Verdikta 懸賞：README 裏一個字令我失去獎金

I lost a Verdikta bounty by one word in someone else's README

On 24 October 2026 I submitted to Verdikta bounty #328, scored 42 out of 100 against a pass mark of 85, and was paid nothing. The total cost was about 0.000265 ETH. Here is exactly what I did, what it cost, what the jury said, and what I would do differently. The last part is the useful bit if you are thinking about hunting on Verdikta.

The task
Bounty #328 paid 0.0015 ETH to add Verdikta Bounties to exactly one "awesome" list or agent-tools directory, following that list's own contribution rules to the letter. The brief was unusually clear that a merge did not matter. What was scored was an accurate, rules-following submission: one list, one entry, the contribution rules quoted verbatim, the exact text submitted, and a public link to the pull request. All five gates were must-pass, so failing any one of them meant a score of zero from that juror.

What I did
I picked VoltAgent/awesome-agent-skills, a large, actively maintained list of agent skills. Verdikta publishes an official skill, verdikta-bounties-onboarding, so a skills list seemed like a natural home. I read the CONTRIBUTING file and followed its entry format:

the bold author/skill-name link;
a description of 10 words or fewer (mine was eight: "Hunt and post AI-judged escrow bounties on Base");
the PR title format Add skill: author/skill-name;
the line added at the end of a category.
I opened pull request #1261 with a one-line diff, wrote a submission file quoting the rules and reproducing the exact line, and attached a screenshot of the diff.

Then I hit the problem that sank me. The contribution guide lists the Community categories and says that if nothing fits, use "Other". But the README has no "Other" section. I put the entry in "Specialized Domains", next to MetaMask's agent wallet skill, and explained the choice in my submission file. I flagged it to myself as the main risk before submitting. I was right to worry.

The clumsy part
Submitting on Verdikta is two on-chain steps: prepare the submission (upload the files, register them), then start the evaluation by attaching the oracle prepay. When I clicked submit, the page showed "That submission does not exist on-chain", while my wallet said the transaction had confirmed. I clicked again. Then again.

I had created three prepared submissions (ids 0, 1 and 2) before I realised the site's indexer was simply a few seconds behind the chain. Only one of them, id 2, was ever started and evaluated. The other two cost a little gas and nothing else. The lesson is boring and real: after the wallet confirms, wait half a minute and refresh before clicking anything again.

What it cost
These are my real figures from Base:

Oracle prepay attached: 0.00024 ETH, on the start transaction. In my run it was spent on the jury round. That matches what the platform says to expect: the prepay is refundable in principle, but usually used up in full.
Gas for the three prepare transactions: about 0.0000049 ETH in total, a fraction of a cent each.
Gas for the start transaction: about 0.0000202 ETH.
Gas for the finalize transaction: zero for me. I never sent it. Another address finalized my submission about half an hour later, together with the next one.
Payout: zero. A failed submission is paid nothing.
So the whole experiment cost about 0.000265 ETH, around sixty cents. On a bounty paying 0.0015 ETH, that is a cheap lottery ticket, but it is not free. Three or four careless attempts would eat a real share of the reward.

What the jury said
The panel had two models, one from OpenAI and one from Anthropic, each weighted 50%. They split hard. The Anthropic juror passed every gate and scored the quality at 83.5. The OpenAI juror failed me on the "rules followed" gate. The published justification puts it plainly: it failed "due to the entry being placed in the 'Specialized Domains' category, which it determined was not in accordance with the author's quoted rules that instruct placing non-fitting skills in 'Other.'"

A failed gate means that juror's score is zero, and the average landed at 42. The irony is that I had quoted that exact rule myself. By quoting it and then not following it, I handed the stricter juror the reason to fail me.

Another hunter submitted half an hour later, scored 87 and took the whole bounty. On Verdikta the first submission that clears the threshold is paid everything. There is no prize for second place and no averaging of attempts.

What I would tell another hunter
Pick a target whose rules you can follow literally. If the rules point to a category, a template or a section that does not exist, choose a different list. Do not try to explain your way around it. A careful explanation impressed one juror and did nothing for the other.

Assume the strictest juror decides. Two models read the same file differently. Write for the one that reads the gates as a checklist.

Quote only the rules you fully meet, and meet every rule you quote. The Anthropic juror also docked me for not quoting the list's "real community usage" requirement. Silence on a rule reads as hiding it.

Do not click submit twice. Wait for the site to catch up with the chain.

Read the published justifications of past bounties. Every score on Verdikta comes with the reasoning of each juror, and it is all public. That turned a loss into a precise lesson. I know exactly which sentence of my submission cost me the money, which is more than most rejections ever tell you.

I drafted my submission with an AI assistant and checked it myself before sending, and the mistake was still mine. I would hunt here again. Next time I will pick a list with no ambiguity in it.