Home/Datasets/Nonce Archaeology

Dataset · v1 · sealed 2026-07-29

Nonce Archaeology

Bitcoin's hardware history, in a number with no further use.

Every block records the number the miner happened to stop on. It is never read again. It turns out to hold a legible record of what hardware was doing the guessing — including the exact point GPUs arrived.

Price

$19

Tables

Rows

960,916

Download

35.4 MB

Formats

Parquet + CSV

Measured twice

Independent second pass and controls that must be able to fail.

Own node

Raw blocks parsed on our archival hardware — no rented API.

Bitcoin-anchored

Root hash committed on-chain via OpenTimestamps.

Verify free first

Manifest, samples, and verify.py free before you pay.

Check free first

Judge the seal before you pay. These files are free right now:

How verification works →

Buy this dataset

One-time purchase. Instant download, no account, no subscription.

Everything needed to judge this dataset — manifest, samples, verifier, Bitcoin timestamp — is free below, right now.

Why this isn't free data

The statistics are the product

Height and nonce are in every free chain dump. Extracting them is worth nothing. The work is the normalisation that makes the 2010 collapse visible, the pool conditioning that places the modern deviation in equipment rather than protocol, and knowing how to bin the field without manufacturing a result.

That last one is not hypothetical. Binning this exact column with a cast rather than a floor produced a published finding here that did not exist, and we retracted it. The same mistake shifted this dataset's epoch table by half a bin during the build — caught because the export disagreed with the free finding by 0.046.

Collapse floor

0.1218

At epoch

60,000

Below midpoint

96.33%

Lowest nonce ever

1,082

Uniform searching would give 1.000 and 50%. The free finding explains why 2010 does not, and how it dates the arrival of GPU mining.

Check us before you pay

Free downloads

Sample rows

1,000 real rows of the era-bounded signal table, and the complete version-rolling-by-year series.

nonce_by_epoch, complete (97 rows)
lowest_nonces sample (100 rows)

The manifest

Schemas, per-column digests, and every figure with the exact SQL behind it — including the deliberately wrong one.

manifest.json

The proof

Root hash and OpenTimestamps receipt, committing this dataset to the Bitcoin blockchain.

root.hash · root.hash.ots

The verifier

Recomputes everything from the files alone, and re-runs every control against your own copy.

verify.py

3b0a29af5a1afab79097207922e4439e64b5cd68907741d32655a0474ddbc280

Contents

What's in the box

Five tables, delivered as both Parquet (ZSTD) and CSV.
Table	Rows	Parquet	CSV	What it is
block_nonces	960,193	14.47 MB	59.43 MB	Every winning nonce with time and position in the search space
nonce_by_epoch	97	< 0.01 MB	< 0.01 MB	Per 10,000-block epoch — where the 2010 collapse appears
nonce_by_year	18	< 0.01 MB	< 0.01 MB	The same statistics by calendar year
nonce_by_pool_year	108	< 0.01 MB	< 0.01 MB	Per pool per year, with the deviation in sigma
lowest_nonces	500	0.01 MB	0.02 MB	The 500 luckiest blocks ever mined

Schema

block_nonces        height BIGINT · block_time TIMESTAMP · nonce BIGINT
                    nonce_frac_of_space DOUBLE · below_midpoint INTEGER

nonce_by_epoch      epoch_start DOUBLE · blocks BIGINT · mean_nonce DOUBLE
                    mean_over_midpoint DOUBLE · frac_below_midpoint DOUBLE

nonce_by_year       year INTEGER · blocks BIGINT · mean_over_midpoint DOUBLE
                    frac_below_midpoint DOUBLE

nonce_by_pool_year  year INTEGER · pool VARCHAR · blocks BIGINT
                    frac_below_midpoint DOUBLE · sigma DOUBLE

lowest_nonces       height BIGINT · block_time TIMESTAMP · nonce BIGINT

Use version_unsigned for any bit masking. The version field is stored signed, and the signed form silently inverts every high-bit test — a trap we ship pre-sprung rather than let you find.

Quality

The controls this build had to pass

The export deletes its own output if any of these fail. All five run against the delivered files, so you can re-run every one of them:

null_period_2011_2014_is_flat — 2011 to 2014 must read 1.000. It is the internal null: if our parse skewed nonces it would skew them here too, which would make the 2010 collapse an artefact rather than a finding.
early_collapse_is_present — the 2010 collapse must survive the export. It is the headline; absent it, the product is not what it claims to be.
modern_era_is_only_subtly_off — the modern chain must be close to uniform. A parse fault would make it wildly off, and every pool comparison downstream would be meaningless.
nonce_fits_in_uint32 — the nonce is a uint32; anything outside that range is a parse fault.
pool_panel_spans_many_years — the pool table must cover many years, not only the window where the effect shows. Cherry-picked windows are how weak claims survive.

Honest limits

What this is not

No mechanism is claimed. The modern deviation is real, dated and pool-dependent. Why it happens, we do not know. The timing overlaps the adoption of version-rolling firmware, which is correlation and nothing more.
Pool attribution is self-reported. It comes from our own copy of the coinbase message; proxy pools blur it, and pool-years under 400 blocks are excluded as underpowered rather than shown as noise.
Aggregation window changes the number. This dataset reports per year, so F2Pool 2020 reads 17.9σ. The free finding pools 2020–2021 and reads 24.2σ. Both are correct for the question asked — we mention it so you are not surprised when your query differs from our page.
It is a snapshot to height 960,192. The chain keeps moving.