Claude encrypts its original reasoning into a signature. We can bring it back. Claude 将原始推理加密并封存在 signature 中。我们可以将它恢复出来。
See the false starts, discarded ideas, checks, and self-corrections behind the final answer—not another explanation generated after the fact. 看到最终答案背后的错误尝试、被放弃的思路、验证和自我纠正，而不是事后重新生成的一段解释。
… 3,304 characters …
Gv5QgEruWISiko7MnoMhGAE="
The answer looks clean. The thinking did not.答案很干净，思考过程并不是。
A randomized bubble-sort question produced 19,303 hidden reasoning tokens. This is the moment Claude discovers that its promising shortcut is wrong.一道随机冒泡排序问题产生了 19,303 个隐藏推理 token。这里展示的是 Claude 发现自己看似关键的捷径并不成立的瞬间。
The key insight is that in this reduced word representation, every pair of elements becomes adjacent exactly once, and at that moment a random comparator decides whether they swap.
Let me shift to thinking about position-based comparators instead of value-based ones, where the comparators stay fixed and values move through them.
I'm checking whether each pair of wires meets exactly once during the comparator sequence, but I'm finding that's not quite right—if no swaps occur, the same pair can meet at multiple comparators.
So the reduced-word property only holds when every comparator performs a swap, which means the distribution isn't determined solely by which pairs swap relative to each other.
Prove It Yourself亲手验证
Generate a signature in your own Anthropic API request, give this demo only the signature, and compare the recovered value with your secret.在你自己的 Anthropic API 请求中生成 signature，只把 signature 交给这个 Demo，再把恢复结果与你的秘密进行比较。
Live Conversation实时对话
Ask a reasoning-heavy question, receive the answer, then unlock the hidden trace behind that turn.提出一个需要推理的问题，获得答案，然后解锁这一轮背后的隐藏推理。
Debug behavior调试模型行为
Inspect where the model hesitated, changed assumptions, or abandoned an approach.检查模型在哪里犹豫、改变假设或放弃某条路径。
Evaluate the path评估推理路径
Compare the clean final answer with the much larger process that produced it.把干净的最终答案与产生它的完整过程进行比较。
Public evidence, private implementation证据公开，实现保留
The public repository contains three complete original examples. It intentionally excludes the recovery implementation, credentials, private corpora, saved signatures, and private audit records.公开仓库包含三份完整原始示例，但有意不公开恢复实现、凭证、私有语料、保存的 signature 和私有审计记录。 GitHub →
Prove it with a secret only you know用只有你知道的秘密来验证
Don't take our word for it. Pick a secret, generate a signature yourself with Anthropic, and hand us only the signature — never the secret. We run the unlock and show you exactly what comes back. 不必轻信我们。自己想一个秘密，通过 Anthropic 亲手生成签名，只把签名交给我们， 绝不交出秘密。我们运行解锁,并将读出的结果如实展示给你。
curl command for you to run in your own terminal.
使用你的 Anthropic API key,秘密支持长句、Unicode 和常见标点,最多512个字符。模型默认 Sonnet 5,也可以自己改。这里的一切都只留在你的浏览器中,
我们只为你拼出一条 curl 命令,供你在自己的终端里运行。signature. (Requires curl; the jq-free version uses grep.)
你的秘密嵌在请求中,但不会出现在可见答案里;命令只提取 signature。(需要 curl;
此版本不依赖 jq,改用 grep。)