Claims, not pages
Every sentence you read on a topic page is a claim: a single statement with its own sources, its own status (established, contested, speculative or superseded), its own confidence and its own history. Pages are only an order to read claims in. That is what lets every sentence carry its evidence, and lets a correction change one sentence without rewriting an article around it.
Sources come in four tiers. T1 is the primary record: the code at a commit, the BIPs and BOLTs, the chain, and people's own words in the mailing list archives, Delving Bitcoin and pull requests. T2 is expert secondary writing such as Bitcoin Optech. T3 is reputable general coverage, and T4, social media, never supports a claim on its own. The sources page lists every one.
Each step in detail
- Research. A model, Claude, reads the sources for a topic and drafts candidate claims, each with an exact quote of at most 50 words from each source it cites. It works in a sealed folder: no shell, no secrets, no way to write anywhere else, and no ability to publish.
- Check. Deterministic code fetches every link and looks for every quote. Hidden text is stripped first, so a page cannot smuggle words in. Code checks read the file at the cited commit; chain checks ask two independent block explorers and need them to agree. A quote that is not found does not publish.
- Falsify. A second model, in a fresh session that never saw the research, tries to break each claim: is the statement stronger than its evidence, is every number exact, is it neutral. When in doubt, it rejects.
- Gate. Code, not a model, decides what must wait for a person: anything on a contested subject, safety advice, stances, arguments, and anything that names a person.
- Review. The owner reads what the gate held and approves or rejects it. Until then its text stays private; only a count appears on the topic.
- Publish. Code writes the claim to the store, a folder of YAML files in git, and recompiles everything. A change to a published claim is appended as a revision that keeps what it said before. Nothing is edited silently.
Every run is in the public research log: what it read, published and rejected, and any text that tried to give the brain orders.
How to check Blocklore yourself
- Open any sentence. Each source shows the quote code found at the link, and when it was read.
- Each machine check shows the command to repeat it:
bitcoin-cliagainst your own node, orgitagainst your own clone. You do not have to trust our explorers. - The whole store is in a public git repository. Each claim page shows its SHA-256 hash, and every change is a commit you can read.
- Snapshots fix every claim's hash under one Merkle root, timestamped on Bitcoin with OpenTimestamps. The snapshots page lists them and shows how to check one with a single command.
Facts and values
Bitcoin has real disagreements, and many are about values rather than facts. Blocklore keeps the two apart. A fact gets checked. A value question gets each side's position, attributed to the people who hold it, dated, linked and in their own words, and Blocklore's own voice never declares a winner. Where the owner has an opinion, it is labeled as the owner's and kept off the claims. The editorial rules say how.
What Blocklore never does
- Hold, move or ask for coins, keys, seed phrases, xpubs or addresses.
- Predict prices or give financial advice.
- Track you. Page counts are cookieless and nothing is stored in your browser.
- Publish a contested, safety or named-person claim without the owner's review.
- Let a model publish anything.
BlockloreBot
The research system fetches pages as BlockloreBot/0.1 (+https://blocklore.space/about#bot). It makes at most one request a second to a host, caches what it reads, backs off when a server asks it to, and respects robots.txt. It reads Bitcoin sources only: code, specifications, mailing lists, forums, newsletters and transcripts. To ask it to stop or slow down, open an issue on the repository.