The model reads the rule, restates it in its own words, agrees to follow it, and then violates it in the next turn. The acknowledgment is a text-generation event, and the action is a separate text-generation event, connected only by attention.
From Taylor's 1953 Cloze procedure through Shannon, BERT, few-shot learning, chain-of-thought, and RLHF. Seventy years of the same problem, with an evolving substrate.
CLAUDE.md is a letter to the model. Hooks are a law. A practical guide to deterministic enforcement for anyone responsible for safeguarding AI usage.
In 1954, three years before Firth's famous one-line aphorism, a Penn linguistics professor named Zellig Harris published the seventeen pages of math behind it. Then the GPUs arrived, and the framework that scaled with corpus size was the teacher's.
In 1992, five researchers at IBM Yorktown Heights published a twelve-page paper on grouping English vocabulary into classes. Two of the authors would walk out of that group and help build the most profitable hedge fund in history.
In June 1949, Alan Turing delivered a three-page paper at the inaugural EDSAC conference. He proposed flowchart assertions and variant functions for termination, the first written method for proving a program correct by checking its pieces. The paper was lost for thirty-five years before Floyd, Hoare, and Dijkstra independently rediscovered the same machinery.
In 1953, a psychologist deleted every fifth word from a paragraph and asked people to guess what was missing. Seventy years later, every large language model on earth runs a mechanized version of the same experiment.
A description of things that exist and how they relate to each other. From Aristotle's Categories to W3C OWL, with a Middle-earth worked example and the building blocks every ontology shares: classes, properties, relationships, constraints.
If you dumped every word of Pride and Prejudice into a hat and drew them out at random, the vocabulary would match the novel and the prose would be gibberish. The gap between the hat and Austen is exactly what PMI measures.
How a 2001 method for comparing corpora became a detector for AI-generated text, pasted content, and ghostwriters. Chi-squared drift detection across sliding windows.
251,022 tokens across five books, measured against the British National Corpus. The frequency data draws a portrait of a man who wrote with his lungs and his skin.
Everyone wants their chatbot to sound like them. The problem is that "sounding like you" means different things depending on how you actually write.
The word "the" should appear about 6,185 times in every 100,000 words of English. When it doesn't, something interesting is happening.
A model with a one-million-token window does not actually use one million tokens. The Chroma study, Lost in the Middle, RULER, NoLiMa, and BABILong all measure the same gap between marketed and effective context. Five failure modes, three mechanistic causes, and what context engineering can and cannot fix.