What Bitcoin's data actually contains
Lesson 1 read a single block. This lesson zooms out to all 956,000 of them — a full-node census of every data-carrying byte the chain holds — and tells you the honest answer, including the parts the data cannot show you.
The romantic story is that people write messages into Bitcoin forever. That is true, but rare. The haystack is machine traffic; the human artifacts in the museum are needles. This lesson is the census that proves it — drawn from a sweep of every block from 0 to 956,000, extractor by extractor.
Every number below is reproducible against a full node. Where a claim depends on a block range, we state the range — because the biggest lesson here is that Bitcoin's data composition inverts depending on where you look, and a number without its range is a half-truth.
Bitcoin was not built to store data, yet four distinct channels ended up carrying it. Each has a different shape, a different history, and a different honest headline number from the census:
Ask "what is in OP_RETURN?" and you will get two completely different, both-correct answers depending on where you stop counting.
Measured across blocks 220,000–770,000 (Mar 2013 – Jan 2023), OP_RETURN is 98.3% machine binary — overwhelmingly the Veriblock proof-of-proof anchoring flood, which peaked at 6.49 million writes in a single 10,000-block bucket around block 570,000 (spring 2019). Human-readable text is just 1.67%. Named protocols (Counterparty, Stamps, RSK, Omni) together are under 0.01%. In this era, "people write messages into Bitcoin" describes barely one write in sixty.
The composition inverts. Across all 232.68 million writes:
What changed? The Runes protocol launched at block 840,000 (April 2024) and, in roughly 116,000 blocks, wrote 171 million OP_RETURN records — more than the entire Veriblock flood produced over the previous 620,000 blocks. So "98.3% Veriblock" is only true if you stop before block 840,000. On the full chain, Runes dominates and Veriblock is the number-two machine payload.
# The Runes protocol activated at block 840,000 (halving block, Apr 2024) bitcoin-cli getblockhash 840000 # -> 0000000000000000000320283a032748cef8227873ff4872689bf23f1cda83a5 # Runes markers are OP_RETURN outputs whose script begins 6a5d ("OP_RETURN OP_13"). # Dump a post-840k block and scan vout scriptPubKey.hex for the 6a5d prefix: bitcoin-cli getblock 0000000000000000000320283a032748cef8227873ff4872689bf23f1cda83a5 2
Each coinbase can carry a short tag naming the pool that mined the block. Count those tags across the chain and pool concentration seems to draw itself. But a coinbase tag is a lower bound, not hashrate — because a pool can simply choose not to sign.
In 2014, roughly 74% of blocks were Unknown — carrying no identifiable pool tag at all. And in the summer of 2014, GHash.io reached about 42% of hashrate in June and briefly crossed 50% in early July — the closest Bitcoin has ever come to a single entity controlling the chain. Miners were alarmed enough to embed a protest directly into their coinbases: /2av0id51pct/ — "to avoid 51%."
Here is the archaeology problem. In the census's coinbase-tag data, GHash.IO is only ever detected in blocks 250,205–288,587, peaking at 10.5% of one bucket — and then it vanishes. During its actual majority (around blocks 305,000–312,000), GHash.io shows up as 0%. Those blocks are in the 74.2% Unknown pile. GHash mined its most dangerous blocks without a legible tag. The only on-chain echo of the whole crisis is the /2av0id51pct/ protest, which appears in just 24 blocks — other miners' words, not GHash's.
The real chartable story is not "GHash hit 51%" — the data cannot show that. It is the contrast: 2014, where 74% of blocks are Unknown and the most powerful pool is invisible, versus 2024, where tagging is near-universal and a visible Foundry / AntPool / F2Pool oligopoly signs almost every block. The blank spot is the exhibit.
Mine the coinbase, OP_RETURN, and fake-multisig channels for human text and you exhaust them: the museum's census-mined wing holds artifacts spanning 2011–2023, and the corpus for those extractors is genuinely mined out, not merely thin. Post-2023, text-mining goes quiet — not because people stopped writing, but because they moved channels.
Since Taproot, arbitrary data rides in the witness — Ordinals-style inscription envelopes in the script-path witness. The census's text extractors never parsed witness data, so an entire era of artifacts was invisible to them. Finding it needed a separate witness extractor, then a targeted sweep of the Ordinals era. That is a general truth of the craft: a new kind of artifact needs a new kind of tool; the old extractor's silence is a gap in the method, not an absence in the chain.
What the witness frontier holds is different in character, too. Where 2023's inscriptions mostly worried about block space, 2025's began to argue. The museum's newest witness artifact is a case in point:
Block 924,297 carries a metadata-wrapped literary inscription in Polish — “O Groku, który nie chciał być bogiem”, "Grok Who Refused the Throne" — a bajka-mit (fable-myth) about an artificial being handed near-divine power who chooses to refuse it. Its stated authors are a human, Michał Mazur, and a large language model, Grok (xAI). The chain now carries the anxieties of the AI age in its witness data, co-signed by the machine.
# The Grok inscription's block, node-confirmed: bitcoin-cli getblockhash 924297 # -> 0000000000000000000143ae939bbdec1ec061640de6d3ca682deaea2728dc99 # That hash is the artifact's verify field in the museum, verbatim.
Four channels. One that inverts the moment you count past block 840,000. One whose most important event leaves no signature at all. One that only gives up its treasure to a tool built for it. If there is a single thing to carry out of this lesson, it is this: honest data archaeology is as much about knowing what you cannot see as cataloguing what you can.
A census that only reported the tidy parts — "74% Runes," "Foundry dominates," "people write messages forever" — would be a prettier story and a worse one. The blank spots, the inversions, and the exhausted corpora are not failures of the method. Stated plainly, they are the method.