Home | Search

2026-08-30

Update

I found better solutions to this problem and have made significant edits to the post.

• The part 1 (signed web crawls) part can be done using merkle proofs, you don't need to involve ZK here as everything is public. Edited post to talk about merkle proofs not ZK proofs.
• The part 2 (trust-minimised verification of AI responses) has a better solution not mentioned in the original post. I realised you will soon (likely in next 1-2 years) be able to generate ZK proofs of AI responses, thus eliminating the need for an oracle in part 2 entirely.

2026-08-18

minor edits 2026-08-30

Signed web crawls might help solve the blockchain oracle problem

Disclaimer

• Quick Note
• Contains politically sensitive info
• I don't endorse actual whistleblowers in high-stakes situations taking the below post too seriously. These are brainstorming notes.

Background context on what problem I want to solve

• I wanted a way for an ethereum smart contract to trustlessly verify whether a specific domain https://alice.com hosted some specific content or not. This could help solve the oracle problem.
• I also wanted the oracle to be censorship-resistant. I didn't want it to honestly respond to some requests and either lie or refuse to answer others based on political pressures. I wanted the oracle to honestly respond to all requests.
• My specific use case for solving the oracle problem was paying people who did certain leaks. But there are many other use cases where solving the oracle problem could help.
• See also: Ethereum smart contracts cannot trustlessly verify authenticity of web data (by verifying TLS and CA chain-of-trust) or emails (by verifying DKIM headers)
• Unfortunately I found out that TLS connections by default do not contain any signatures that can be proven to people outside the connection.

Potential solution using signed web crawls

• Insight: A partial solution here might be to trust signatures on top of web crawls?
• The full setup is quite convoluted but here it goes.
• Part 1 of full setup:
  • Oracle question: Did there exist a US govt whistleblower in the year 2017 who was morally supported by atleast 2 out of 3 organisations among the EFF, the ACLU and FPF?
  • Link to ACLU page on Reality Winner, Link to EFF page on Reality Winner
  • CommonCrawl has probably crawled both of these pages. Link to CC crawl of ACLU page on Reality Winner, Link to CC crawl of EFF page on Reality Winner
  • Ask a multisig or proof-of-stake DAO (like UMA optimistic oracle) to report what the correct merkle root of CC-MAIN-2023-06 is. Ask the same multisig or PoS DAO to report what the correct merkle root of CC-MAIN-2017-47 is. Trust that the oracle is 51% honest.
  • Generate a merkle proof that these specific pages (ACLU page, EFF page) are part of the same archives whose merkle roots the multisig or DAO has attested to. Use an ethereum smart contract to trustlessly verify these merkle proofs.
• Part 2 of full setup: (This part is optional and uses AI)
  • Give these webpages along with hard-coded prompt to an AI.
    • Hard-coded prompt example: Does the attached webpage provide conclusive proof that the answer to the question is "Yes"? Question: Did there exist a US govt whistleblower in the year 2017 who was morally supported by atleast 2 out of 3 organisations among the EFF, the ACLU and FPF?
    • AI model could be hard-coded to be gpt-oss-120b for example.
  • Ask a multisig or proof-of-stake DAO to report what the AI answered. Trust that this oracle is 51% honest. Generate a ZK proof of this AI response, and verify it using a smart contract.
  • Use this response to actually resolve the original question.
  • Use the resolved question to actually trigger a payment to the leaker/whistleblower. (Obviously we can't go back in time and pay Reality Winner now, but the idea is to use this setup for whistleblowers in future.)

Why I like this solution?

Simple answer is censorship resistance.

• At no point did we ask the oracle Did Reality Winner whistleblow against US govt in 2017?
• If you ask the oracle a question this specific, and the people operating the oracle know that their answer will be used to release the payment, they might 51% collude and lie or refuse to answer this question. They could still remain honest on other questions.
• Instead we asked What is the merkle root of CC-MAIN-2023-06? If there are many questions that all resolve based on this merkle root, it is hard for the oracle to lie to this question.
• This makes part 1 quite censorship-resistant.
• There is trusted dependency on CommonCrawl, for instance CommonCrawl could refuse to crawl any pages that mention Reality Winner, if they know their pages will be used by the oracle to release the payment. You can mitigate this somewhat by relying on N different crawls, such as CommonCrawl, Internet Archive and others, and trust that 51% of the crawls are both honest and contain the relevant pages.
• Unfortunately, part 2 is still not censorship-resistant. The oracle can 51% collude and lie to the question What will gpt-oss-120b respond if you give it XYZ input? for some inputs, but be honest on other inputs.
• You can solve this by distributing among N different inference providers who all sign their requests, and trusting 51% of those to be honest, and asking the DAO/multisig to simply report the pubkeys of these N providers, and trusting the DAO/multisig to also be 51% honest here.
• The major bottleneck for this however, is that as of 2026-08, there do not exist that many AI inference providers who will sign their responses.

Subscribe

Enter email or phone number to subscribe. You will receive atmost one update per month.

Alternatively, you can subscribe via RSS, or a third-party service like blogtrottr

Comment

Enter comment