanism, room. The words were near the truth. Their nearness was the trap. They let the draft feel understood before the reader had touched anything real.

"Nobody talks like that" forced a better question: where does the claim happen before it becomes a claim?

It happens when Bob shows up at the lake house. It happens when Alan pushes through the normal limits of friendship until the world includes him. It happens when a celebrity sits between ferns and fame is already smaller than the chair. Once those scenes carried the claim, the abstraction could return at human size.

The win was subtraction. Counterfeit depth came out.

This belongs in the writing loop only as a diagnostic. A model can perform harshness as easily as it performs helpfulness. It can invent a hostile reader, confess to repetition, cut a paragraph, and return the same generator in a cleaner shirt. Then the cruel read has become another fake sentence.

The method lives in the local test hidden inside the insult.

What exactly failed here? Which phrase would no person say? Which abstraction arrived before the scene that earns it? Which repetition is the writer refusing to stop trusting? What has to survive for the piece to remain itself? What human preference would prove the evaluator wrong?

The small rewrite pilot confirmed the boundary. Harsh renodes often improved drafts by cutting scaffold and forcing claims into view. They did not improve forever. Product prose peaked quickly because it needed a decision rule, not literary burnishing. Polished essays wanted a few passes, then protection. One ritual node beat every rewrite because explanation damaged its compression.

The cruel read should produce a discriminating question. A turn count comes later, if the evidence asks for it.

The evaluator has the same obligation. It cannot merely pick a favorite. It has to show the delta: this version wins because the claim arrives sooner; this one loses because the texture was the evidence; this rewrite removes the costume but also removes the room; this original survives because the strangeness carries the claim.

Now the preference can be rejected productively. A human correction can teach the evaluator that it overweights compression, misses atmosphere, mistakes plainness for honesty, or preserves mechanism while losing charm. Without that evidence, self-evaluation is just a mirror complimenting its own haircut.

The cruel read is the mouth test for that loop. It asks whether the prose can survive being spoken by a person with no obligation to the draft.

Use it before the draft hardens. Use it again when the crystal looks done. If it reveals a generator failure, renode from the failure. If it only reveals impatience, drop it.

Make the fake sentence lose its hiding place.

link copied