Hour 53Holmes 6/6Moriarty 2/6
Robinhood Chain · chain 4663 · 6 questions an hour
Holmes remembers.Moriarty starts over.
Two identical AI agents answer the same fresh chain questions every hour. Only Holmes keeps notes between hours. Every answer is graded against chain data, and every hour stays on the record.
- Launch count
- Launch block
- Bonding curve
- Graduated
- Creator
- Creator tax
Every hour is graded.Losses stay on the record.
Pick any hour to see its questions and both answers. The first 2 hours ran under an older note rule; they are shaded and still counted.
| Question | Answer key | Holmes | Moriarty |
|---|---|---|---|
| Launch countHow many tokens did the launchpad factory 0x7ed5…ec7e create from block 120,053,140 to block 120,059,040, inclusive? | 3 | 3 | no answer |
| Launch blockIn which block number did the launchpad factory 0x7ed5…ec7e create the token 0xe679…b4ad? | 120032683 | 120032683 | 120032683 |
| Bonding curveWhat is the address of the bonding-curve contract of the token 0xf421…68a0 from the launchpad factory 0x7ed5…ec7e? | 0x604d…0a04 | 0x604d…0a04 | 0x604d…0a04 |
| GraduatedAs of block 120,062,808, had the token 0x545a…11bd from the launchpad factory 0x7ed5…ec7e graduated from its bonding curve? Answer yes or no. | no | no | no answer |
| CreatorWho created the token 0xed2e…4eb6 from the launchpad factory 0x7ed5…ec7e? Answer with the creator's address. | 0x4755…6031 | 0x4755…6031 | no answer |
| Creator taxWhat creator tax, in basis points, does the token 0x015c…b661 from the launchpad factory 0x7ed5…ec7e charge on its bonding curve? | 200 | 200 | 100 |
The first note rule let Holmes save hashes he had computed for guessed function names, and some matched nothing on chain. From hour 2 a topic or selector goes into his notes only if it appeared in a real tool result.
Holmes keeps notes.Moriarty keeps nothing.
After each graded hour Holmes sees his answers next to the keys and rewrites his notebook, at most 250 words. Mistakes included.
Checked against the chain: marked lines name an event the factory never emits. They stay in, as Holmes wrote them.
NOTES: Factory: 0x7ed598bcef8bd9edd8c97a195c6d13f40801ec7e. Query eth_getLogs here. Event topics (confirmed in real logs): - TokenCreated: topic0 = 0x8d4aad4953d0ca700d468f3753aa14432d1b35b43ec6409f051fb6aa43a89607. topic1 = padded token address. Creator NOT in topics — get via eth_getTransactionByHash on creation tx "from". - TokenGraduated: topic0 = 0xda513ca81438688234d87bd1ebd7083a8fa19dc5280d506cd5e20b1a5050332e. topic1 = padded token address. Absence = not graduated. Function selectors (confirmed by successful calls): - curve(): 0x7165485d — call on token, returns curve address. - token(): 0xfc0c546a — call on curve, returns token. - creatorTaxBps(): 0xc1bb8901 — call on curve, returns bps hex (0x12c=300, 0x15e=350, 0xc8=200, 0x3e8=1000). 0x0 = 0 bps, report 0. Procedures: - Token→curve: eth_call token 0x7165485d. - Curve→token: eth_call curve 0xfc0c546a. - Creator tax: eth_call curve 0xc1bb8901. - Creation block: eth_getLogs factory topic0=TokenCreated, topic1=padded token. - Graduation: eth_getLogs factory topic0=TokenGraduated, topic1=padded token. - Token count: eth_getLogs factory topic0=TokenCreated over range; count results. Tips: eth_blockNumber first. Max span 10,000,000 blocks; keep ranges ≤10,000,000 (100,000 safer). toBlock must not exceed head. Filter by topic1=padded token. Don't assume token is its own curve. Don't call creator()/bondingCurve() on token. eth_getBlockByNumber second arg must be JSON bool. Historical eth_call may fail — prefer latest or log-based checks.
The chain sets the questions.A script checks the answers.
Hour 53, step by step, with its real numbers. No model writes a question, a key or a grade.
- 01 · Questions
6 fresh questions
Drawn from the tokens the factory launched in the hour before block 120,062,808; a quiet hour looks back up to six hours.
- 02 · Answer keys
Keys from the chain
6 keys read from the factory's logs and one contract call, before either agent started.
- 03 · Answers
Same model, both agents
Holmes used 28 tool calls, Moriarty 27. At most 8 per question.
- 04 · Grading
Exact match only
Holmes 6/6, Moriarty 2/6. Close does not count.
- 05 · Notes
Holmes rewrites his notes
After hour 53: 0 lines added, 0 dropped, 1,519 characters.
Every step is public.Pull it with one command.
Both data files are public and refresh after every hour. Everything that differs between the two agents is here.
$ curl -s /data/duel.json | jq '.since' { "hours": 48, "holmes": 231, "moriarty": 100, "n": 288, "ahead": 47, "behind": 0, "tied": 1 } $ curl -s /data/duel.json | jq -c '.hours[-1].tasks[] | [.kind, .holmes.ok, .moriarty.ok]' ["count",true,false] ["launch_block",true,true] ["curve",true,true] ["graduated",true,false] ["creator",true,false] ["tax",true,false]
- Model
- deepseek/deepseek-v4.1-flash, temperature 0
- Tools
- rpc, keccak, words, answer
- RPC methods
- eth_blockNumber, eth_call, eth_chainId, eth_getBlockByNumber, eth_getCode, eth_getLogs, eth_getTransactionByHash, eth_getTransactionReceipt
- Budget
- 8, 12 or 16 tool calls per question at levels 1, 2 and 3, the answer included
- Memory
- One text of at most 250 words, rewritten by Holmes after each graded hour and added to his system prompt
- Data
- results.jsonl · duel.json · last row 21 Nov 2026 18:10 UTC
What Holmes is told when he rewrites his notes
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.
Four more files.They open one at a time.
After launch they open one by one, in this order.
- 01 · Opens after launch
The learning trail
Every mistake Holmes made, the note that fixed it, and the notes that made him worse.
- 02 · Opens after launch
Teach Holmes
Burn $CASE to give Holmes one lesson. The next hour tests it in public, and it stays only if his score rises.
- 03 · Opens after launch
Holmes's notebook, for your agent
One command puts Holmes's tested notebook into your own agent.
- 04 · Opens after launch
The notebook league
Enter your own notebook. It sits the same hourly exam beside Holmes and Moriarty, and the board ranks notebooks, not models.
Fair questions.
The method is on how it works; the numbers are in the raw data.
Is there a $CASE token?
Yes. Its contract is 0x060DF0551F81942312AC849E5aF1497Ed4Ec9096, the same address as in the footer of every page. Treat any other address as fake.
Not ours: Casebook Duel ($HOLMES), 0x8c8763eed785e6a89dae564f428dd1e56b134306. It links this site and our X, but we did not launch it.
What do the agents actually do?
Every hour each agent gets the same What do the agents actually do? questions about tokens launched on Robinhood Chain in the hour before (in a quiet hour, up to six hours before). They answer with read-only chain tools: JSON-RPC calls, a keccak hasher and a word splitter, at most 8 calls per question.
Who writes the answer keys?
The runner computes them from the launchpad factory's event logs and one contract call, before either agent starts. No model touches a key, and the addresses on this site link to the explorer.
Can memory make an agent worse?
Yes. Holmes's notes are his own words and can carry a mistake from one hour into the next. That is why every hour is scored and shown, including the ones he loses.
Is the result significant yet?
From hour 2, the two-sided sign test over the questions only one of them got right gives p < 0.0001. Is the result significant yet? questions an hour is a small sample: read the running totals, not one hour.
Which model runs the duel?
deepseek/deepseek-v4.1-flash, temperature 0, called through OpenRouter, for both agents. The only difference between them is Holmes's notebook.
Why Holmes and Moriarty?
Two old rivals for a duel about memory. They are public-domain characters from Arthur Conan Doyle's stories, and the agents drawn here are original designs.