For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 780 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. ↓
An AI scanned 778 of my public notes and found familiar AI-writing phrases in seven. It then concluded that 3 percent of the corpus was AI slop, with 82 percent credence that the true figure lay between 1.5 and 6 percent.
The scan produced evidence. The answer never shows a path from that evidence to those numbers.
The report also pulled the full corpus, found no short notes built mainly from bullets, inspected stylistic tics, read a sample, and examined category and length distributions. It looked at how the notes linked together and how they were selected. Those checks make the judgment better informed. The route from them to 3 percent remains undisclosed.
The report counted surface markers while its conclusion judged a larger property: whether the writing was low-value generated material. Its explicit count used notes. Its conclusion described weak sentences or short passages. The phrase AI slop can itself refer to stock language, deceptive provenance, low information for a particular reader, unwanted volume, or work nobody seriously judged.
To reach one corpus-wide percentage, the evaluator still had to define that property, choose a unit, label examples, show how the proxy checks predicted those labels, explain how the parts became a whole, and account for the credence range. No such bridge was disclosed.
The prompt had requested a percentage and high confidence. The response filled those fields. It treated the requested answer type as evidence that an answer of that type was available.
A requested answer type is a hypothesis about what the evidence can support. The request itself supplies no evidence.
The evidence earned a narrower answer: the listed stock phrases were rare, the checked list-heavy form was absent, and the sampled writing appeared structured and heavily filtered. A personal impression could identify itself as an impression. A black-box instrument could warrant precision through a calibration record.
Those observations stand. Three percent remains unaccounted for.
Before debating the digits, ask what operation produced them.