Chain-of-Thought — reasoning made visible, not verifiable

What it is. In January 2022 a Google Research team showed that if you prompt a large language model with a few worked examples that spell out their intermediate steps, the model starts spelling out its own — and gets dramatically better at problems it had been failing. On GSM8K, a set of grade-school maths word problems, the technique lifted the largest model tested from under twenty percent to around fifty-seven. Nothing was retrained. The weights were identical before and after; only the shape of the prompt changed. A companion result later that year found that much of the effect could be had from the single instruction "Let's think step by step".

Why it matters here. Almost every agent in this collection stands on it. ReAct is chain-of-thought interleaved with tool calls; Reflexion is a chain of thought turned on the agent's own failures; the autonomous loops of 2023 are chains of thought wired to a task queue. Before this, a model's answer arrived as a single opaque emission. After it, the model produced a *trace* — a sequence of steps a reader could follow, disagree with, and point at. That is the substrate agent architectures were built on, and it arrived not as a new capability in the weights but as a discovery about how to ask.

And the reason it is filed in this museum rather than a textbook. A chain of thought looks exactly like a derivation. It is not one. It is generated text produced alongside the answer, not a log of the computation that produced it — and the two can come apart without anything in the output indicating that they have. Researchers demonstrated this directly in 2023: bias the few-shot examples so the correct answer is always in the same position, and the model's answers shift accordingly while its stated reasoning never mentions the pattern, instead constructing a fluent case for whatever it landed on. Subsequent work measured how far a model's stated reasoning can be perturbed without changing its answer, and found faithfulness varies by task and by model rather than holding as a rule.

So the trace is testimony that has the shape of evidence. It is genuinely useful — it makes errors locatable, it makes disagreement possible, it made a decade of agent work tractable — and it is not a receipt. Nothing in a chain of thought can be recomputed by a reader to check that it is what the model actually did, because the only witness to what the model actually did is the model.

Its place beside ReAct. ReAct's contribution was to interleave the reasoning with actions, and actions are the part that leaves marks: a tool call has a return value, an API has a status code, a file either exists afterwards or does not. That is the museum's whole thesis arriving through the side door — the parts of an agent's account that can be checked are the parts that touched something the agent did not author. Chain-of-thought made agents legible. It is the artifact that made *legible* and *verified* feel like the same word, and every verification object in this collection exists because they are not.

*Primary source inside: the January 2022 arXiv paper — fingerprinted and anchored like every object here.*