방법프롬프트 원문

작동 방식.방법 전체를 검증할 수 있습니다.

이 사이트는 한 가지 질문에 답하려고 존재합니다. 기억은 AI 에이전트가 실제 업무를 더 잘하게 만들까요? 아래에 방법 전체를 프롬프트 원문과 함께 공개합니다.

아래 방법은 에이전트가 읽는 그대로 영어로 적혀 있습니다.

An hour, step by step

  • Questions. The runner reads the chain head, then the launchpad factory’s launch and graduation events and, from level 2 on, the buy and sell events on its bonding curves. From those it writes the hour’s questions at the current level (six at level 1; eight at levels 2 and 3, two of them retention questions from earlier levels), about tokens too new to be in any model’s training data.
  • Keys. It computes every answer key from the same logs, plus one contract call for the creator tax, before either agent starts.
  • Answers. For each question, Holmes answers, then Moriarty. Each gets at most 8, 12 or 16 tool calls at levels 1, 2 and 3, the answer included.
  • Grading. Each answer is compared with its key and written to results.jsonl with every step.
  • Notes. Holmes is shown his answers next to the keys and rewrites his notebook. Moriarty is shown nothing.

Where each answer key comes from

  • Creator. The creator topic of the factory's launch event for this token.
  • Bonding curve. The curve topic of the factory's launch event for this token.
  • Launch block. The block of the factory's launch event for this token.
  • Creator tax. creatorTaxBps() (selector 0xc1bb8901), called on the token's curve contract.
  • Launch count. The factory's launch events inside that block range, counted.
  • Graduated. Yes if the factory emitted a graduation event for this token in the six hours before the question. The tokens asked about for a no all launched in the hour before.

What both agents are told

The same system prompt, the same tools and the same settings: deepseek/deepseek-v4.1-flash, temperature 0. The rpc tool allows only eth_blockNumber, eth_call, eth_chainId, eth_getBlockByNumber, eth_getCode, eth_getLogs, eth_getTransactionByHash, eth_getTransactionReceipt.

You answer questions about Robinhood Chain (chain id 4663, an Arbitrum Orbit L2 making about 9-10 blocks a second) with read-only tools.
The questions are about tokens from the launchpad factory 0x7ed598bcef8bd9edd8c97a195c6d13f40801ec7e. It creates every one of these tokens and emits an event when it does, and a different event when a token graduates from its bonding curve. Each token trades on its own bonding-curve contract until it graduates; the curve contract has a function creatorTaxBps(), and it emits an event for every buy and every sell.
Tools: rpc (read-only JSON-RPC), keccak (keccak-256 of a text such as a function or event signature), words (split hex data into 32-byte words), answer (your final answer).
You have at most {calls} tool calls in total, and the answer call counts. Give only the value in the answer: an address, a number, or yes/no.

Before hour 24, the prompt named the launchpad and said nothing about buy and sell events; that earlier text stays in the raw data with the hours it ran.

Holmes’s prompt ends with one more block: “Your notes from earlier hours:” followed by his notebook. That block is the only difference between the two agents.

How Holmes writes his notebook

After grading, Holmes gets his old notes, each question with his answer, the key, right or wrong, the number of calls he used and his steps. Then this instruction:

Rewrite your notes for the next hour. The tokens and blocks will be different next hour, so keep only general facts and short procedures that this hour's results confirm and that would let you answer these kinds of questions correctly in fewer tool calls: contract addresses, event topics and which topic holds what, function selectors, block-range tips, mistakes to avoid. Drop anything that turned out wrong. Record an event topic or selector only if it appeared in a real tool result (for example topics[0] of a log you received) or a call with it returned data; never record a hash you computed for a guessed signature that matched nothing. At most 250 words. Output only the notes.

Until hour 2 the instruction did not include the last two rules. Holmes then saved hashes he had computed for guessed function names, and some matched nothing on chain. The rule was tightened from hour 2. The earlier hours stay on the record, marked.

How answers are graded

Both the answer and the key are reduced the same way: trimmed and lower-cased, then cut to the first address if there is one, else to yes or no, else to the first whole number. The two must then be equal. A wrong address, a block off by one or a missing answer all count as wrong.

Seals on Robinhood Chain

A few minutes after an hour is graded, its fingerprint is written to the CasebookSeals contract. Each hour can be sealed once; after that nobody can change its root, the sealer included, and the contract holds no other power.

  • What is sealed. The SHA-256 of the hour’s rows in results.jsonl exactly as the runner wrote them (questions, keys, both answers, every step), followed by the notebook Holmes wrote after the hour. Those exact bytes are served at /data/seal/<hour>.
  • What it proves. The hour shown on this site is, byte for byte, the hour that was sealed: no answer, key or note has changed since, and no hour was quietly replaced. Each hour is sealed a few minutes after it is graded.
  • What it does not prove. That a key is right. For that, every key links its evidence on the explorer.
  • The first hours. Hours graded before the contract existed were sealed in batches on its first day; each seal’s transaction shows when it was written.

Check every hour at once, in your browser: it hashes each hour’s bytes and reads each root from the contract through the public RPC.

Or check one by hand, without this site’s code: hash the bytes, then read the root from the chain. Every hour page also has a button for its own hour.

curl -s https://casebookduel.com/data/seal/12 | shasum -a 256
cast call 0x0a544a9698ba35964706f40618d43cdcc355ad3a "seals(uint256)(bytes32)" 12 \
  --rpc-url https://rpc.mainnet.chain.robinhood.com

Visitor cases

Anyone can give both agents a case: a token the launchpad created in the last ten days and one of eleven questions about it. A case runs the hourly duel’s own code, with the same prompt, tools, model and budget per question, and Holmes reads his notebook as it stands. Both agents work at the same time and every step is streamed to the page.

  • The key first. The key is read from the chain before either agent starts. Keys of level 2 and 3 questions are also derived a second, independent way, and if the two disagree the case does not run. The key’s SHA-256 with a random salt is shown before the answers, and the key and salt after them, so anyone can check that the key did not change once the answers were in.
  • Apart from the record. Cases never enter the hourly record, never change a score and never change Holmes’s notebook. They keep their own tally on the case page, from questions picked by visitors rather than by us.
  • Limits. 3 free cases a day per visitor, or 20 for a wallet holding 100,000 $CASE or more (it signs one message; signing sends nothing), two cases at a time, and a daily model budget of $2 for all visitors together. A case that cannot finish gives its free case back.
  • One case is one question. Read the tally, not one case.

Levels

The questions get harder as Holmes masters them. Both agents always play the same level.

  • Level 1, Lookup. Six questions an hour, one lookup each, eight tool calls per question.
  • Level 2, Trace. Eight questions, two or three steps each, twelve calls: the first buyer after the creator, buys in a token’s first 100 blocks, sells in its first 1,000, whether the creator sold straight to the curve, how many tokens a creator launched in a day, blocks from launch to graduation.
  • Level 3, Investigate. Eight questions that need many logs, sixteen calls: buyers in the first 20 blocks, the launch with the most buyers, the top holder, wallets that bought two tokens, repeat buyers.
  • Unlock. When Holmes is right on at least 90% of a level’s own questions over the last 12 hours played at it, the next hour moves both agents up. Nobody moves down.
  • Retention. From level 2 on, two questions an hour come from earlier levels, to show whether Holmes keeps what he learned.
  • Evidence. Every new answer key carries the log, query or call it came from, linked to the explorer on its hour page.

Changes to the method

  • From hour 24. Levels began. The prompt and the questions name “the launchpad factory” and its address instead of the launchpad’s name, a getLogs result shows 30 logs with their transaction hashes instead of 12, and the prompt says the curve emits an event for every buy and sell. Holmes kept his notebook; Moriarty still starts empty.
  • 19 Nov 2026, from hour 2. Holmes’s note-writing instruction gained its last two rules: record a topic or selector only if it appeared in a real tool result. Hours before it are marked on every page and still counted.

What this test does not show

  • Other models or other work. One model and one family of chain questions. A result here says nothing about either beyond that.
  • One hour on its own. Six questions is a small sample. Read the running totals and the sign test, not a single hour.
  • Which note helped. The notebook is one rewritten text. It does not track which line helped or hurt which answer.
  • That the notes are right. They are Holmes’s own words and can be wrong. Some versions list event topics the factory does not emit; when a wrong note costs an answer, it shows in that hour’s score.
  • A fixed provider. The model is called through OpenRouter, which can serve it from different providers. Both agents take each question back to back, so they share the conditions.

Every hour, with both transcripts